feat: Raise ValueError for alpha > 1 in sigmoid_focal_loss

[developer0hye]

Overview

This PR updates the sigmoid_focal_loss function to explicitly raise a ValueError when alpha > 1.
In the focal loss formulation, alpha is meant to balance positive and negative examples and should
generally lie within the range (0, 1) or be set to -1 to ignore the balancing factor.
Allowing alpha to exceed 1 can lead to unexpected behaviors.

Changes

• Added a condition to check if alpha > 1. If true, it raises a ValueError.
• Clarified in the docstring that valid values of alpha are within (0, 1) or -1 for ignoring.

Reasoning

• According to the focal loss paper and its implementations, alpha in the range of (0, 1) (or -1)
  is standard practice.
• Ensuring alpha does not exceed 1 prevents users from accidentally passing invalid parameters,
  thus avoiding silent errors or unstable training behavior.

Additional Notes

• If there is a special use case for greater, we can discuss supporting that explicitly.
  However, for most applications, restricting alpha to (0, 1) or -1 remains consistent with the
  original focal loss approach.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/8882

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.