[WIP] Add VITS-2

[ezerhouni]

Add VITS-2 Recipe for LJSpeech.
For the moment, I am facing a bug tracked here :
#1508

I will add the License etc later on, once the PR work

[csukuangfj]

Does torch.min(inputs, dim=None) work for you?

[ezerhouni]

For now, I am getting

  File "/vits2/egs/ljspeech/TTS/vits2/transform.py", line 118, in rational_quadratic_spline
    if torch.min(inputs, dim=None) < left or torch.max(inputs) > right:
RuntimeError: Please look up dimensions by name, got: name = None.

Debugging as we speak

Edit: The previous crash happened at epoch 918. I think it might come from the transformer layer I am adding

[csukuangfj]

What is the output of

print(type(inputs))

[ezerhouni]

What is the output of

print(type(inputs))

<class 'torch.Tensor'>

[csukuangfj]

As a last attempt, could you replace

torch.min(inputs, dim=None)

with

torch.min(inputs.reshape(-1), dim=0)

---

I just searched for the new error, and found
pytorch/pytorch#70925

Not sure if inputs is a named tensor.

[ezerhouni]

@csukuangfj Thanks for your help, I have replaced by:

    if torch.min(inputs.reshape(-1), dim=0)[0] < left or torch.max(inputs.reshape(-1), dim=0)[0] > right:
        raise ValueError("Input to a transform is not within its domain")

and it seems to run. It takes ~ 24h on my machine to train a model for 1000 epochs. I will let you know once the training is done

[ezerhouni]

@csukuangfj I will try to debug more, but I think the issue might comes from there

[ezerhouni]

@csukuangfj I tried with the fix that you proposed and with the new tokenization, both throw the same error.

[csukuangfj]

Could you check that inputs is not empty in

    if torch.min(inputs.reshape(-1), dim=0)[0] < left or torch.max(inputs.reshape(-1), dim=0)[0] > right:

?

If it is empty, could you find why it is empty?

[ezerhouni]

@csukuangfj The training works fine. The quality is not optimal for the moment, so I am playing a bit with the parameters and will let you know

[ezerhouni]

@csukuangfj I am still having issues to train a model. I will close this PR and re-open one once I have a more stable branch.