RuntimeError with SegFormer and multilabel FocalLoss

[simonreise]

When I train SMP SegFormer with multilabel FocalLoss this RuntimeError appears:

     58 if self.mode in {BINARY_MODE, MULTILABEL_MODE}:
     59     y_true = y_true.view(-1)
---> 60     y_pred = y_pred.view(-1)
     62     if self.ignore_index is not None:
     63         # Filter predictions with ignore label from loss computation
     64         not_ignored = y_true != self.ignore_index

RuntimeError: view size is not compatible with input tensor's size and stride (at least one dimension spans across two contiguous subspaces). Use .reshape(...) instead.

It does not appear in multiclass mode.

I checked my y_pred and y_true shapes and they are

torch.Size([16, 14, 256, 256])
torch.Size([16, 14, 256, 256])

After changing

 loss = self.loss(inputs, labels)

to

 loss = self.loss(inputs.contiguous(), labels.contiguous())

the error disappeared.

Maybe it is related to #998 where .contiguous() is removed from SegFormer decoder.

As mode logic is usually the same in every loss, this error may appear in other losses too, and in binary mode too.

Maybe replacing view with reshape can be a good fix? Or it will slow computations down like contiguous() did?

[qubvel]

Hey @simonreise, thanks for reporting!
I suppose we can replace it with reshape - it should not be a big deal in terms of slowing down.

Let me know if you have bandwidth to submit the fix 🤗