About ResNet18 experiment on CIFAR 10

[ruili-pml]

Hi,

Thank you for your excellent work. I have two questions about the ResNet18 experiments on CIFAR10 and would be great if you could answer them.

Question 1: I was trying to reproduce your results in PyTorch and I found that the hyperparameters you provided in paper's appendix were not the same as the ones on github. In paper's appendix priorprec is set to 10 but in the repo it's set to 40?

Question 2: I'm not very familiar with Jax so I'm not sure how did you deal with BatchNorm's running statistics during training. As b-SAM have multiple forward pass within one update, I think the running statistics needs to be turned off and turned on. I was wondering have you experimented with when to turn it on and off?

Thanks,
Rui