Matching torch.stft and torch.istft

[mpariente]

Those are just first steps but it seems it won't be too hard.

Until now:

• STFT seems to pass easily, with various windows and window sizes. (More tests needed to cover edge cases etc..)
• ISTFT passes with square window (without counting the first and last frames where the overlap is not complete).

More work needs to be done to support w**2 OLA correction after the ISTFT (highpri), center and normalize (lowpri).

onesided will always supposed to be True and we won't support the onesided argument.

Would solve #1

[mpariente]

Cc @Baldwin-disso and @faroit

[mpariente]

The ISTFT looks in a good shape as well. I need to write more informal tests with different windows to make sure it works, but given the C++ code, I'm pretty sure it'll be fine.

[mpariente]

torch.isttf raises RuntimeError for most windows, so we need to use center=True to avoid boundary effects on the w**2 OLA. The forward works with center=True apparently (to be check with other strides etc.. but looks in a good shape)

[jonashaag]

FYI there is a ton of tests here: https://github.com/pytorch/pytorch/blob/master/test/test_spectral_ops.py probably overkill to duplicate all of them but maybe for inspiration! :)

[mpariente]

Thanks for that, it will be a good source of inspiration indeed.

[mpariente]

So the ISTFT also works with center=True, when we don't have the length argument.. And now comes a new problem, we should be able to pass kwargs to the hooks in order to pass the length. But how to organize that?

[mpariente]

Finally passing! Only works for kernel_size == n_filters. I won't spend more time on this.

[faroit]

@mpariente thanks will test that soon