fix: use return value of torch.clamp for weight clamping

[Mr-Neutr0n]

Summary

• torch.clamp() is not in-place — it returns a new tensor. In both GE2ELoss.forward() and AngleProtoLoss.forward() in TTS/encoder/losses.py, the call torch.clamp(self.w, 1e-6) discards the return value, so the weight parameter self.w is never actually clamped to a minimum of 1e-6.
• Replaced with self.w.data.clamp_(1e-6) which correctly clamps the parameter's underlying data in-place, ensuring the weight scaling factor stays positive as intended.

Details

The bug exists in two locations in TTS/encoder/losses.py:

GE2ELoss.forward() (line 115):

# Before (bug): return value discarded, self.w unchanged
torch.clamp(self.w, 1e-6)

# After (fix): in-place clamp on parameter data
self.w.data.clamp_(1e-6)

AngleProtoLoss.forward() (line 159):

# Before (bug): return value discarded, self.w unchanged
torch.clamp(self.w, 1e-6)

# After (fix): in-place clamp on parameter data
self.w.data.clamp_(1e-6)

Without this fix, if self.w becomes negative during training (via gradient updates), the cosine similarity matrix scaling would be inverted, potentially destabilizing speaker encoder training.

Test plan

• Verify self.w stays non-negative during GE2E loss training
• Verify self.w stays non-negative during AngleProto loss training
• Run existing encoder loss unit tests

[CLAassistant]

All committers have signed the CLA.

[Mr-Neutr0n]

closing this out — saw the note about the repo being unmaintained. will redirect to the active fork if still relevant