Gradient Pertub

This repo is the implementation of Alleviating Data Imbalance Issue with Perturbed Input During Inference paper

Method

Train

The training config is setting in the file of param.json

python main.py --model_dir ./experiments/example

Evaluation

After the model is well-trained, the model is stored in the experiments. When evaluate the model, we add gradient perturbation with small magnitube (e.g., 0.001).

python evaluate.py --model_dir ./experiments/example --epsilon 0.001

Citation

@inproceedings{chen2021alleviating,
  title={Alleviating Data Imbalance Issue with Perturbed Input During Inference},
  author={Chen, Kanghao and Mao, Yifan and Lu, Huijuan and Zeng, Chenghua and Wang, Ruixuan and Zheng, Wei-Shi},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={407--417},
  year={2021},
  organization={Springer}
}