AUST-Hansen / cls_KD

'NKD and USKD' (ICCV 2023) and 'ViTKD'

Knowledge Distillation for Image Classification

This repository includes official implementation for the following papers:

• ICCV 2023: NKD and USKD: From Knowledge Distillation to Self-Knowledge Distillation: A Unified Approach with Normalized Loss and Customized Soft Labels

• ViTKD: ViTKD: Practical Guidelines for ViT feature knowledge distillation

It also provides unofficial implementation for the following papers:

• KD, DKD, WSLD
• MGD, SRRL

If this repository is helpful, please give us a star ⭐ and cite relevant papers.

Install

• Prepare the dataset in data/imagenet

• # Set environment
  pip install torch==1.10.1+cu111 torchvision==0.11.2+cu111 torchaudio==0.10.1 -f https://download.pytorch.org/whl/cu111/torch_stable.html
  pip install -r requirements.txt
  

• This repo uses mmcls = 1.0.0rc6. If you want to use lower mmcls version for distillation, you can refer branch master to change the codes.

Run

• Please refer nkd.md and vitkd.md to train the student and get the weight.
• You can modify the configs to choose different distillation methods and pairs.
• The implementation details of different methods can be seen in the folder dis_losses.

Citing NKD and USKD

@article{yang2023knowledge,
  title={From Knowledge Distillation to Self-Knowledge Distillation: A Unified Approach with Normalized Loss and Customized Soft Labels},
  author={Yang, Zhendong and Zeng, Ailing and Li, Zhe and Zhang, Tianke and Yuan, Chun and Li, Yu},
  journal={arXiv preprint arXiv:2303.13005},
  year={2023}
}

Citing ViTKD

@article{yang2022vitkd,
  title={ViTKD: Practical Guidelines for ViT feature knowledge distillation},
  author={Yang, Zhendong and Li, Zhe and Zeng, Ailing and Li, Zexian and Yuan, Chun and Li, Yu},
  journal={arXiv preprint arXiv:2209.02432},
  year={2022}
}

Acknowledgement

Our code is based on the project MMPretrain.