STFT.inverse() fails on magnitude spectrograms when called multiple times about nnaudio HOT 2 CLOSED

Hi tasercake, thanks for pointing out this problem and providing a possible fix.
The reason for having the self.output_format in STFT.forward() is that Griffin-Lim (another inverse function inside STFT) requires the complex output.
But you are right, we should figure out a way to avoid this bug while allowing Griffin-Lim to work.

Since this fix will also affect Griffin-Lim, I think I need more time later to investigate.
Alternatively, would you like to make a pull request? So that I can just check your fix to see if it works for Griffin-Lim too.

I will post the instruction on how to do unit test if you are interested in doing a pull request.

from nnaudio.

I've created a PR with the changes I made on my end (though I haven't changed/added any tests yet) - #53

At least in my specific use case of inverting a magnitude spectrogram via Griffin-Lim, these changes didn't seem to cause any issues

from nnaudio.