Autograd tests for Transforms

[mthrok]

Until recent, we have been assumed that ops provided in torchaudio support autograd just because they are implemented with PyTorch. However, this assumption was not correct all the time. For example, in #704, it was pointed out that lfilter does not support autograd, and this was resolved in #1310 with proper unit tests by community contribution. Similarly, as a part of #1337, I have added autograd test to some transforms in #1340. We would like to extend the autograd testing to as many functionals and transforms as possible. We would like to ask your help.

Steps

1. Pick a transform from the list bellow and see if someone is working on it. If not, leave a comment in this thread that you will be working on it.
2. Add a test in AutogradTestMixin similar to the existing tests.
   NOTE Please first try adding test without providing nondet_tol to see if the transform's backward pass is deterministic or not. If you see Backward is not reentrant error message, that means it is not deterministic. Please report it back here, so that we can discuss how to handle it.
3. Run test with (cd test && pytest torchaudio_unittest/transforms/autograd_*_test.py).
4. Make a PR and add cc @mthrok in the description (or comment)

Note: We are not sure if all the transforms actually support autograd. If you find a Transform does not support autograd please report back in this thread.

For the instruction to set up development environment, please refer to CONTRIBUTING.md.

Transforms

• AmplitudeToDB Add autograd test for T.AmplitudeToDB #1447
  is used by MFCC so it should support but having an independent test is be nice
• Spectrogram Add autograd test for T.Spectrogram/T.MelSpectrogram #1340
• GriffinLim Add autograd test for T.GriffinLim #1421
• MelScale Add autograd test for T.MelScale #1467
  is used by MelSpectrogram so it should support but having an independent test is be nice
• InverseMelScale
  this transform most likely does not support autograd
• MelSpectrogram Add autograd test for T.Spectrogram/T.MelSpectrogram #1340
• MFCC Autograd tests for Transforms MFCC #1415
• MuLawEncoding
• MuLawDecoding
MuLaws are quantization operation, thus they do not support autograd.
• Resample Add Autograd test for T.Resample #1416
• ComputeDeltas Add autograd test for T.ComputeDeltas #1422
• TimeStretch Add autograd test to T.TimeStretch (and F.phase_vocoder) #1420
  NOTE: TimeStretch requires coordinate transform from cartesian to polar which uses atan2 and atan2 is not differentiable around zero.
• Fade Add autograd test for T.Fade #1424
• FrequencyMasking Add autograd tests for TimeMasking/FrequencyMasking #1498
• TimeMasking Add autograd tests for TimeMasking/FrequencyMasking #1498
• Vol Add autograd test for T.Vol #1460
• SlidingWindowCmn Adding test for T.SlidingWindowCmn #1482
• SpectralCentroid Add autograd test for T.SpectralCentroid #1425
• Vad
  this transform most likely does not support autograd

[yoyololicon]

@mthrok
I have some experiences with spectrogram inversion before, so I would like to take on GriffinLim.
Also ComputeDeltas.

[mthrok]

@mthrok
I have some experiences with spectrogram inversion before, so I would like to take on GriffinLim.

@yoyololicon

Thanks. Let me know if you need help there.

[krishnakalyan3]

I will be working on MFCC

[krishnakalyan3]

Working on resample, SpectralCentroid and Fade

[yoyololicon]

I don't think Mulaw*coding needs autograd cuz it quantized value to integers, right?

[mthrok]

I don't think Mulaw*coding needs autograd cuz it quantized value to integers, right?

Hi @yoyololicon
Yes, you are right. Scratching them out. Thanks for the pointer.

[dhthompson]

I'm taking a look at Vol

[pavithranrao]

I will try SlidingWindowCmn Thanks!

[kiri11]

I'm going to take a stab at FrequencyMasking/TimeMasking, the only ones left!

[mthrok]

Now all the Transforms with autograd supports are properly tested. Thanks for the help!