mo666666
commented
10 months ago Hello, Gatsby666! Thank you for your interest in our work!
Our experiments are actually performed on CIFAR-10 dataset instead of CIFAR-100. I think the worse robustness may attribute to the larger number of class for CIFAR-100 than those of CIFAR-10. It increases the difficulty of ViTs to defend adversarial attacks. Similar observations are observed on CNNs. You can refer to robustbench for more results.
Hope my responses can address your concerns!
Gatsby666
hello~ I want to reproduce your code on the CIFAR-100 dataset, using vanilla adversarial defense methods on vit_base_patch16_224 with pre-training, to achieve similar effects in the paper.
When I modified the data_loader, imported the pre-trained model and ran it successfully. However, the robustness only reached about 20%. I want to know if this is normal or is it an issue with some hyperparameter settings.
my settings: Namespace(model='vit_base_patch16_224', method='AT', dataset='cifa100', run_dummy=False, accum_steps=1, grad_clip=1.0, test=False, log_interval=10, batch_size=64, AA_batch=128, crop=32, resize=32, load=False, load_path='./checkpoint', scratch=False, n_w=10, attack_iters=10, patch=16, ARD=False, PRM=False, drop_rate=1.0, beta=6.0, eval_restarts=1, eval_iters=10, data_dir='./data', epochs=40, lr_min=0.0, lr_max=0.1, weight_decay=0.0001, momentum=0.9, epsilon=8, labelsmoothvalue=0, alpha=2, delta_init='random', out_dir='./logs/vit-B-16-v2-32_AT', seed=0, mixup=0.8, cutmix=1.0, cutmix_minmax=None, mixup_prob=0.3, mixup_switch_prob=0.5, mixup_mode='batch', output_dir='output')