[Bug] MelGAN based vocoders do not use feature matching loss even if it is enabled.

[erogol]

Describe the bug
The condition checking for enabling feature_matching loss in

TTS/TTS/vocoder/layers/losses.py

Line 263 in 4a3cc8d

if self.use_feat_match_loss and not feats_fake:

is always False.

Probably all our trained models are affected by this bug and caused suboptimal results.

Specifically, I observed that it caused the metallic noise in the model outputs.

Expected behavior
The models should use feat_matching loss

Additional context
For anyone who needs an instant fix, the line indicated above needs to be updated as follows.

        if self.use_feat_match_loss and not feats_fake is None:

[erogol]

fixed on dev