You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Add sigmoid/softmax interface for AsymmetricUnifiedFocalLoss #8603
Is your feature request related to a problem? Please describe.
It seems that AsymmetricUnifiedFocalLoss has not yet been updated to the latest interface like FocalLoss. It is a general loss function and we should make it easier for everyone to use. We can also allow multiple classes segmentation since it only supports binary one accordingly
Describe the solution you'd like
Follow the interface of FocalLoss. Also, we don't need to reinvent the wheel. I am sure we can use the available loss function to replace the implementation.
Hi @ding3820 , I've submitted a PR for this: #8863. It adds use_softmax and use_sigmoid parameters to AsymmetricUnifiedFocalLoss, following the same pattern as FocalLoss. Fully backward compatible — both default to False, so existing code that passes probabilities still works unchanged. Happy to adjust if needed.
Activity
AlexanderSanin commentedon May 18, 2026
Hi @ding3820 , I've submitted a PR for this: #8863. It adds
use_softmaxanduse_sigmoidparameters toAsymmetricUnifiedFocalLoss, following the same pattern asFocalLoss. Fully backward compatible — both default toFalse, so existing code that passes probabilities still works unchanged. Happy to adjust if needed.