Attention weights with partial flat line (non-english)

[SornrasakC]

Hi, I have been trying to train this model with Thai-dataset (1 speaker, ~5 hour).

After ~80k Steps (batch size = 1, ~31 epoch), the attention weights turns out like this

Is it normal to see partial flat lines like this? all the issues I looked through only sees entire flat line or just straight diagonal... Or am I being too impatient? it's just 80k steps after all.

Here's some additional info

(Is this even correct?)

The above result comes from me warm starting the model from flowtron_ljs.pt with the flow=1 config file (speaker_embedding.weight ignored)

Things I have done

• Trim the start and end of the sound using Librosa, also filtered out any data with a duration longer than 10 secs.
• Set the p_arpabet=0 and instead of using Thai symbols, I convert all my filelists into IPAs first, and then add those IPAs symbols to the symbols.py
• Change the cleaner so that they don't transliterate.
• Add this line to ignore embedding.weight since they have different shape during warmstart.

Additional Questions

• Does the above gate output make sense? does this means the model think all sounds end at 350-ich frame?
• What happens if I train the model with only one flow? can I still do the inference and/or style transfer with lower score? or would it just goes un-usable due to absence of reverse mapping(?)
• To my understanding, the process I should be going through is first, train the model flow=1 until attentions aligned, seconds, same but flow=2, and then third, turns the attn_prior off to attends the speaker. What's the sign to look for during third step? how do I know if the model has attended?
• In default config, the ctc_loss starts at 10k iters, do I need to change this? does starting this earlier or later affects anything?
• Does warm starting from flowtron_ljs really helps in learning different language? I'm wondering which parts did it helps with, the decoder?
• Regarding emotion transfer, If I happened to get this Thai-dataset working, can I use an English dataset to transfer the emotion into this? Or do I also need Thai-dataset with emotion labeled as well?

Thank you for reading and would really appreciate any answers or suggestions.

[Bahm9919]

I think The problem is representation of your symbols with text. You must get good alignment in 10k steps. If don't its mean something get wrong.

Yes. Warmstart helping.

I don't know whats happening, but yes you can do inference with only 1 flow.

[SornrasakC]

I think The problem is representation of your symbols with text. You must get good alignment in 10k steps. If don't its mean something get wrong.

Yes. Warmstart helping.

I don't know whats happening, but yes you can do inference with only 1 flow.

I guess.. I will try changing my symbols to ASCII, will come update soon.

[SornrasakC]

Sorry for a very late reply.

I tried changing my symbols to characters such as @A1 @A2 along with a new filelists that is already converted to the added symbols.

Here's the results after 20k iters, which I honestly think doesn't have any differences... I wonder where did it goes wrong.

[SornrasakC]

Turn out, 20 out of ~2k audio files was just pure noise, removing them solves the problem.

[SornrasakC]

@Bahm9919 Sorry for the @ and for commenting on the closed issue. I saw that it seems like you just recently able to do the inference, would it be possible for you to kindly share how you did it?

Especially

• Torch version
• Any changes in inference.py
• Which WaveGlow weight
• Does your submodule has this commit? Submodule path 'tacotron2': checked out '6f435f7f29c3e1553cf2dd7ca2daf56903b20c39'

Or anything else, surprisingly? Very much appreciated.

[Bahm9919]

@Bahm9919 Sorry for the @ and for commenting on the closed issue. I saw that it seems like you just recently able to do the inference, would it be possible for you to kindly share how you did it?

Especially

• Torch version
• Any changes in inference.py
• Which WaveGlow weight
• Does your submodule has this commit? Submodule path 'tacotron2': checked out '6f435f7f29c3e1553cf2dd7ca2daf56903b20c39'

Or anything else, surprisingly? Very much appreciated.

I see your issue, will answer there.