Pytorch Implementation of ROCL . Based on the paper:
Adversarial Self-Supervised Contrastive Learning
Minseon Kim, Jihoon Tack, Sung Ju Hwang
arXiv:2006.07589
- Write the ROCL implementation
- Write the RELIC loss implementation
- Train and check if ROCL code eproduces the result: In progress
- Commit RELIC loss file
- [] Write a training script for training relic loss
Model name | Accuracy | robustness |
---|---|---|
Official | 83.71 % | 40.27% |
Ours | 71.91 % | 20.21% |
I only trained for 200 epochs instead of 1000 in the paper. Please run python train.py and resume training from the 200th epoch.