Bug fix: remove softmax layer from model loading code

We should only add softmax when we explicitly require it (as when
exporting to ONNX), because CrossEntropyLoss implicitly computes
softmax on the logits it receives as input.

This cade was left there by mistake and should have never been
pushed to git.

models/__init__.py

@@ -75,9 +75,5 @@ def create_model(pretrained, dataset, arch, parallel=True, device_ids=None):
     elif parallel:
         model = torch.nn.DataParallel(model, device_ids=device_ids)
 
-    # explicitly add a softmax layer, because it is useful when exporting to ONNX
-    model.original_forward = model.forward
-    softmax = torch.nn.Softmax(dim=1)
-    model.forward = lambda input: softmax(model.original_forward(input))
     model.cuda()
     return model