Stop loss jumps during gradual training

[JRMeyer]

>>> mathematiguy
[June 5, 2020, 12:48am]

I've noticed a pattern where the stop loss jumps discontinuously when
decreasing the number of decoder frames + batch size during training.
Has anyone else seen this?

So far as I can tell, it doesn't seem to greatly diminish inference
quality (note the alignment and loss decoder scores aren't affected
much). But I was curious if it had some sort of explanation.

[This is an archived TTS discussion thread from discourse.mozilla.org/t/stop-loss-jumps-during-gradual-training]

[JRMeyer]

>>> georroussos
[June 6, 2020, 4:08pm]
Is your r 1? [Maybe that is why. I had the sam]e problem, but I think it
was because I forgot to switch off the transliteration switch and I was
training in a language with special characters, so it might have messed
with the stopnet as well. Have you trimmed silences?

Alternatively, if your r is 2, try training with 3 a bit longer.

[Archived Post]

[JRMeyer]

>>> erogol
[June 8, 2020, 9:24am]

have you checked how it performs in practice?

[Archived Post]