Griffin-Lim Transformation Implementation

[jaeyeun97]

Implementation of Griffin-Lim Transformation for feature request #351
@vincentqb I am not quite sure how I might go about writing the test code for this.
Meanwhile, I have a simple comparison of the waveforms of the steam-train-whistle-daniel_simon.mp3 file, which by the looks and sounds seem to be working properly.

[vincentqb]

Thanks for submitting the PR!

• We put the computation part of a transform in torchaudio/functional.py and then make the code in torchaudio/transforms.py just a wrapper around the functional.
• torchaudio's code can be compared against librosa's output. And torchaudio's code is also tested for jitability. See examples in test/test_functional.py and test/test_transforms.py.
• Python 2 support will be dropped after the upcoming release, so this pull request only needs to support Python 3.

[vincentqb]

To make sure the code works as expected, the following test will need to be added.

• compare against librosa's output (e.g. here), or a fixed example (say by exporting the output from librosa and loading in test).
• jitability, e.g. here

See other examples in test/test_functional.py and test/test_transforms.py.

[jaeyeun97]

@vincentqb Since torchaudio.functional.istft is not jitable, neither is this one--should I omit the latter test?

EDIT: made istft jitable instead :)

[jaeyeun97]

@vincentqb As a part of implementing jitability for GriffinLim, I ended up writing jitability for istft as well. It passes the tests on my machine, but please double check?

[vincentqb]

Is there a reason to create the tensor and then move to a different device?

[jaeyeun97]

It has to do with the JITability of the function: the signature of torch.ones seem not to allow for device and dtype setting when the out parameter is not specified. A workaround was to create the tensor then to move it.

[vincentqb]

The ones and eye tensor can be created on device directly. No need for two separate steps.

[jaeyeun97]

The last time I checked, the function signatures required an out parameter when specifying the device. Since out needed to be of type Tensor and window is of Optional[Tensor], it complained. I might be misinterpreting the errors, though.

[vincentqb]

@vincentqb As a part of implementing jitability for GriffinLim, I ended up writing jitability for istft as well. It passes the tests on my machine, but please double check?

Thanks! jitability test looks good to me! and thanks for taking care of istft.

[vincentqb]

LGTM. Waiting on test to run, and will merge.

• I removed the extra spectrogram step in test, and used smaller error bound.
• I matched the signature of the transform to the docstring.