GitLab

Precomputing and caching the resampling kernel in transforms provides speed improvements for resample, but no longer handles the automatic device and dtype recognition and construction based on input waveform. This is BC-breaking if users do not manually move the transforms object to the correct device and dtype, in which case calls to resample will fail if the input waveform is on gpu, or not of float32 dtype. Precomputing the kernel additionally results in very minor precision differences from previous implementation.

It was reported in #1478 that spectrogram masking operations were done in-place and modified the original input tensors. This PR fixes this behavior and adds tests to ensure that the input tensor is not changed.

This reverts commit 8d2eeb11.

* Update spectrogram to use complex
* Update autograd test
* Update TS test
* Update librosa test

1. `F.phase_vocoder` accepts Tensor with complex dtype.
    * The implementation path has been updated from #758 so that they share the same code path by internally converting the input Tensor to complex dtype and performing all the operation on top of it.
    * Adopted `torch.polar` for simpler Tensor generation from magnitude and angle.
2. Updated tests
    * librosa compatibility test for complex dtype and pseudo complex dtype
        * Extracted the output shape check test and moved it to functional so that it will be tested on all the combination of `{CPU | CUDA} x {complex64 | complex128}`
    * TorchScript compatibility test for `F.phase_vocoder` and `T.TimeStretch`.
    * batch consistency test for `T.TimeStretch`.

This PR additionally adds batching to kaldi compliance resample interface.

* Change the name of the specgram named `waveform`

`F.sliding_window_cmn` takes a spectrogram as input (of shape
`(..., freq, time)`). However, this spectrogram is named `waveform`.
This appears to be an error, so rename this (and the output tensor) to
reflect that both are spectrograms.

* Correct tensor description in docstring

The output tensor of `F.sliding_window_cmn` is also a spectrogram.
Update the description to reflect this.

modified amplitude_to_DB to clamp per-item when a batch is provided
Co-authored-by: Vincent QB <vincentqb@users.noreply.github.com>

* Resolves https://github.com/pytorch/audio/issues/1095

* Add griffinlim and DB_to_amplitude
* Fix Dataset docstring
* Fix other formatting
Co-authored-by: krishnakalyan3 <skalyan@cloudera.com>

Co-authored-by: krishnakalyan3 <skalyan@cloudera.com>

* Suggested changes to speed up phaser

* Checkpoint

* Checkpoint

* Checkpoint

* Checkpoint

* removing todo items
Co-authored-by: Vincent QB <vincentqb@users.noreply.github.com>

* Warn if create_fb_matrix produces a column whose weights are all zeros

See also https://github.com/librosa/librosa/issues/478