Loss function

[liushunyu]

I find that you use nll_loss as the loss function in PGDtraining, why you don't use cross_entropy?

In ResNet, you use the cross_entropy to train the model

[liushunyu]

Whether the same effect can be achieved with these two loss functions？

[YaxinLi0-0]

You can refer to this link to see the difference of torch.nn.NLLLoss and torch.nn.CrossEntropyLoss : https://discuss.pytorch.org/t/difference-between-cross-entropy-loss-or-log-likelihood-loss/38816

[YaxinLi0-0]

also: https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html

[liushunyu]

So I think that you should use cross_entropy as the loss function in PGDtraining?

Because there is not log_softmax in your resnet, vgg, densenet and so on.

[YaxinLi0-0]

Yes, you are right:) There is a mismatch between the CNN model and other models. I would use CrossEntropyLoss instead of nll_loss and update CNN model to make it unified. Thank you for reporting this issue.