GitLab

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.

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

* Remove `class FunctionalComplex` header accidentally re-introduced in #1490 

* Update VAD docstring and check for input shape length

* Update docstring in forward for transform

* Address review feedback: merge tests, update wording

Since the update of Xcode, some of HTTP tests are failing on macOS, due to `reqests.get` fails to establish a connection to local server. This change increase the wait time to 2 seconds before test starts.

Co-authored-by: Vincent Quenneville-Belair <vincentqb@gmail.com>

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.

Replace the prototype RNNT implementation (using warp-transducer) with one without external library dependencies

- Put functional test logic into one place, `functional_impl.py`
- Tidy imports

* Fixed floor_divide deprecation warnings seen in pytest output

* Fixed warning in test_flanger_triangle_linear

The VAD function trims the input tensor to the first instance of voice
activity on any channel or item. Trimming batches this way may be
undesirable as the item with earliest activity will dominate. Either
way, the batch behaviour does not match the itemwise behaviour.

The VAD batch consistency tests currently pass out of coincidence, but
they specify incorrect behaviour. This commit removes them.