Closed mpariente closed 3 years ago
Cc @Baldwin-disso and @faroit
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.
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)
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! :)
Thanks for that, it will be a good source of inspiration indeed.
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?
Finally passing!
Only works for kernel_size == n_filters
.
I won't spend more time on this.
@mpariente thanks will test that soon
Those are just first steps but it seems it won't be too hard.
Until now:
More work needs to be done to support w**2 OLA correction after the ISTFT (highpri),
center
andnormalize
(lowpri).onesided
will always supposed to beTrue
and we won't support theonesided
argument.Would solve #1