Fix boolean operator precedence bugs in GeneratorLoss

[Mr-Neutr0n]

Summary

• Fix hinge loss guard: not scores_fake is not None was parsed as (not scores_fake) is not None due to Python operator precedence, which is always True — the hinge loss block ran unconditionally even when scores_fake was None, causing AttributeError at runtime.
• Fix feature matching loss guard: not feats_fake is None was parsed as (not feats_fake) is None, which is always False — the feature matching loss was silently never applied.
• Both conditions now use the idiomatic x is not None form, consistent with the existing MSE loss check on the preceding line.

Details

In TTS/vocoder/layers/losses.py, the GeneratorLoss.forward() method had two boolean expressions affected by Python operator precedence:

Line	Before (buggy)	Parsed as	Result
295	not scores_fake is not None	(not scores_fake) is not None	Always True (bool is never None)
301	not feats_fake is None	(not feats_fake) is None	Always False (bool is never None)

The not unary operator binds more tightly than is / is not, so these expressions do not test what was intended.

After fix: both lines use scores_fake is not None and feats_fake is not None, matching the pattern already used for the MSE loss guard on line 289.

Test plan

• Verify hinge GAN loss is now correctly computed when scores_fake is provided
• Verify feature matching loss is now correctly applied when feats_fake is provided
• Run existing vocoder tests: python -m pytest tests/vocoder_tests/

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions. You might also look our discussion channels.