wzekai99/dm-improves-at

Code for the paper "Better Diffusion Models Further Improve Adversarial Training" (ICML 2023)

dm-improves-at's Issues

Number of samples per epoch?

I just wanted to confirm my understanding because I didn't see this in the paper: for each generated dataset (e.g. 1m, 10m, 50m, etc), you don't actually iterate over the full augmented datasets in each epoch, right? You're simply sampling from the true and generated datasets such that each epoch corresponds to the same number of training steps for the original dataset (e.g. for CIFAR10, you'd only see 50k training examples), albeit different samples each epoch due to the random sampling from the real and generated datasets.

i.e. here:

https://github.com/wzekai99/DM-Improves-AT/blob/main/core/data/semisup.py#L15

about the open checkpionts accuracy did not inconsistent with yours.

I have tried to evaluate your wrn-76-16 on cifar-10 under linf norm, but the clean accuracy is 73.71%, not 93% as you describe, how did it comes?