HTQ: Exploring the High-Dimensional Trade-Off of Mixed-Precision Quantization

This repository contains the official PyTorch implementation for the paper "HTQ: Exploring the High-Dimensional Trade-Off of Mixed-Precision Quantization".

This repository includes the core components of the algorithm, including the calculation of the sensitivity and the generation of the 3D Pareto frontier.

Below are instructions for reproducing the results.

Install

• To install HTQ and develop locally:

  git clone https://github.com/zkkli/HTQ.git
  cd HTQ

Reproduce the results

• You can reproduce the 3D Pareto frontier of ResNet18 on ImageNet using the following command:

python main.py --model resnet18 --dataset imagenet --dataset_path [/Path/to/ImageNet/] 

optional arguments:
--model: Model architecture.
--dataset: Dataset type.
--dataset_path: path to dataset.

Citation

We appreciate it if you would please cite the following paper if you found the implementation useful for your work:

@article{li2024htq,
  title={HTQ: Exploring the High-Dimensional Trade-Off of Mixed-Precision Quantization},
  author={Li, Zhikai and Long, Xianlei and Xiao, Junrui and Gu, Qingyi},
  journal={Pattern Recognition},
  year={2024},
  publisher={Elsevier}
}