franciumlee / LLSQ-Unofficial-

The PyTorch implementation of of Neural Networks for Low-precision Integer Hardware (LLSQ) in ICLR2020 (unofficial)

LLSQ-Unofficial-

The PyTorch implementation of Linear Symmetric Quantization of Neural Networks for Low-precision Integer Hardware (LLSQ) in ICLR2020 (unofficial)

I am working on reproducing this paper. This project is based on the code.

Experiment

----- VGG-small Cifar10 Accuracy -----

	This Project	Paper
fp32	88.47 %	93.34 %
w4a4	90.25 %	94.34 %

• fp32 training 400 epochs
• w4a4 training 100 epochs (Based on fp32)

Reference

@inproceedings{
Zhao2020Linear,
title={Linear Symmetric Quantization of Neural Networks for Low-precision Integer Hardware},
author={Xiandong Zhao and Ying Wang and Xuyi Cai and Cheng Liu and Lei Zhang},
booktitle={International Conference on Learning Representations},
year={2020},
url={https://openreview.net/forum?id=H1lBj2VFPS}
}