Add from_logits support to FocalLoss and eps for API consistency with DiceLoss

[Harsh-2005d]

Summary

This PR adds a from_logits and eps parameter to FocalLoss to align its API with DiceLoss.

Currently, DiceLoss supports both logits and probability inputs via the from_logits flag, while FocalLoss always assumes raw logits. This creates inconsistency when using models configured with an activation function (e.g., activation="softmax"), requiring users to manually remove activations or wrap the loss.

This change introduces a from_logits flag to FocalLoss to support both logits and probability inputs in a consistent manner.

---

Motivation

Adresses the Issue #1263
Example of current behavior:

model = smp.create_model(..., activation="softmax")
outputs = model(x)  # probabilities

dice_loss = smp.losses.DiceLoss(mode="multiclass", from_logits=False)  # works
focal_loss = smp.losses.FocalLoss(mode="multiclass")  # expects logits → incorrect

Because FocalLoss always assumes logits, using it with probability outputs results in incorrect loss computation.

This PR resolves that inconsistency.

---

Implementation Details

• Added from_logits: bool = True parameter to FocalLoss.

• When from_logits=False:

  • For binary/multilabel modes:

    • Probabilities are converted to logits using inverse sigmoid.

  • For multiclass mode:

    • Softmax probabilities are converted to log-space before applying the existing focal computation.

• No changes were made to the underlying focal formulation.

• Backward compatibility is preserved (from_logits=True by default).

---

Tests

Added tests to verify:

• FocalLoss works correctly with from_logits=False.
• Probability-based input produces finite and consistent results.
• Good predictions yield lower loss than bad predictions.

All existing tests pass.

---

Backward Compatibility

• Default behavior remains unchanged.
• Existing code using logits is unaffected.
• New functionality only activates when from_logits=False.

---

[Harsh-2005d]

@qubvel also probably a better implementation of focal loss would be something like this by open-mm lab, if u want i can do that also

[qubvel]

Thanks for your contribution! Please see the comments, looks good otherwise.

Re open-mm lab loss: if you can provide some details why it would be a better option I would be happy to consider adding it, thanks!

[Harsh-2005d]

The current fix keeps the existing implementation unchanged, but it still reconstructs logits from probabilities, which is not strictly equivalent to the original logits formulation. The OpenMMLab implementation handles logits and activated probabilities separately and also provides an optimized CUDA implementation. I kept this PR minimal, but I can open a separate issue to discuss the differences between the two implementations if that would be helpful.

[qubvel]

Please run make fixup to fix style test, otherwise looks good.

also provides an optimized CUDA implementation. I kept this PR minimal

thanks for keeping it minimal and easy to review, I would prefer to merge it as is, don't want to add optimized CUDA kernels atm

[Harsh-2005d]

what's the issue here?

[qubvel]

No issues, merging!