The torch deep learning model should not use softmax last layer + CrossEntropyLoss

[yxu02]

Torch CrossEntropyLoss expects logits instead of normalized probabilities.

[Graeme Malcolm]

Thanks - fixed