How to use attentionentropyloss

[JRMeyer]

>>> Had
[April 14, 2020, 2:05pm]

I have found it here slash

github.com

#### mozilla/TTS/blob/master/layers/losses.py slash #L86

<br/> x curl-run-all.sh discourse.mozilla.org html-to-markdown.sh ordered-posts ordered-posts~ TTS.cdx tts.commands tts-emails.txt TTS.pages tts-telegram.txt TTS.warc.gz mask, target curl-run-all.sh discourse.mozilla.org html-to-markdown.sh ordered-posts ordered-posts~ TTS.cdx tts.commands tts-emails.txt TTS.pages tts-telegram.txt TTS.warc.gz mask, reduction='none')<br/> loss = loss.mul(out_weights.to(loss.device)).sum()<br/> else:<br/> mask = mask.expand_as(x)<br/> loss = functional.mse_loss(<br/> x curl-run-all.sh discourse.mozilla.org html-to-markdown.sh ordered-posts ordered-posts~ TTS.cdx tts.commands tts-emails.txt TTS.pages tts-telegram.txt TTS.warc.gz mask, target curl-run-all.sh discourse.mozilla.org html-to-markdown.sh ordered-posts ordered-posts~ TTS.cdx tts.commands tts-emails.txt TTS.pages tts-telegram.txt TTS.warc.gz mask, reduction='sum')<br/> loss = loss / mask.sum()<br/> return loss<br/> <br/> class AttentionEntropyLoss(nn.Module):<br/> # pylint: disable=R0201<br/> def forward(self, align):<br/> '''<br/> Forces attention to be more decisive by penalizing<br/> soft attention weights<br/> <br/> TODO: arguments<br/> TODO: unit_test<br/> '''<br/> entropy = torch.distributions.Categorical(probs=align).entropy()<br/>

What is the expected tensor size?

[This is an archived TTS discussion thread from discourse.mozilla.org/t/how-to-use-attentionentropyloss]

[JRMeyer]

>>> erogol
[April 15, 2020, 9:41am]

I've used it before but it is not active in the current code base. If
you like you need to run it over the attention alignment vector.

[Archived Post]

[JRMeyer]

>>> Had
[April 15, 2020, 5:49pm]

I want to make alignments for fastspeech train.

So every text token should consist with mel spectrogram frame. slash
My alignment looks like this with multi head attention. slash
It tend to skip whitespaces and other punctuation. slash
Default tacotron is not perfect too. slash

So I want to try this function to make the strict token-to-token
alignment.

Is it fine for slash [batch, width, height slash ]? slash
Or I need to run it for every sample like slash [loss(i) for I in batch slash ]? slash
Maybe I need something else?

Thanks for attention.

[Archived Post]

[JRMeyer]

>>> erogol
[April 16, 2020, 10:06am]

entropy loss would not do that. It works more like a guided attention
loss which forces the alignment to be more diagonal.

Alignment looks quite noisy. Pls share the config.json so I can guess
better.

[Archived Post]