License

[JuanFMontesinos]

Hi,
Do you plan to add open-source license? It would be super cool.

[seungwonpark]

Hi, thanks for your interest.
Actually, these codes are not my own - this repo originates from Keunwoo Choi’s work: keunwoochoi/torchaudio-contrib#27
If you plan to use his work(istft_irfft.py), I recommend you to ask him about license there. Are you planning to use my implementation, istft_deconv.py?

[JuanFMontesinos]

Hi, It's the most clean one I've seen, following PyTorch semantics and so on. I was thinking that it would be good to generate a functional from this function and a nn.Module registering the kernel of the convolution as a buffer (not to be recomputed each time you apply an istft)

I have also seen the fft version but I don't think it were faster than convolutional version (did you compared it?).

Anyway, even if the code comes from keunwoochoi, I saw you've applied several changes and copyright "only" covers the usage of exactly the same code. As you rewrote his code it belongs to you in practice.

Edit:
I performed some reconstruction test and error is usually <1e-3, using torch.stft as stft generator. Thus yep, I would like to use it if you don't mind.

[seungwonpark]

Wow, thank you for saying so.

I was thinking that it would be good to generate a functional from this function and a nn.Module registering the kernel of the convolution as a buffer (not to be recomputed each time you apply an istft)

-> Yes, I've already done that privately and works seemlessly. The only reason I didn't show that work here is to compare my code's speed performance with keunwoochoi's implementation. As his code doesn't generate nn.Module, I didn't do that here.

I have also seen the fft version but I don't think it were faster than convolutional version (did you compared it?).

-> Yes, I've experimented that in inspection.ipynb, and found that my implementation is 2~5x slower than keunwoochoi's. If speed matters to you, you'd better use istft_irfft.py. If code semantics matter, then you'll prefer using istft_deconv.py, which is mine ;)

As you rewrote his code it belongs to you in practice.

-> I'm glad to hear that. I'll be right back with commit that adds license.

[seungwonpark]

I've added a license. da9a824
Feel free to reopen this issue if you need further discussion about the license.

[JuanFMontesinos]

Hi, just to be sure, have you considered at the time of timing that cuda is asynchronous?
I don't know if the way you did it is fair. (Soz, before inspection file wasn't loading for me, and not loading now again).
You may want to have a look at https://discuss.pytorch.org/t/how-to-measure-time-in-pytorch/26964
It would be good to move allocate tensors on cuda outside the function to be more precise.

[seungwonpark]

Seems like GitHub can't load IPython notebook these days. :(
I've exported inspection.ipynb to inspection.py at here.

I just did %timeit command in IPython. As you pointed out, I think it would not be a fair comparison. I have almost no knowledge on PyTorch time profiling.

If you create a Pull Request that adds code for fair comparison or precision improvement, I would love to merge it!