GitLab

Following https://github.com/pytorch/pytorch/blob/master/docs/cpp/source/notes/inference_mode.rst#migration-guide-from-autononvariabletypemode 

Resolves #1522 

There are some issues on CMake-based integration after 2021-06-19.

```
  Imported target "torch" includes non-existent path
    "/opt/conda/conda-bld/pytorch_1624259172741/work/torch/lib"
  in its INTERFACE_INCLUDE_DIRECTORIES.  Possible reasons include:
  * The path was deleted, renamed, or moved to another location.
  * An install or uninstall procedure did not complete successfully.
  * The installation package was faulty and references files it does not provide.
```

This is being worked on PR like https://github.com/pytorch/pytorch/pull/60403.
The nightly builds are still causing the error, so as a workaround, 
we pin the nightly build version for Linux/macOS unittest.
The issue still happens on Windows unittest and binary builds.

Mel scale frequency is in general defined/used on power spectrogram and as far as we know, it is not defined for raw (complex value) spectrogram. When `power=None` (when intermediate spectrogram is complex-valued), `MelSpectrogram` fails because of extra dimension (when using pseudo complex) or type mismatch between `torch.float` and torch.cfloat` (when using native complex).

This commit changes type annotation so that power is only float.

Since 0.9.0-RC1, `T.Resample` precomputes and caches resampling kernel for performance improvement. (10x improvement).

The implementation from 0.8.0 computed the kernel on-the-fly on the same `device`/`dtype` as the input Tensor, 
but in the newer version, the kernel is precomputed at the construction time and is cached with `float32` first.
This causes degradation if one wants to perform resampling on `float64`, because `sinc` values computed on `float32`s are not good enough for resampling in `float64`.

The reason why we decided to use `float32` for initial caching is to keep the UX disruption minimum, and there were no way to make it work for `float64`. This PR adds `dtype` argument, that can be used for overwriting the cache precision.

`torchaudio.compliance.kaldi.resample_waveform` has been replaced with `torchaudio.funcitonal.resample`.

* [BC-Breaking] Default to native complex type when returning raw spectrogram

Part of https://github.com/pytorch/audio/issues/1337 .

- This code changes the return type of spectrogram to be native complex dtype,
when (and only when) returning raw (complex-valued) spectrogram.
- Change `return_complex=False` to `return_complex=True` in spectrogram ops.
- `return_complex` is only effective when `power` is `None`. It is ignored for
cases where `power` is not `None`. Because the returned Tensor is power spectrogram,
which is real-valued Tensors.

* autograd test from carolineechen/audio#2

* fix numpy backward: be careful to not modify inplace.

* Use `bibtex` for paper citations.
  * add `override.css` for fixing back reference.
  * wav2vec2
  * wav2letter
  * convtasnet
  * deepspeech
  * rnnt-loss
  * griffinlim
* Fix broken references in `filtering`.
* Fix note in soundfile backends.
* Tweak wav2vec2 example.
* Removes unused `pytorch_theme.css`

* Update compatibility matrix
* Remove some obsolete information

These arguments are deprecated in 0.8 release for legal reasons, and not used anymore. Now we can remove the arguments.

Deprecating `torchaudio.compliance.kaldi.resample_waveform`, which is exactly same as `torchaudio.functional.resample`.

* add rocm binary

use the term cu_version

update pkg_helpers

add docker wheel option

update config yml

add force_cuda option

* update config.yml

* remove conda builds

- TorchScript-able `Wav2Vec2Model` class
- Factory functions for three configurations presented in the paper 
  - `wav2vec2_base`
  - `wav2vec2_large`
  - `wav2vec2_large_lv60k`