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.
Solving Oscillation Problem in Post-Training Quantization Through a Theoretical Perspective (CVPR 2023) (Link)
Overview
Prerequisites
Getting Started
Requirements
Data preparation
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.
Results
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
@articlema2023solving,title=SolvingOscillationProbleminPost−TrainingQuantizationThroughaTheoreticalPerspective,author=YuexiaoMaandHuixiaLiandXiawuZhengandXuefengXiaoandRuiWangandShileiWenandXinPanandFeiChaoandRongrongJi,journal=arXivpreprintarXiv:2303.11906,year=2023