OmniShift

Overview

OmniShift is a framework that converts any supported CNN backbone into a multiply-free network by applying four independently toggleable quantization techniques:

Component	Description	Effect
Sparse Shift	W in {0, +/-2^p}	Conv multiplications -> bit-shifts + skip-zero
PoT-BN	gamma/sigma -> +/-2^q	BN scale multiplication -> shift
PoT-Act	Post-ReLU -> {0} U {2^p}	Activation quantization to log-uniform grid
EWGS	Element-Wise Gradient Scaling	Replaces STE backward for smoother training

Energy model (45nm CMOS): mul = 3.7 pJ, add = 0.9 pJ, shift = 0.13 pJ

Quick Start

pip install -r requirements.txt

cd OmniShift

# Sanity check - all 8 methods
python3 -c "
from src.models.resnet_cifar import build_model
from src.utils.ops_counter import count_mul_add_shift
from src.quantize.pot_bn import set_bn_epoch
import torch

for method in ['fp32', 'deepshift', 'apot', 'denseshift', 's3shift', 'fogzo', 'aptq', 'omnishift']:
    m = build_model('resnet20', method, num_classes=10)
    set_bn_epoch(m, 999)
    out = m(torch.randn(2, 3, 32, 32))
    ops = count_mul_add_shift(m)
    print(f'{method:12s}  shape={out.shape}  energy={ops[\"energy_GpJ\"]:.4f} GpJ')
"

# Run experiment
python scripts/run_experiment.py --config configs/omnishift.yaml
python scripts/run_experiment.py --config configs/omnishift.yaml --method deepshift --dataset svhn

# Print results table
python scripts/summarize_results.py

# Plot training curves -> saves 12 images to assets/
python scripts/plot.py

Supported Backbones & Datasets

Backbones: resnet20, resnet56

Datasets: cifar10, svhn, stl10

Baselines

Method	Paper	Authors	ArXiv	Venue
`fp32`	-	-	-	-
`deepshift`	DeepShift: Towards Multiplication-Less Neural Networks	Elhoushi et al.	1905.13298	CVPR-W 2021
`apot`	Additive Power-of-Two Quantization	Li et al.	1909.13144	ICLR 2020
`denseshift`	DenseShift: Towards Accurate and Efficient Low-Bit Power-of-Two Quantization	Li et al.	2208.09708	ICCV 2023
`s3shift`	S3: Sign-Sparse-Shift Reparametrization for Effective Training of Low-Bit Shift Networks	Li et al.	2107.03453	NeurIPS 2021
`fogzo`	FOGZO: First-Order-Guided Zeroth-Order Gradient Descent for Quantization-Aware Training	Yang & Aamodt	2510.23926	NeurIPS 2025
`aptq`	APTQ: Adaptive Global Power-of-Two Ternary Quantization	Liu et al.	-	Sensors (MDPI) 2024
`omnishift`	OmniShift (this work)	-	-	-

apot uses a single-term PoT grid (step = alpha / 2^(n_bits-1)) rather than the original additive multi-term construction. aptq has no public arXiv preprint; DOI: 10.3390/s24010181.

Project Structure

OmniShift/
├── src/
|   ├── quantize/
|   |   ├── shift.py           # ShiftConv2d - W in {+/-2^p} (DeepShift-PS)
|   |   ├── sparse_shift.py    # SparseShiftConv2d - W in {0, +/-2^p}
|   |   ├── s3shift.py         # S3ShiftConv2d - sign x sparse x shift
|   |   ├── fogzo.py           # FogzoShiftConv2d - ZO-augmented STE
|   |   ├── aptq_ternary.py    # APTQTernaryConv2d - two-sub-filter PoT ternary
|   |   ├── apot.py            # APoTConv2d - additive PoT grid
|   |   ├── denseshift.py      # DenseShiftConv2d - sign x shift, no zero
|   |   ├── pot_bn.py          # PoTBatchNorm2d, set_bn_epoch
|   |   ├── pot_act.py         # PoTActivation
|   |   └── ewgs.py            # EWGS variants of all quantizers
|   |
|   ├── methods/
|   |   ├── __init__.py        # get_factories(), METHODS list
|   |   ├── fp32.py            # FP32 baseline
|   |   ├── deepshift.py       # DeepShift-PS
|   |   ├── apot.py            # APoT
|   |   ├── denseshift.py      # DenseShift
|   |   ├── s3shift.py         # S3
|   |   ├── fogzo.py           # FOGZO
|   |   ├── aptq_ternary.py    # APTQ
|   |   └── omnishift.py       # OmniShift full pipeline
|   | 
|   ├── models/
|   |   └── resnet_cifar.py    # ResNetCIFAR, build_model()
|   |
|   ├── data/
|   |   └── loaders.py         # get_dataloaders (cifar10, svhn, stl10)
|   |
|   ├── training/
|   |   ├── train.py           # train_one_epoch, evaluate, EarlyStopping
|   |   ├── scheduler.py       # cosine_lr_schedule
|   |   └── regularize.py      # L1 sparsity regularization
|   |
|   └── utils/
|       ├── ops_counter.py     # hook-based backbone-agnostic op counter
|       ├── seed.py            # set_seed, clear_memory
|       └── checkpoint.py      # save_checkpoint, save_log
|   
├── configs/
|   └── omnishift.yaml         # unified config (edit method/backbone/dataset)
|   
├── scripts/
|   ├── run_experiment.py      # training entry point (CLI + Python API)
|   ├── summarize_results.py   # print results table from logs/
|   ├── update_readme.py       # auto-update Results section from logs/
|   ├── plot.py                # generate training curve plots -> assets/
|   ├── fpga_estimate.py       # Xilinx 7-series resource estimation
|   └── trt_benchmark.py       # TensorRT FP16 benchmark
|   
└── notebooks/
    └── omnishift.ipynb        # Kaggle notebook (setup / config / train / results)

Configuration

Edit configs/omnishift.yaml or pass --method/--dataset flags:

experiment:
  method:   "omnishift"   # fp32 | deepshift | apot | denseshift | s3shift | fogzo | aptq | omnishift
  backbone: "resnet20"    # resnet20 | resnet56
  dataset:  "cifar10"     # cifar10 | svhn | stl10
  seed:     42

training:
  epochs:          200
  batch_size:      256
  lr:              0.1      # cosine decay
  sparsity_lambda: 0.0001   # L1 regularization (learnable sparse mode)

Method-specific opts via Python API:

from scripts.run_experiment import build_cfg, run

cfg = build_cfg('omnishift', 'resnet20', 'cifar10', seed=42, epochs=200,
                sparse_mode='learnable', bn_warmup=30)
result = run(cfg)

Outputs

Each run saves two files under backbone-specific subdirectories:

checkpoints/{backbone}/{run_name}_{dataset}_seed{seed}.pt   # best weights + metadata
logs/{backbone}/{run_name}_{dataset}_seed{seed}.json        # per-epoch loss/acc log

Results

ResNet20

CIFAR-10

Method	Test Acc	Sparsity	Energy (GpJ)	vs FP32
omnishift	83.01%	89.03%	0.0068	27.6x
aptq	91.35%	74.90%	0.0134	14.1x
apot	91.27%	0.00%	0.0445	4.2x
deepshift	92.23%	0.00%	0.0445	4.2x
denseshift	90.56%	0.00%	0.0445	4.2x
fogzo	90.72%	0.00%	0.0445	4.2x
s3shift	90.58%	0.00%	0.0445	4.2x
fp32 (baseline)	92.17%	0.00%	0.1887	1.0x

SVHN

Method	Test Acc	Sparsity	Energy (GpJ)	vs FP32
omnishift	94.78%	93.43%	0.0050	37.7x
aptq	96.50%	84.61%	0.0094	20.1x
apot	96.54%	0.00%	0.0445	4.2x
deepshift	96.73%	0.00%	0.0445	4.2x
denseshift	96.47%	0.00%	0.0445	4.2x
fogzo	95.98%	0.00%	0.0445	4.2x
s3shift	96.36%	0.00%	0.0445	4.2x
fp32 (baseline)	96.82%	0.00%	0.1887	1.0x

STL-10

Method	Test Acc	Sparsity	Energy (GpJ)	vs FP32
omnishift	65.92%	64.29%	0.0171	11.0x
aptq	65.50%	51.17%	0.0233	8.1x
apot	66.88%	0.00%	0.0445	4.2x
deepshift	67.54%	0.00%	0.0445	4.2x
denseshift	68.46%	0.00%	0.0445	4.2x
fogzo	59.03%	0.00%	0.0445	4.2x
s3shift	68.11%	0.00%	0.0445	4.2x
fp32 (baseline)	67.62%	0.00%	0.1887	1.0x

ResNet56

CIFAR-10

Method	Test Acc	Sparsity	Energy (GpJ)	vs FP32
omnishift	83.59%	96.89%	0.0067	87.0x
aptq	91.84%	91.83%	0.0151	38.5x
apot	92.69%	0.00%	0.1336	4.3x
deepshift	93.80%	0.00%	0.1336	4.3x
denseshift	92.53%	0.00%	0.1336	4.3x
fogzo	92.50%	0.00%	0.1336	4.3x
s3shift	92.40%	0.00%	0.1336	4.3x
fp32 (baseline)	93.67%	0.00%	0.5809	1.0x

SVHN

Method	Test Acc	Sparsity	Energy (GpJ)	vs FP32
omnishift	95.03%	97.97%	0.0053	109.7x
aptq	96.38%	96.16%	0.0095	61.1x
apot	96.68%	0.00%	0.1336	4.3x
deepshift	96.73%	0.00%	0.1336	4.3x
denseshift	96.69%	0.00%	0.1336	4.3x
fogzo	96.34%	0.00%	0.1336	4.3x
s3shift	96.74%	0.00%	0.1336	4.3x
fp32 (baseline)	96.75%	0.00%	0.5809	1.0x

STL-10

Method	Test Acc	Sparsity	Energy (GpJ)	vs FP32
omnishift	60.62%	73.77%	0.0365	15.9x
aptq	63.19%	64.40%	0.0505	11.5x
apot	65.49%	0.00%	0.1336	4.3x
deepshift	67.29%	0.00%	0.1336	4.3x
denseshift	67.27%	0.00%	0.1336	4.3x
fogzo	56.23%	0.00%	0.1336	4.3x
s3shift	67.22%	0.00%	0.1336	4.3x
fp32 (baseline)	66.95%	0.00%	0.5809	1.0x

Training Curves

All 8 methods per chart. Dashed lines = train, solid lines = val.

Run python scripts/plot.py to regenerate these images from logs/.

ResNet-20

Dataset	Loss	Accuracy
CIFAR-10
SVHN
STL-10

ResNet-56

Dataset	Loss	Accuracy
CIFAR-10
SVHN
STL-10

Hyperparameters

Param	Default
Epochs	200
Batch size	256
LR	0.1 (cosine decay)
Momentum	0.9
Weight decay	5e-4
Sparsity lambda	1e-4 (learnable mode)
BN warmup	30 epochs
EWGS lambda	0.02
PoT-Act levels	8

Val split: 10% of train, torch.Generator(seed=42).

References

DeepShift - Elhoushi et al., CVPR-W 2021
APoT - Li et al., ICLR 2020
DenseShift - Li et al., ICCV 2023
S3 - Li et al., NeurIPS 2021
EWGS - Lee et al., CVPR 2021
FOGZO - Yang & Aamodt, NeurIPS 2025
APTQ - Liu et al., Sensors 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

OmniShift

Overview

Quick Start

Supported Backbones & Datasets

Baselines

Project Structure

Configuration

Outputs

Results

ResNet20

CIFAR-10

SVHN

STL-10

ResNet56

CIFAR-10

SVHN

STL-10

Training Curves

ResNet-20

ResNet-56

Hyperparameters

References

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Name		Name	Last commit message	Last commit date
Latest commit History 15 Commits
assets		assets
checkpoints		checkpoints
configs		configs
logs		logs
notebooks		notebooks
scripts		scripts
src		src
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
requirements.txt		requirements.txt

Folders and files

Latest commit

History

Repository files navigation

OmniShift

Overview

Quick Start

Supported Backbones & Datasets

Baselines

Project Structure

Configuration

Outputs

Results

ResNet20

CIFAR-10

SVHN

STL-10

ResNet56

CIFAR-10

SVHN

STL-10

Training Curves

ResNet-20

ResNet-56

Hyperparameters

References

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages