Xu Haoming
Results
2
comments of
Xu Haoming
> @dhkim0225 Thank you for your input! So eventually it will be like > > ``` > with torch.no_grad(), torch.cuda.amp.autocast(enabled=False): > cka = CKA(model1, model2, > model1_name=model1_name > model2_name=model2_name >...
> @HaomingX I suspect probably it is due to the explosion of computation (e.g. gradients) that leads to nan. Try to reduce the amount of computations needed such as layers...