[Warning] stft with return_complex=False is deprecated.

[Alienpups]

Describe the bug

If I start to train my model I always get this warning message. Is it just me or is this a thing?

TRAINING (2023-05-29 00:46:49)
C:\TTS-Training\lib\site-packages\torch\functional.py:641: UserWarning: stft with return_complex=False is deprecated. In a future pytorch release, stft will return complex tensors for all inputs, and return_complex=False will raise an error.
Note: you can still call torch.view_as_real on the complex output to recover the old return format. (Triggered internally at ..\aten\src\ATen\native\SpectralOps.cpp:867.)
return _VF.stft(input, n_fft, hop_length, win_length, window, # type: ignore[attr-defined]

To Reproduce

I run this command: C:\TTS-Training>python train_vits.py --continue_path "vits_louise-voice-April-21-2023_05+38PM-0000000".
Training starts but I always get warned about some sft error.

TRAINING (2023-05-29 00:46:49)
C:\TTS-Training\lib\site-packages\torch\functional.py:641: UserWarning: stft with return_complex=False is deprecated. In a future pytorch release, stft will return complex tensors for all inputs, and return_complex=False will raise an error.
Note: you can still call torch.view_as_real on the complex output to recover the old return format. (Triggered internally at ..\aten\src\ATen\native\SpectralOps.cpp:867.)
return _VF.stft(input, n_fft, hop_length, win_length, window, # type: ignore[attr-defined]

Expected behavior

No response

Logs

No response

Environment

{
    "CUDA": {
        "GPU": [],
        "available": false,
        "version": null
    },
    "Packages": {
        "PyTorch_debug": false,
        "PyTorch_version": "2.0.0+cpu",
        "TTS": "0.14.0",
        "numpy": "1.23.5"
    },
    "System": {
        "OS": "Windows",
        "architecture": [
            "64bit",
            "WindowsPE"
        ],
        "processor": "Intel64 Family 6 Model 151 Stepping 2, GenuineIntel",
        "python": "3.10.11",
        "version": "10.0.22631"
    }
}

Additional context

No response

[Edresson]

Hi @Alienpups,

Yeah, it is a thing.

To fix it, we need to change our STFT class to use return_complex=True and replace:

M = o[:, :, :, 0]
P = o[:, :, :, 1]
S = torch.sqrt(torch.clamp(M**2 + P**2, min=1e-8))

by

o = torch.view_as_real(o)
S = torch.norm(o, p=2, dim=-1)

A PR would be welcome :).

[Alienpups]

To fix this warning, do I just need to make the changes you mentioned in torch_transforms.py? Did I understand that correctly? And what exactly does PR mean? :)

So I replace these three lines...
M = o[:, :, :, 0]
P = o[:, :, :, 1]
S = torch.sqrt(torch.clamp(M2 + P2, min=1e-8))

with the following? ...
o = torch.view_as_real(o)
S = torch.norm(o, p=2, dim=-1)

[Edresson]

To fix this warning, do I just need to make the changes you mentioned in torch_transforms.py? Did I understand that correctly? And what exactly does PR mean? :)

So I replace these three lines... M = o[:, :, :, 0] P = o[:, :, :, 1] S = torch.sqrt(torch.clamp(M2 + P2, min=1e-8))

with the following? ... o = torch.view_as_real(o) S = torch.norm(o, p=2, dim=-1)

Yes, you also need to use "return_complex=True" on the torch.stft function call.

Sorry, PR means pull request. If you like you can do these changes and then send a pull request and turn a 🐸 TTS contributor :).

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions. You might also look our discussion channels.

[pivolan]

I have the same problem, and fix by change edit file python3.10/site-packages/torch/functional.py
at line number 641, just insert this before return statement:
you should find this string:

return _VF.stft(input, n_fft, hop_length, win_length, window,  # type: ignore[attr-defined]

and insert this before:

    if not return_complex:
        return torch.view_as_real(_VF.stft(input, n_fft, hop_length, win_length, window, normalized, onesided, return_complex=True))

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions. You might also look our discussion channels.