Asymmetric Loss Functions for Learning with Noisy Labels

This repository is the official implementation of Asymmetric Loss Functions for Learning with Noisy Labels [ICML 2021] and Asymmetric Loss Functions for Noise-tolerant Learning: Theory and Applications [T-PAMI].

Requirements

Python >= 3.6, PyTorch >= 1.3.1, torchvision >= 0.4.1, numpy>=1.11.2, tqdm >= 4.50.2, seaborn >= 0.11.0, tensorboardX >= 2.5

Learning with Noisy Labels (LNL)

The main running file is main.py with arguments as follows:

• noise_type: symmetric | asymmetric
• noise_rate: noise rate
• loss: AGCE | AUL | AEL | CE (Cross Entropy) | FL (Focal Loss) | MAE | GCE | SCE | NFL | NCE | ...

The detailed implementation about the proposed asymmetric losses for classification can be found in ./lnl/losses.py

Example for 0.4 Symmetric noise rate with AUL loss

# CIFAR-10
$  python3  main.py --noise_type      symmetric           \
                    --noise_rate      0.4                 \
                    --loss            AUL                 \

Self-supervised Image Denoising

The main running file is main.py with arguments as follows:

• exp: n2c | n2n | n2s
• style: gauss | bernoulli | saltpepper | impulse
• loss: heat | poisson | lp | mse ...

The detailed implementation about the proposed asymmetric losses for regression can be found in ./denoising/losses.py

Example for using the negative heat kernel loss for Gaussian denoising with noise2self

$  python3  main.py --exp       n2s               \
                    --style     gauss15           \
                    --loss      heat0.1           \

Reference

For technical details and full experimental results, please check the paper. If you have used our work in your own, please consider citing:

@InProceedings{zhou2021asymmetric,
  title = 	 {Asymmetric Loss Functions for Learning with Noisy Labels},
  author =       {Zhou, Xiong and Liu, Xianming and Jiang, Junjun and Gao, Xin and Ji, Xiangyang},
  booktitle = 	 {Proceedings of the 38th International Conference on Machine Learning},
  pages = 	 {12846--12856},
  year = 	 {2021},
  editor = 	 {Meila, Marina and Zhang, Tong},
  volume = 	 {139},
  series = 	 {Proceedings of Machine Learning Research},
  month = 	 {18--24 Jul},
  publisher =    {PMLR}
}

@ARTICLE{10039708,
  author={Zhou, Xiong and Liu, Xianming and Zhai, Deming and Jiang, Junjun and Ji, Xiangyang},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={Asymmetric Loss Functions for Noise-Tolerant Learning: Theory and Applications}, 
  year={2023},
  volume={},
  number={},
  pages={1-16},
  doi={10.1109/TPAMI.2023.3236459}
}

Moreover, we thank the code implemented by Ma et al. (classification) and Zhang et al. (DnCNN).