Tacotron2: where is the denoiser?

[NileZhou]

in file inference.ipynb,
`ModuleNotFoundError Traceback (most recent call last)
in ()
16 from train import load_model
17 from text import text_to_sequence
---> 18 from denoiser import Denoiser

ModuleNotFoundError: No module named 'denoiser'`

if the denoiser is in apex or other python models?

[akshitmittal1]

denoiser is a .py file in the repository itself.

[NileZhou]

denoiser is a .py file in the repository itself.

thanks. but I still don't understand, i can't find the file in tacotron2

[akshitmittal1]

okay, I don't know why is that happening but you can just copy its code and make a .py file denoiser itself to make it work

[vikrantsharma7]

It's in the waveglow repo.
sys.path.append('waveglow/')

If still not found

cd waveglow/
git pull origin master

[NileZhou]

It's in the waveglow repo.
sys.path.append('waveglow/')

If still not found

cd waveglow/
git pull origin master

thanks, i just delete the code to avoid it, with out denoiser, i can generate voice by using nvidia's pre-trained model. however, my training's train loss is always oscillation around 0.4

[rafaelvalle]

Remember to run Initialize submodule: git submodule init; git submodule update
Assuming you have the same setup as the one in this implementation, you need to anneal the learning rate to further decrease the loss.

[orena1]

Thanks @rafaelvalle but I think that in the waveglow submodule Denoiser is not there:
https://github.com/NVIDIA/waveglow/tree/4b1001fa3336a1184b8293745bb89b177457f09b

[vikrantsharma7]

The denoiser was added after the commit you mention.
Pull the latest master branch and you will find it.

[orena1]

I understand @vikrantsharma7 but as far as I see the git of tactron2 points to this commit

[vikrantsharma7]

Yes, it does and it hasn't been updated to use a commit that includes the denoiser, so you will need to manually pull it, or run inference without the denoiser.

[SomeUserName1]

can confirm that after
git submodule init; git submodule update
one needs also to
git pull origin master; git checkout 6188a1d106a1060336040db82f464d6441f39e21
(If you dont checkout this commit there will be an AttributeError object NM does not have attribute cond_layer)
further the pretrained waveglow weights file needs to be renamed (cut off the _ljs_vX.pt)
Then it should work out of the box

[NoahDrisort]

I have a same problem, have you solve this, thank you

[ColeMurray]

Experienced the above issue. @SomeUserName1 suggestion to pull in the latest resolved this.