DMLP

Source code for our CVPR paper Learning from Noisy Labels with Decoupled Meta Label Purifier

Learning from Noisy Labels with Decoupled Meta Label Purifier (CVPR 2023)

This is the pytorch implementation of the paper (accepted by CVPR 2023).

<img src='DMLP.png'>

Fig 1.DMLP framework

Training

First download the features and corresponding noisy labels in labels to the folder labels_cifar10.

Then purify the noisy labels of CIFAR-10 with symmetric 20% noise rate:

python DMLP.py

Or you can directly use the purified labels of CIFAR-10, the corresponding files are in labels.

To achieve better performance, you can combine the purified labels with existing methods to boost their accuracy (i.e., C2D and DivideMix).

Performance

<img src='Clothing1m.png'>

Training Logs

Training logs on CIFAR-10/100 of DMLP+DivideMix across all the noise rates can be found in link.

TODO

[ ] Release the pretrained weights of MoCo on Clothing1M.

[ ] Release the purified labels on CIFAR-100 and Clothing1M.

Videos

For the introduction of the paper, you can refer to bilibili or youtube for more details.

Citation

If you find DMLP useful in your research, please consider citing.

@inproceedings{tu2023learning,
  title={Learning from noisy labels with decoupled meta label purifier},
  author={Tu, Yuanpeng and Zhang, Boshen and Li, Yuxi and Liu, Liang and Li, Jian and Wang, Yabiao and Wang, Chengjie and Zhao, Cai Rong},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={19934--19943},
  year={2023}
}

Reference

Pretrained-weights of SimCLR (CIFAR-10/CIFAR-100) and MoCo (Clothing1M) for the feature extraction can be referred in Simclr and MoCo.

For C2D and DivideMix, you can refer to C2D and DivideMix and combine them with our purified labels. Thanks for their great work!