rollback torch.norm() in spectrogram()

[lbjcom]

rollback torch.norm() in spectrogram() to v0.4.0 because torch.norm() is very slow in CPU mode as mentioned in following issues:

• Very slow speed in complex_norm() function. #740
• torch.norm is 113x slower than torch.sqrt(a**2 + b**2) #455

[mthrok]

Hi @lbjcom

Thanks for submitting the PR. Looks like you forgot to remove the slow implementation.

[vincentqb]

Thanks for submitting this! It is indeed enough to address the original concern. However, as mentioned here, this affects any function invoking complex_norm (since the latter uses torch.norm). I would prefer fixing complex_norm here instead of spectrogram. Is there a reason why you modified spectrogram instead?

[lbjcom]

Thanks for submitting this! It is indeed enough to address the original concern. However, as mentioned here, this affects any function invoking complex_norm (since the latter uses torch.norm). I would prefer fixing complex_norm here instead of spectrogram. Is there a reason why you modified spectrogram instead?

That's a good point! The reason I changed the spectrogram() is that the difference between v0.4.0 and v0.5.0 is from that function:

But there's no change in complex_norm() between v0.4.0 and v0.5.0. So I changed spectrogram() since I don't know all history of torchaudio. If there's no problem, as you mentioned, fixing complex_norm() is the right solution 😄

[vincentqb]

Thanks for being careful! In this case though, changing complex_norm is the better solution: not only will it fix your problem, but it will also make complex_norm faster :)

[lbjcom]

Thanks for being careful! In this case though, changing complex_norm is the better solution: not only will it fix your problem, but it will also make complex_norm faster :)

I updated the code as you mentioned and found out that there's no problem in my training script. Thanks!

[vincentqb]

LGTM. Waiting on tests to finish before merging :)