vill-lab/2023-icml-dpl

Pytorch implementation for DPL (ICML 2023).

2023-ICML-DPL

This is the pytorch implementation of the paper.

Getting start

python main.py [--dataset DATASET] [--data_dir DATA_DIR] 
               [--net NET] [--batch_size BATCH_SIZE] [--gpu GPU] 
               [--lr LR] [--epoch EPOCH] [--resume] 
               [--alpha ALPHA] [--G_size G_SIZE] [--varepsilon VAREPSILON] [--rep_aug rep_aug]

• batch size: 128
• learning rate: 0.1
• training epoch: 200
• the hyperparameter $\alpha$: 0.005
• the size of Guide Set: 5% of the size of training set
• the hyperparameter $\varepsilon$: 0.04
• the approaches of replacement and augmentation: augmentation

Citation

@inproceedings{song2023deep,
  title={Deep perturbation learning: enhancing the network performance via image perturbations},
  author={Song, Zifan and Gong, Xiao and Hu, Guosheng and Zhao, Cairong},
  booktitle={International Conference on Machine Learning},
  pages={32273--32287},
  year={2023},
  organization={PMLR}
}