Masked MSE loss - Githubissues

NVIDIA / tacotron2

Tacotron 2 - PyTorch implementation with faster-than-realtime inference

BSD 3-Clause "New" or "Revised" License

5.03k stars 1.37k forks source link

Masked MSE loss #244

Open njellinas opened 5 years ago

njellinas commented 5 years ago

Since the input to the decoder are batched sequences of different lenghts, shouldn't there be a masked MSE loss, not taking into account the padding frames and the extra frames that are produced after the true sentence length?

rafaelvalle commented 5 years ago

Yes! This is something in our to-do list. Please put a PR if you decide to do so.

njellinas commented 5 years ago

Well I finally found it in another repository (Mozilla's TTS) :P Is it possible to include this code in your repo or not? (because of license implications)

rafaelvalle commented 5 years ago

Yes, Mozilla's license is very permissive. https://github.com/mozilla/TTS/blob/master/LICENSE.txt Just put the PR and will review it.