Chapter 12 - Bug Using Softmax Output on CrossEntropyLoss

[liereynaldo]

Since CrossEntropyLoss on Pytorch expecting Logit input is incompatible with Softmax outputting probability, the loss will be Loss=−log(Softmax(Softmax(logits))), so to resolve this either:

1. The last layer of the model (softmax) should be removed & keep CrossEntropyLoss
2. Change the last layer into logsoftmax and use NLLLoss as loss function