Skip to content

Releases: facex-engine/facex

facex_nano recognition (ONNX)

14 May 23:56
@bauratynov bauratynov
facex-nano-1.0
8846599

Choose a tag to compare

View all tags
facex_nano recognition (ONNX) Latest
Latest

nano variant of our MobileFaceNet recognition stack. 200K params, 800KB, LFW 95.62% mean over 10 folds (std 1.11%).

input: 112x112 RGB float32 normalized to [-1, 1] (HWC, then transposed to CHW). face must be 5-point ArcFace aligned. output: 512-dim L2-normalized embedding.

reproduce LFW numbers with the bundled lfw_eval_onnx.py:

pip install onnxruntime pillow numpy
python lfw_eval_onnx.py facex_nano.onnx lfw_aligned.bin

bigger variants (tiny / standard / xs at 96.85 / 98.25 / 99.07 %) are not in this release. those are the licensed ones, ping bauratynov@gmail.com if you need them for verification, no NDA.

Assets 4
Loading

v1.0.0 — Initial Release

26 Apr 02:03
@bauratynov bauratynov
v1.0.0
edad256

Choose a tag to compare

View all tags
v1.0.0 — Initial Release

First public release of FaceX.

Engine: 3.0ms median CPU inference, 148 KB library, zero dependencies.
WASM: 7ms in browser, 44 KB module.

Downloads

  • edgeface_xs_fp32.bin — Model weights (7 MB). Place in data/ folder.

Quick start

git clone https://github.com/facex-engine/facex
cd facex
./download_weights.sh
make
make test
Loading