RAPQ: Rescuing Accuracy for Power-of-Two Low-bit Post-training Quantization

Authors: Hongyi Yao, Pu Li, Jian Cao, Xiangcheng Liu, Chenying Xie, Bingzhang Wang

IJCAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on Image Net prove the excellent performance of our proposed method.
Researcher Affiliation Academia Hongyi Yao , Pu Li , Jian Cao , Xiangcheng Liu , Chenying Xie and Bingzhang Wang Peking University yhy@stu.pku.edu.cn, spurslipu@pku.edu.cn, caojian@ss.pku.edu.cn, liuxiangcheng@stu.pku.edu.cn, 402600293@qq.com, 13919334117@163.com
Pseudocode Yes Algorithm 1 : RAPQ Power-of-Two Quantization
Open Source Code Yes The code1 was released. 1https://github.com/Bill Amihom/RAPQ
Open Datasets Yes Extensive experiments based on Image Net [Russakovsky et al., 2015] dataset to demonstrate the superiority of our method.
Dataset Splits No The paper mentions 'Image Net' and '1024 images for PTQ calibration' but does not provide specific train/validation/test dataset splits (percentages, counts, or predefined citations) for reproducibility.
Hardware Specification Yes This scheme takes only 10 minutes to quantize Resnet18 with Intel i9-10980XE + Nvidia RTX3090.
Software Dependencies No The paper does not provide specific version numbers for key software components or libraries (e.g., Python, PyTorch, TensorFlow, CUDA) used in their implementation beyond general references to concepts like backpropagation.
Experiment Setup Yes We randomly pick a total of 1024 images for PTQ calibration in each experiment. To be fair, we optimized each model with 80,000 weight iterations and 5000 activation iterations in order to fully converge. At this point, the parameter α corresponding to Equation (19) is set to 0.9 and β is set to 1. ... Quick Mode with only 20,000 weight iterations and 1,000 activation iterations. Correspondingly, the parameter α in Equation (19) is set to 0.1 and β is set to 1.