Hi @ChiangE Thank you very much for reporting this issue! We'll take a closer look as soon as possible and I will follow up with colleagues to identify the reason for the current implementation.

Is the line return_complex=True required for your solution?

Hi @beat-buesser, you are welcome.
return_complex=True is suggested by Pytorch but is not a must. However, leaving out this parameter will cause a warning. And different return_complex choices correspond with different ways to calculate transformed_delta . WIth return_complex=False or leaving out this parameter, the following solution should also work:

# Return STFT of delta
delta_stft = torch.stft(
    delta,
    n_fft=self.n_fft,
    hop_length=self.hop_length,
    win_length=self.win_length,
    center=False,
    window=window_fn(self.win_length).to(self.estimator.device),
    return_complex=False    # or omit this parameter
).to(self.estimator.device)

# Take abs of complex STFT results
transformed_delta = torch.sum(torch.square(delta_stft), -1)

# Compute the psd matrix
psd = ((8.0 / 3.0 / self.win_length) ** 2) * transformed_delta 
psd = (
    torch.pow(torch.tensor(10.0).type(torch.float64), torch.tensor(9.6).type(torch.float64)).to(
        self.estimator.device
    )
    / torch.reshape(torch.tensor(original_max_psd).to(self.estimator.device), [-1, 1, 1])
    * psd.type(torch.float64)
)

return psd

@ChiangE Ok, now I see. I think that makes sense. Would you be interested to open a pull request with your solution? (If don't have time I would create it instead.)

Are you sure that transformed_delta doesn't need to be squared for return_complex=True? I'm asking because torch.abs of a complex tensor returns sqrt(real**2 imag**2).

@beat-buesser Excuse me, I make a mistake. You are right, transformed_delta should be squared for return_complex=True. So I have another solution for return_complex=True. I do not use torch.abs anymore since it might do redundant computations as you say.

By the way, I am happy to open a pull request.

# Return STFT of delta
delta_stft = torch.stft(
    delta,
    n_fft=self.n_fft,
    hop_length=self.hop_length,
    win_length=self.win_length,
    center=False,
    window=window_fn(self.win_length).to(self.estimator.device),
    return_complex=True
).to(self.estimator.device)

# Take abs of complex STFT results
transformed_delta = torch.real(delta_stft)**2   torch.imag(delta_stft)**2

# Compute the psd matrix
psd = ((8.0 / 3.0 / self.win_length) ** 2) * transformed_delta 
psd = (
    torch.pow(torch.tensor(10.0).type(torch.float64), torch.tensor(9.6).type(torch.float64)).to(
        self.estimator.device
    )
    / torch.reshape(torch.tensor(original_max_psd).to(self.estimator.device), [-1, 1, 1])
    * psd.type(torch.float64)
)

return psd

Is it correct now?
I also write a PoC to verify that this way does not produce nan grad when delta_stft==0,

import torch

real = torch.FloatTensor([0])
imag = torch.FloatTensor([0])

real.requires_grad = True
imag.requires_grad = True

x = torch.complex(real, imag)
y = torch.real(x)**2   torch.imag(x)**2
print(y)
# >>> tensor([0.], grad_fn=<AbsBackward0>)

y.backward()
print(real.grad, imag.grad)
# >>> tensor([0.]) tensor([0.])

import torch

real = torch.FloatTensor([0])
imag = torch.FloatTensor([0])

real.requires_grad = True
imag.requires_grad = True

y = torch.sqrt(real**2   imag**2)**2
print(y)
# >>> tensor([0.], grad_fn=<SqrtBackward0>)

y.backward()
print(real.grad)
# >>> tensor([nan])

Hi @ChiangE I agree, I think this solution should work. Thank you very much for planning to open a pull request. You can target the PR to branch dev_1.11.0 for the next release of ART. Please let me know if you have any questions about ART PR procedures.

@beat-buesser Sorry, I have problems when I created a pull request. Would you please do that for me?

Hi @ChiangE I have noticed the pull request that you opened and closed, what was the problem there?

The DCO check failed and gave the error message that:

Author: ChiangE, Committer: GitHub; The sign-off is missing.

I tried to commit with git commit -s, but it did not seem to work. Do I just need to commit with git commit -s? Do I need to cancel and remove the previous commit or pull request (without sign-off) thoroughly?

Sorry, actually I am fresh to pull request :(

Hi @ChiangE Sorry for the delay. Did you configure your git with your username and email e.g. with:

git config user.name "FIRST_NAME LAST_NAME"
git config user.email "[email protected]"

After that, could your pleas open a new pull request showing your commits for me to take a look?