Solving Oscillation Problem in Post-Training Quantization Through a Theoretical Perspective (CVPR 2023) (Link)
Overview
Prerequisites
- Linux
- Python 3
- CPU or NVIDIA GPU + CUDA CuDNN
Getting Started
Requirements
- MQBench
- pyyaml
- easydict
Data preparation
- ImageNet
Add the imagenet path to the "data.path" in the config file.
Run MRECG
-
Run the following command to quantize the model using the MRECG algorithm, all yaml files can be found in ./config. You can change the bit width, batchsize, pre-trained model path and other quantization parameters in the yaml file.
python ptq_main.py --config configs/qdrop/mbv2_2_4.yaml
Results
| Method\Model | W/A | Res18 | Res50 | MBV2×1.0 | MBV2×0.75 | MBV2×0.5 | MBV2×0.35 | Reg600M |
|---|---|---|---|---|---|---|---|---|
| Full Prec. | 32/32 | 71.01 | 76.63 | 72.20 | 69.95 | 64.60 | 60.08 | 73.52 |
| Pretrained | - | link | link | link | link | link | link | link |
| Ours+BRECQ | 4/4 | 69.06 | 74.84 | 68.56 | 64.55 | 55.26 | 50.67 | - |
| Ours+BRECQ | 2/4 | 65.61 | 70.04 | 58.49 | 52.50 | 41.16 | 35.46 | - |
| Ours+BRECQ | 3/3 | 65.64 | 70.68 | 57.14 | 50.21 | 35.11 | 30.26 | - |
| Ours+BRECQ | 2/2 | 52.02 | 43.72 | 13.84 | 9.46 | 3.43 | 3.22 | - |
| Ours+QDROP | 4/4 | 69.46 | 75.35 | 68.84 | 64.39 | 55.64 | 50.94 | 71.22 |
| Ours+QDROP | 2/4 | 66.18 | 70.53 | 57.85 | 53.71 | 40.09 | 35.85 | 65.16 |
| Ours+QDROP | 3/3 | 66.30 | 71.92 | 58.40 | 51.78 | 38.43 | 32.96 | 66.08 |
| Ours+QDROP | 2/2 | 54.46 | 56.82 | 14.44 | 11.40 | 4.18 | 3.09 | 43.67 |
Due to the presence of random numbers in the experiment, the actual model accuracy may be slightly high or low.
Acknowledgements
Our code relies on the MQBench package.
Reference
@article{ma2023solving,
title={Solving Oscillation Problem in Post-Training Quantization Through a Theoretical Perspective},
author={Yuexiao Ma and Huixia Li and Xiawu Zheng and Xuefeng Xiao and Rui Wang and Shilei Wen and Xin Pan and Fei Chao and Rongrong Ji},
journal={arXiv preprint arXiv:2303.11906},
year={2023}
}