Skip to content

Bug in the generalized Wasserstein Dice loss #4650

Closed

Description

@LucasFidon

Describe the bug
The values returned by the loss are not correct for batch sizes larger than 1

To Reproduce
In the binary case and for M=[[1,0],[0,1]] the generalized Wasserstein Dice loss should be the same as the Dice loss.
For batch size > 1 it is not the case.

Expected behavior
See above. For this special case the GWDL and the DL should return the same values.

Additional context
I am the author of the generalized Wasserstein Dice loss paper and the bug has been already fixed in my git repository
https://github.com/LucasFidon/GeneralizedWassersteinDiceLoss
I will fix the bug in MONAI too and add unit tests.

Activity

  1. hongjie-qiu commented on Feb 19, 2026

    @hongjie-qiu
    Contributor

    Hey! I noticed the same problem, found this existing open issue, and went ahead to put together a fix.
    The root cause is that alpha_extended keeps shape (B, 1, S) after the torch.gather in _compute_generalized_true_positive and _compute_denominator, so when it's multiplied with wasserstein_distance_map (B, S) it silently broadcasts to (B, B, S) for any batch size > 1. The fix squeezes that extra dim after gather, matching the reference implementation. Also fixed the reduction="none" path which had a related reshape issue.

    cc @KumoLiu would appreciate a review when you get a chance

  2. added a commit that references this issue on Feb 27, 2026
    4b1777f
  3. added a commit that references this issue on May 25, 2026
    3b0751a
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions