Skip to content

Latest commit

 

History

History
37 lines (25 loc) · 1.39 KB

README.md

File metadata and controls

37 lines (25 loc) · 1.39 KB
Raw

RobustMW-Net

WACV'21: Do We Really Need Gold Samples for Sample Weighting Under Label Noise?

This is the official code for the paper: Do We Really Need Gold Samples for Sample Weighting Under Label Noise?
Aritra Ghosh, and Andrew Lan Presented at WACV 2021.

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

@InProceedings{Ghosh_2021_WACV,
    author    = {Ghosh, Aritra and Lan, Andrew},
    title     = {Do We Really Need Gold Samples for Sample Weighting Under Label Noise?},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2021},
    pages     = {3922-3931}
}

Running RobustMW-Net on benchmark datasets.

To run on cifar10 (cifar100) dataset with uniform (flip2/flip) noise with noise rate 0.4 and noisy (clean) meta dataset with robust loss (or CE loss like MWnet), run

python trainer.py --dataset cifar10 (cifar100) --corruption_type unif (flip2/flip) --corruption_prob 0.4 --noisy 1 (0) --meta_loss mae (cross)

Acknowledgements

We thank the Pytorch implementation on MWNet(https://github.com/xjtushujun/meta-weight-net).

Contact: Aritra Ghosh (aritraghosh.iem@gmail.com).