CRCKD

This repository is an official PyTorch implementation of the paper "Categorical Relation-Preserving Contrastive Knowledge Distillation for Skin Lesion Classification" paper from MICCAI 2021.

Installation

Dependencies

• Python 3.6
• PyTorch >= 1.5.0
• numpy
• pandas
• scipy
• sklearn
• tensorboardX
• torchvision

Usage

Baseline 1: DenseNet 121

python3 train_CRCKD.py --exp sup4_base

Baseline 2: DenseNet 121 + conventional Mean Teacher (MT)

python3 train_CRCKD.py --ema_consistency 1  --exp sup4_pred_MT

CRCKD: baseline + pred_MT + 0.1 * CRD_20pos + CRP

python3 train_CRCKD.py --ema_consistency 1 --CCD_distill 1 --CRD_b4_weight 0.1 --mode multi_pos --nce_p 20 --CRP_distill 1 --exp sup4_pred_MT_0.1CRD_20pos_CRP

Main Scripts

· cv_splits.py: Split all data into five folds for cross validation.

· train_CRCKD.py: Model training.

· validation.py: Evaluation of the model predictions.

· dataloader/dataset.py: Dataset with memory bank and contrastive samples.

· utils/memory.py: Construct memory banks that supply positive and negative samples.

· utils/CRD_CRP_loss.py: Definition of the CCD and CRP modules proposed in our method.

Citation

If you find the codes useful, please cite the following publication:

@article{xing2021categorical,
  title={Categorical Relation-Preserving Contrastive Knowledge Distillation for Medical Image Classification},
  author={Xing, Xiaohan and Hou, Yuenan and Li, Hang and Yuan, Yixuan and Li, Hongsheng and Meng, Max Q-H},
  journal={arXiv preprint arXiv:2107.03225},
  year={2021}
}

Contact

If you have any problems in the codes, please contact xhxing@link.cuhk.edu.hk.