RUPQ: Improving low-bit quantization by equalizing relative updates of quantization parameters
Valentin Buchnev (Huawei Technologies Co. Ltd.),* Jiao He (huawei company), Fengyu Sun (Huawei), Ivan Koryakovskiy (Huawei Technologies Co., Ltd.)
The 34th British Machine Vision Conference
Abstract
Neural network quantization is a model compression technique using low-bit width representation of floating-point weights and activations. Although quantization can significantly reduce power consumption and inference time, it often leads to worse target metrics, such as accuracy or peak signal-to-noise ratio. To improve upon the target metrics, particularly at low-bit width computations, we propose a new Relative Update-Preserving Quantizer. This quantizer stabilizes relative updates for all parameters during training. Our experimental results show that the proposed quantizer performs consistently better than LSQ+ and, compared to full-precision models, significantly reduces the gap in quality for low-bit SRResNet, EDSR and YOLO-v3 models. The method is easy to implement and use in practice.Video
Citation
@inproceedings{Buchnev_2023_BMVC,
author = {Valentin Buchnev and Jiao He and Fengyu Sun and Ivan Koryakovskiy},
title = {RUPQ: Improving low-bit quantization by equalizing relative updates of quantization parameters},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0307.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by