/Learning-to-Purify-Noisy-Labels-via-Meta-Soft-Label-Corrector

[AAAI 21] Utilizing meta-learning to correct the noisy labels.

Primary LanguagePythonMIT LicenseMIT

Learning-to-Purify-Noisy-Labels-via-Meta-Soft-Label-Corrector

AAAI'21: Learning to Purify Noisy Labels via Meta Soft Label Corrector (Official Pytorch implementation for noisy labels).

This is the code for the paper: Learning to Purify Noisy Labels via Meta Soft Label Corrector
Yichen Wu, Jun Shu, Qi Xie, Qian Zhao, Deyu Meng* To be presented at AAAI 2021.

If you find this code useful in your research then please cite

@article{wu2020learning,
  title={Learning to Purify Noisy Labels via Meta Soft Label Corrector},
  author={Wu, Yichen and Shu, Jun and Xie, Qi and Zhao, Qian and Meng, Deyu},
  journal={arXiv preprint arXiv:2008.00627},
  year={2020}
}

Setups

The requiring environment is as bellow:

  • Linux
  • Python 3+
  • PyTorch 0.4.0
  • Torchvision 0.2.0

Running our method on benchmark datasets (CIFAR-10 and CIFAR-100).

Here is an example:

python main.py --dataset cifar10 --corruption_type unif(flip2) --corruption_prob 0.6

The default network structure is Resnet34

Acknowledgements

We thank the Pytorch implementation on glc(https://github.com/mmazeika/glc) and learning-to-reweight-examples(https://github.com/danieltan07/learning-to-reweight-examples).

Contact: Yichen Wu (wuyichen.am97@gmail.com); Deyu Meng(dymeng@mail.xjtu.edu.cn).