Contrastive Consistent Representation Distillation

This repository is PyTorch code for our proposed CoCoRD.

The illustration of our proposed CoCoRD. The task-specific loss (e.g. classification loss) is omitted for clear presentation. The details about our proposed CoCoRD can be found in our main paper.

Contents

1. Requirements
2. Train
3. Results
4. Citation
5. Acknowledgements

Requirements

• Python 3 (Tested with Python 3.10)
• PyTorch (Tested with PyTorch 2.0, should be runnable with PyTorch 1.x)
• tensorboard (pip install tensorboard)
• PyYAML (pip install PyYAML)

Train

1. Fetch the pretrained teacher models provided by CRD.

2. Modify ./options/train_student.yml for your needs. The hyperparameters are explained in the yaml file.

3. Run CoCoRD distillation with the following command:

   python main.py -opt ./options/train_student.yml -t 1
   

   where the flags are explained as:

   • -opt: specify the path to the training yaml file.
   • -t: specify the experimental id to differentiate between multiple runs.

4. You can monitor the distillation process in ./experiments.

Note: this code is for a single-gpu training on CIFAR-100 dataset. The dataset will be automatically downloaded.

Results

The performance is measured by Top-1 classification accuracy (%).

The red arrow denotes underperforming KD. The green arrow denotes outperforming KD.

1. Teacher and student are of the same architectural style.

   

2. Teacher and student are of the different architectural style.

   

Citation

If you find our work useful for your research, please consider citing the paper

@inproceedings{CoCoRD_Fu_bmvc23,
  author       = {Shipeng Fu and Haoran Yang and Xiaomin Yang},
  title        = {Contrastive Consistent Representation Distillation},
  booktitle    = {34th British Machine Vision Conference 2023, {BMVC} 2023},
  pages        = {300--303},
  publisher    = {{BMVA} Press},
  year         = {2023},
}

For any questions, please contact Shipeng Fu ([email protected]).

Acknowledgements

• Thanks to Haoran Yang for helping develop this repository.
• Thanks to the authors of RepDistiller. They provide many useful codes which facilitate our work.
• Thanks also go to authors who make their code publicly available.