The code for NoiLIn: Improving Adversarial Training and Correcting Stereotype of Noisy Labels (TMLR 22 accept)
- Python (3.8)
- PyTorch (1.11)
- CUDA (11.3)
- AutoAttack (This is a package for implementing AutoAttack. You can install it via
pip install git+https://github.com/fra31/auto-attack) - Numpy
(1) SAT-NoiLIn/TRADES-NoiLIn using ResNet18/Wide ResNet on CIFAR10/CIAFR100/SVHN
python SAT-NoiLIn.py --dataset='dataset' --noise_type='noise_type' --net='net_name'
python TRADES-NoiLIn.py --dataset='dataset' --noise_type='noise_type' --net='net_name'
(2) TRADES-AWP-NoiLIn using WRN-34-10 on CIFAR10
cd TRADES-AWP-NoiLIn
python TRADES-AWP-NoiLIn.py
(3) SAT-NoiLIn with extra training data using WRN-28-10 on CIFAR10
You need to download ti_500K_pseudo_labeled.pickle which contains 500K pseudo-labeled TinyImages from this link (Auxillary data provided by Carmon et al. 2019). Then, store ti_500K_pseudo_labeled.pickle into the folder ./data. In addition, we recommend to using mutilple GPUs.
cd NoiLIn_ExtraData
python SAT-NoiLIn-ExtraData.py --gpu='0,1,2,3' --aux_data_filename='ti_500K_pseudo_labeled.pickle'
(4) Evaluate the performance of trained models
(4.1) Obtain the learning curve of natural and robust accuracy
python eval.py --all_epoch --start_epoch=1 --end_epoch=120 --model_dir='model_dir'
(4.2) Obtain natural and robust accuracy of a given model
python eval.py --model_dir='model_dir' --pt_name='model_pt_name'
@article{
zhang2022noilin,
title={NoiLin: Improving adversarial training and correcting stereotype of noisy labels },
author={Jingfeng Zhang and Xilie Xu and Bo Han and Tongliang Liu and Lizhen Cui and Gang Niu and Masashi Sugiyama},
journal={Transactions on Machine Learning Research},
year={2022},
url={https://openreview.net/forum?id=zlQXV7xtZs},
}
Please contact jingfeng.zhang@riken.jp and xuxilie@comp.nus.edu.sg if you have any questions on the codes.