[CVPR 2023] NIPQ: Noise Proxy-Based Integrated Pseudo-Quantization [Paper] [Slides] [Video]

This is official code for the paper NIPQ: Noise Proxy-Based Integrated Pseudo-Quantization. NIPQ samples accurate pseudo-quantization noise(PQN), while integrating learnable truncation parameter in pseudo-quantization training(PQT) pipeline. Bit-width is automatically allocated considering layer-wise sensitivity.

How to use

You can reproduce the result of MobileNet-v2 for avgbit and bops options

• Optimize average bit-width

sh run_mv2_imagenet_ts_avgbit.sh

• Optimize bops

sh run_mv2_imagenet_ts_bops.sh

For experimental details and hyper-paramters, please refer to the paper and train.py file

Installation

• Python verseion >= 3.7.13
• Pytorch >= 1.12.1
• ImageNet Dataset (Set data option of train.py)
• Using docker:

docker run -v {local_code_loc}:{container_code_loc} -v {local_dataset_loc}:{container_dataset_loc} -it --gpus=all pytorch/pytorch:1.12.1-cuda11.3-cudnn8-devel 

Citation

@InProceedings{Shin_2023_CVPR,
    author    = {Shin, Juncheol and So, Junhyuk and Park, Sein and Kang, Seungyeop and Yoo, Sungjoo and Park, Eunhyeok},
    title     = {NIPQ: Noise Proxy-Based Integrated Pseudo-Quantization},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {3852-3861}
}