Omnipotent Adversarial Training in the Wild [pdf]
Introduction: This paper introduced a new practical challenge in adversarial training, i.e., how to train a robust model on a label-noisy and imbalanced dataset. Our proposed OAT can effectively address such a problem. Notice: This paper is not related to using adversarial training to solve label noise and long-tail problems. It is actually a new research area.
- pytorch == 1.12.0
- torchvision
- numpy
- tqdm
- PIL
python train.py --arch resnet --dataset [cifar10, cifar100] --imb [imbalanced ratio] --nr [noise ratio] --noise_type [sym, asy] --save [the name you want to save your model] --exp [experiment name]