Potential Interface Inconsistency between SegmentationLosses and DataParallelCriterion

First of all, thanks a lot for your excellent code! I found a potential bug in calculating segmentation loss.

If not using CUDA

This line

self.criterion = DataParallelCriterion(self.criterion).cuda()

is not executed, and self.criterion is a SegmentationLosses object. When it calls the criterion at this position,

outputs = self.model(image)
loss = self.criterion(outputs, target)

wouldn't it cause error in this case?

def forward(self, *inputs):
    ...
    pred1, se_pred, pred2, target = tuple(inputs)

Since it is trying to execute pred1, se_pred, pred2, target = (outputs, target).

In the case of using CUDA

self.criterion is a DataParallelCriterion object. It calls SegmentationLosses in a proper way

output = module(*(input + target), **kwargs)

Conclusion

There is interface inconsistency between SegmentationLosses and DataParallelCriterion. I think their calling should be different, e.g.

if args.cuda:
    loss = self.criterion(outputs, target)
else:
    loss = self.criterion(*(list(outputs) + [target]))

Start with experiments/segmentation/train.py around the criterion setup and call, then compare encoding/nn/loss.py with encoding/parallel.py to trace the CPU and CUDA argument shapes. Reproduce segmentation training without CUDA and verify that the criterion interface works consistently in both modes without an argument-unpacking error.