You should pass raw logits to nn.CrossEntropyLoss, since the function itself applies F.log_softmax and nn.NLLLoss() on the input.
If you pass log probabilities (from nn.LogSoftmax) or probabilities (from nn.Softmax()) your loss function won’t work as intended.
6 Likes