[Bugfix] Fix bug in cross entropy loss (#3457)

Thanks for your contribution and we appreciate it a lot. The following
instructions would make your pull request more healthy and more easily
get feedback. If you do not understand some items, don't worry, just
make the pull request and seek help from maintainers.

## Motivation

Fixes #3412

## Modification

We just need to replace tensor creation using torch.stack() instead of
torch.tensor().

## BC-breaking (Optional)

Does the modification introduce changes that break the
backward-compatibility of the downstream repos?
If so, please describe how it breaks the compatibility and how the
downstream projects should modify their code to keep compatibility with
this PR.

## Use cases (Optional)

If this PR introduces a new feature, it is better to list some use cases
here, and update the documentation.

## Checklist

1. Pre-commit or other linting tools are used to fix the potential lint
issues.
2. The modification is covered by complete unit tests. If not, please
add more unit test to ensure the correctness.
3. If the modification has potential influence on downstream projects,
this PR should be tested with downstream projects, like MMDet or
MMDet3D.
4. The documentation has been modified accordingly, like docstring or
example tutorials.

mmseg/models/losses/cross_entropy_loss.py

@@ -63,8 +63,9 @@ def cross_entropy(pred,
 
  else:
  # the average factor should take the class weights into account
- label_weights = torch.tensor([class_weight[cls] for cls in label],
- device=class_weight.device)
+ label_weights = torch.stack([class_weight[cls] for cls in label
+ ]).to(device=class_weight.device)
+
  if avg_non_ignore:
  label_weights[label == ignore_index] = 0
  avg_factor = label_weights.sum()

tests/test_models/test_losses/test_cross_entropy_loss.py

@@ -0,0 +1,28 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+import torch.nn.functional as F
+
+from mmseg.models.losses import CrossEntropyLoss, weight_reduce_loss
+
+
+def test_cross_entropy_loss_class_weights():
+ loss_class = CrossEntropyLoss
+ pred = torch.rand((1, 10, 4, 4))
+ target = torch.randint(0, 10, (1, 4, 4))
+ class_weight = torch.ones(10)
+ avg_factor = target.numel()
+
+ cross_entropy_loss = F.cross_entropy(
+ pred, target, weight=class_weight, reduction='none', ignore_index=-100)
+
+ expected_loss = weight_reduce_loss(
+ cross_entropy_loss,
+ weight=None,
+ reduction='mean',
+ avg_factor=avg_factor)
+
+ # Test loss forward
+ loss = loss_class(class_weight=class_weight.tolist())(pred, target)
+
+ assert isinstance(loss, torch.Tensor)
+ assert expected_loss == loss