GitLab

This is a part of PRs to add new "sox_io" backend. #726 and depends on #718 and #728 .

This PR adds `load` function to "sox_io" backend, which is  tested on the following audio formats;
 - `wav`
 - `mp3`
 - `flac`
 - `ogg/vorbis` *

By default, "sox_io" backend returns Tensor with `float32` dtype and the shape of `[channel, time]`. The samples are normalized to fit in the range of `[-1.0, 1.0]`.

Unlike existing "sox" backend, the new `load` function can handle WAV file natively, when the input format is WAV with integer type, (such as 32-bit signed integer, 16-bit signed integer and 8-bit unsigned integer) by providing `normalize=False`, this function can return integer Tensor, where the samples are expressed within the whole range of the corresponding dtype, that is, `int32` tensor for `32-bit PCM`, `int16` for `16-bit PCM` and `uint8` for `8-bit PCM`. This behavior follows [scipy.io.wavfile.read](https://docs.scipy.org/doc/scipy/reference/generated/scipy.io.wavfile.read.html). `normalize` parameter has no effect for other formats and the load function always return normalized value with `float32` Tensor.

__* Note__ The current binary distribution of torchaudio does not contain `ogg/vorbis` and `opus` codecs. To handle these files, one needs to build torchaudio from the source with proper codecs in the system.

__Note 2__ Since this PR, `scipy` becomes required module for running test. 

* packaging: Add test channels to pytorch dependency resolution
Signed-off-by: Eli Uriegas <eliuriegas@fb.com>

* .circleci: Add test channel to smoke tests
Signed-off-by: Eli Uriegas <eliuriegas@fb.com>

* .circleci: Put pytorch-test into a higher priority

pytorch-nightly was getting prioritized over pytorch-nightly which
shouldn't be the case
Signed-off-by: Eli Uriegas <eliuriegas@fb.com>

Signed-off-by: Eli Uriegas <eliuriegas@fb.com>

In #728, linux unit test switches to libsox provided by apt.
For CPU jobs this is fine because all the job steps share the same Docker container,
but on CPU job, each job step runs a script in a new Docker container, so
libsox installed in a step is not available to the subsequent steps.

To fix this, this PR moves the installation of libsox and sox to Docker build.

This PR refactors cache generation mechanism by introducing dedicated command
and bust cache on daily basis.

At this moment, Windows unittest job for 3.6 and 3.7 are broken because of
broken scipy but the environment is cached this persists until the next week.

As we have nightly build, we do not need to keep cache for one week.

* add unittest in test_models

* update test method

* remove unittest main function
Co-authored-by: Ji Chen <jimchen90@devfair0160.h2.fair>

This PR fixes the wrong member name of SignalInfo introduced in #718. 

 - `num_samples` == `num_frames` * `num_channels`.

This is a part of PRs to add new "sox_io" backend #726, and depends on #718.

This PR adds `info` function to "sox_io" backend, which allows users to fetch some metadata of an audio file. 
At this moment, the information retrieved are;

 - Number of samples in the audio file
 - Sampling rate
 - Number of channels

This is a part of PRs to add new "sox_io" backend. https://github.com/pytorch/audio/pull/726

This PR adds `SignalInfo` structure, which is data exchange interface between Python and C++ in coming TorchScript-based sox IO backend.
For the case, where C++ extension is not available (i.e. Windows), this PR also adds dummy class and module that will be substituted.
This logic is implemented in `torchaudio.extension` moduel.

* Make tests backend aware by introducing TorchaudioTestCase and reset backend for each TestCase.

* Set backends for the test cases that require specific backend.

* spacing and typing adjustment

* remove import list
Co-authored-by: Ji Chen <jimchen90@devfair0160.h2.fair>

* Add MelResNet Block

* add default value

* update model and test

* rebase and small changes

* add pad variable

* update format

* update reference in docstrings

* add underscore name
Co-authored-by: Ji Chen <jimchen90@devfair0160.h2.fair>

Co-authored-by: Ji Chen <jimchen90@devfair0160.h2.fair>

* Refactor backend switching

1. Do not rely on global variables for backend switch
   So that load/save/info/load_wav functions will be torchscript-able
2. Add no_backend module to for the case there is no backend module available
   [bonus] This allows the whole codebase importable on systems that do not have torchaudio C++ extension nor soundfile.

`type` used in `common_utils` generates test class definition in `common_utils` and
this modifies the module state after it's imported. This is anti-pattern.
This PR get rid of the related utility functions and define test suite manually.

* Change parameterized testing system to be compatible with unittest

Summary: The previous implementation of parameterized testing worked by modifying test.common_utils inplace.  This doesn't work in general because unittest's contract with test modules is such that it must be able to load the module and run the test itself.  Because the previous implementation needed to load the module and modify it, it is incompatible.

Reviewed By: mthrok

Differential Revision: D21964676
Co-authored-by: Ben Mehne <bmehne@fb.com>

* Add cmu_arctic dataset

* add dataset name

* update audio test file with whitenoise.wav file

* add test text file

* update text method and file name

* update comment

* change datasets order in doc

* add line length
Co-authored-by: Ji Chen <jimchen90@devfair0160.h2.fair>

There are couple of aspects of this PR that overall improves the maintainability of the code base, based on "decoupling" and "separation of concerns".

First, `sox_effects` functionalities can be either available or unavailable. From the viewpoint of `torchaudio` main module, the looser the connection between the `torchaudio` module and `torchaudio.sox_effects`, the more manageable the code base become because you can change the two modules independently. This is mostly accomplished when the definitions of `initialize_sox` and `shutdown_sox` were moved from `torchaudio.__init__` to `torchaudio.sox_effects`, but the initialization of sox effects are still happening in `torchaudio.__init__`. If we move the initialization to `sox_effects` module, the responsibility of sox initialization is moved to `sox_effects` module, along with the required module availability check etc. The main `torchaudio` module can be carefree about how the `sox_effects` module should work.

In addition to that, I found that `initialize_sox` and `shutdown_sox` are confusing because it sound like they are required for `libsox` based I/O. To make it clear, I renamed them to include `sox_effect` in function name.

Also moving functions from the original places are BC breaking itself, therefore, these functions are re-imported in `torchaudio.__init__` and renamed to match the original names. Therefore the PR is not BC breaking.

It is strange that `sox_backend` depends on custom classes defined in `soudfile_backend`. This PR introduces `backend.common` module where these custom classes are defined and have both `sox_backend` and `soundfile_backend` depend on it.

* Make `_generate_wave_table` device and dtype aware.
Signed-off-by: Bhargav Kathivarapu <bhargavkathivarapu31@gmail.com>

* kaldi compliance files cleanup for spec, fbank, mfcc

* kaldi compliance tests removal for spec, fbank, mfcc
Signed-off-by: Bhargav Kathivarapu <bhargavkathivarapu31@gmail.com>

* also deletes duplicated func

For better cohesion and coupling, move sox_effects related functions to sox_effects module.
Then perform initialization in __init__.py.

See the following for cohesion and coupling.
https://en.wikipedia.org/wiki/Coupling_(computer_programming)

* Replace `backed_guard` with `requires_module`

* Remove backend_guard

* Migrate spectrogram

* Update spectrogram in kaldi.py to support device and dtype

* Remove failing tests
Signed-off-by: Bhargav Kathivarapu <bhargavkathivarapu31@gmail.com>

* Add version compatibility matrix