AlignerNet instead of MAS

[codeghees]

Is it possible to use AlignerNet (aligner.py in pflow-tts repo) instead of MAS in VITS2?

What should be changed in the code? I am a bit confused on what the inputs should be.

[p0p4k]

Check out pflow repo for guidance.

[codeghees]

Hey, thanks for replying!

I did take most of the code from that repo. I am trying to debug why my alignment curve looks like this:

align_loss is being added to duration loss.
Inputs:

 aln_hard, aln_soft, aln_log, aln_mask = self.aligner(
            m_p.transpose(1,2), x_mask, y, y_mask
            )
attn = aln_mask.transpose(1,2).unsqueeze(1)
align_loss = self.aligner_loss(aln_log, x_lengths, y_lengths)
m_p is returned by TextEncoder.

Appreciate any insights!

[p0p4k]

It might be that it is still learning the durations. It think MAS is good enough. What is important is the duration_predictor module.

[codeghees]

So this above graph is after training for several days and thousands of steps. It seems like some bug - maybe in shape size or sth.

The output is basically the first word is legible but the rest is basically gibberish.

[p0p4k]

I see. Maybe it needs some fixing to do, if you can start a PR we can debug this together. I am busy with other stuff.

[Tera2Space]

Probably same problem as we have in pflow)

[codeghees]

@Tera2Space What was the problem?

[Tera2Space]

@Tera2Space What was the problem?

Code geass nice, problem was that it generate wrong aligment, i just though of reason: p0p4k/pflowtts_pytorch#24 (comment)

in your model, what shape is input to alignernet?

[codeghees]

@Tera2Space @p0p4k I added a basic PR of the changes I have so far: #82

[codeghees]

For a single batch - shapes look something like this:

m_p torch.Size([32, 192, 74])
x_mask torch.Size([32, 1, 74])
y_mask torch.Size([32, 1, 293])
m_p torch.Size([32, 192, 74])
x torch.Size([32, 192, 74])
y torch.Size([32, 80, 293])
m_p.transpose(1,2) torch.Size([32, 74, 192])

[Tera2Space]

For a single batch - shapes look something like this:

m_p torch.Size([32, 192, 74])
x_mask torch.Size([32, 1, 74])
y_mask torch.Size([32, 1, 293])
m_p torch.Size([32, 192, 74])
x torch.Size([32, 192, 74])
y torch.Size([32, 80, 293])
m_p.transpose(1,2) torch.Size([32, 74, 192])

hm than my guess was wrong/there are more problems.
I tested my idea on pflow and it didn't work too, so probably we need to check for another problems, I think we should use something like https://github.com/lucidrains/naturalspeech2-pytorch as reference.

[p0p4k]

Iirc, I had yanked the alignernet from there.

[codeghees]

Yeah currently trying out various inputs to Aligner; possibly an input issue. It might be an issue with the masks.

[p0p4k]

still not convinced about putting efforts in aligner net, we must focus on a better TextEncoder instead.

[codeghees]

do you think that's a bottleneck right now?

[p0p4k]

For vits2 it should be the duration predictor, and for pflow it should be both textencoder and duration predictor. MAS gives good alignments during training, it is during inference that these models perform worse.

[Tera2Space]

pflow it should be both textencoder

how do you think we can improve pflow's encoder?

[nicemanis]

Hey, thanks for replying!

I did take most of the code from that repo. I am trying to debug why my alignment curve looks like this:

align_loss is being added to duration loss. Inputs:

 aln_hard, aln_soft, aln_log, aln_mask = self.aligner(
            m_p.transpose(1,2), x_mask, y, y_mask
            )
attn = aln_mask.transpose(1,2).unsqueeze(1)
align_loss = self.aligner_loss(aln_log, x_lengths, y_lengths)
m_p is returned by TextEncoder.

Appreciate any insights!

Is padding included in the encoder timesteps (it seems to me that it is)? You can remove the padded part from the plot.