Home

Awesome

SANM

Source code for our CVPR paper Learning with Noisy labels via Self-supervised Adversarial Noisy Masking

Learning with Noisy labels via Self-supervised Adversarial Noisy Masking (CVPR 2023)

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

<img src='SANM.png'>

Fig 1.SANM framework

Training

First you need to install dependencies by running pip install -r requirements.txt.

Then, please create a folder named <i>checkpoint</i> to store the results.

mkdir checkpoint

Next, run

python Train_{dataset_name}.py --data_path <i>path-to-your-data</i>

Performance

<img src='Clothing1m.png'>

Videos

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

Citation

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

@inproceedings{tu2023learning,
  title={Learning with Noisy labels via Self-supervised Adversarial Noisy Masking},
  author={Tu, Yuanpeng and Zhang, Boshen and Li, Yuxi and Liu, Liang and Li, Jian and Zhang, Jiangning and Wang, Yabiao and Wang, Chengjie and Zhao, Cai Rong},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={16186--16195},
  year={2023}
}

Reference

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