Thanks for your work, it helps me a lot. I notice the delta in your algorithm is gradually accumulate. but that in 《Towards Deep Learning Models Resistant to Adversarial》 is randomly reset. Is there any difference between these two methods ?
Hi, @ashafahi , thanks for your implementation, when evaluating your model during training, it seems like you did not switch off the batch norm, will that cause some difference in the validation results?
Could you share your pre-trained models? Of course, we can train them from scratch, but it will be largely convenient if pre-trained models can be provided. Thanks in advance.
I would really appreciate if you would share the trained robust cifar10 model. Despite the adversarial training is "free", it still requires 10+ GPU hours.