timestep scheduling with np.linspace

[ytl0623]

Fixes #8600 .

Description

This np.linspace approach generates a descending array that starts exactly at 999 and ends exactly at 0 (after rounding), ensuring the scheduler samples the entire intended trajectory.

Types of changes

• Non-breaking change (fix or new feature that would not break existing functionality).
• Breaking change (fix or new feature that would cause existing functionality to change).
• New tests added to cover the changes.
• Integration tests passed locally by running ./runtests.sh -f -u --net --coverage.
• Quick tests passed locally by running ./runtests.sh --quick --unittests --disttests.
• In-line docstrings updated.
• Documentation updated, tested make html command in the docs/ folder.

[coderabbitai]

Caution

Review failed

The pull request is closed.

Walkthrough

This PR refactors three loss classes in unified_focal_loss.py by replacing the to_onehot_y parameter with include_background, removing legacy num_classes and weight parameters, and restructuring forward logic to handle background exclusion and per-class asymmetric computations. Additionally, timestep generation in ddpm.py and ddim.py schedulers switches from a ratio-based approach to np.linspace, fixing endpoint coverage to reach num_train_timesteps.

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~60 minutes

• Loss classes refactoring: Three interdependent classes (AsymmetricFocalLoss, AsymmetricFocalTverskyLoss, AsymmetricUnifiedFocalLoss) undergo substantial signature and logic changes; AsymmetricUnifiedFocalLoss now instantiates and delegates to sub-loss components with new activation and parameter handling.
• Parameter changes: include_background replaces to_onehot_y; new parameters (sigmoid, softmax, lambda_focal, focal_loss_gamma/delta, tversky_loss_gamma/delta) expand configurability; requires verification of backward compatibility impact.
• Scheduler fixes: Straightforward algorithm substitution in set_timesteps methods—verify np.linspace rounding behavior and device placement match original intent and resolve the endpoint issue.

Areas requiring extra attention:

• Verify the asymmetric loss formulation correctness in refactored logic paths
• Confirm per-class reduction handling (mean, sum, none) works as intended across all three loss classes
• Check AsymmetricUnifiedFocalLoss weighted combination of sub-losses produces expected results
• Validate timestep endpoint now reaches num_train_timesteps - 1 as intended in both schedulers

✨ Finishing touches

• 📝 Generate docstrings

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment

---

📜 Recent review details

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Cache: Disabled due to data retention organization setting

Knowledge base: Disabled due to Reviews -> Disable Knowledge Base setting

📥 Commits

Reviewing files that changed from the base of the PR and between e72145c and ddf6898.

📒 Files selected for processing (3)

• monai/losses/unified_focal_loss.py (1 hunks)
• monai/networks/schedulers/ddim.py (1 hunks)
• monai/networks/schedulers/ddpm.py (1 hunks)

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}