The CrossEntropyLoss contains softmax(),so it does not need to add softmax() before the loss funciton