GitLab

Summary:
Pull Request resolved: https://github.com/pytorch/audio/pull/2096

run: `arc lint --apply-patches --paths-cmd 'hg files -I "./**/*.py"'`

Reviewed By: mthrok

Differential Revision: D33297351

fbshipit-source-id: 7bf5956edf0717c5ca90219f72414ff4eeaf5aa8

Summary:
Adds implementation of Conformer module.

Adapted from sravyapopuri388's implementation for fairseq at https://github.com/fairinternal/fairseq-py/pull/2770.

Pull Request resolved: https://github.com/pytorch/audio/pull/2068

Reviewed By: mthrok

Differential Revision: D33236957

Pulled By: hwangjeff

fbshipit-source-id: 382d99394996ff5249522b5899e1a4b4a95de9e6

Summary:
## bug description

When a 24 bits-par-sample audio is loaded via file-like object,
the loaded Tensor is wrong. It was fine if the audio is loaded
from local file.

## The cause of the bug

The core of the sox's decoding mechanism is `sox_read` function,
one of which parameter is the maximum number of samples to decode
from the given buffer.

https://fossies.org/dox/sox-14.4.2/formats_8c.html#a2a4f0194a0f919d4f38c57b81aa2c06f)]

The `sox_read` function is called in what is called `drain` effect,
callback and this callback receives output buffer and its size in
byte. The previous implementation passed this size value as
the argument of `sox_read` for the maximum number of samples to
read. Since buffer size is larger than the number of samples fit in
the buffer, `sox_read` function always consumed the entire
buffer. (This behavior is not wrong except when the input is
24 bits-per-sample and file-like object.)

When the input is read from file-like object, inside of drain
callback, new data are fetched via Python's `read` method and
loaded on fixed-size memory region. The size of this memory region
can be adjusted via `torchaudio.utils.sox_utils.set_buffer_size`,
but the default value is 8096.

If the input format is 24 bits-per-sample, the end of memory region
does not necessarily correspond to the end of a valid sample.
When `sox_read` consumes all the data in the buffer region, the data
at the end introduces some unexpected values.
This causes the aforementioned bug

## Fix

Pass proper (better estimated) maximum number of samples decodable to
`sox_read`.

Pull Request resolved: https://github.com/pytorch/audio/pull/2084

Reviewed By: carolineechen

Differential Revision: D33236947

Pulled By: mthrok

fbshipit-source-id: 171d9b7945f81db54f98362a68b20f2f95bb11a4

Summary:
Our Griffin-Lim autograd tests take a long time to run. This PR adjusts some parameters to shorten the run time.

For one of the four tests:
Before:
```
test/torchaudio_unittest/transforms/autograd_cpu_test.py . [100%]

======================== 1 passed in 517.35s (0:08:37) =========================
```

After:
```
test/torchaudio_unittest/transforms/autograd_cpu_test.py . [100%]

======================== 1 passed in 104.59s (0:01:44) =========================
```

Pull Request resolved: https://github.com/pytorch/audio/pull/2037

Reviewed By: mthrok

Differential Revision: D32726213

Pulled By: hwangjeff

fbshipit-source-id: c785323ab380aea4b63fb1683b557c8ae842f54e

Summary:
Adds beam search decoder for RNN-T implementation ``torchaudio.prototype.RNNT`` that is TorchScript-able and supports both streaming and non-streaming inference.

Pull Request resolved: https://github.com/pytorch/audio/pull/2028

Reviewed By: mthrok

Differential Revision: D32627919

Pulled By: hwangjeff

fbshipit-source-id: aab99e346d6514a3207a9fb69d4b42978b4cdbbd

Summary:
The sox_effects test in `concurrent.future.ThreadPoolExecutor` started failing since couple of days. While investigate this, skipping the test.

Pull Request resolved: https://github.com/pytorch/audio/pull/2025

Reviewed By: nateanl

Differential Revision: D32615933

Pulled By: mthrok

fbshipit-source-id: 4f7301c0d3c0d11f687011e42e06d9c87ce4197f

Summary:
Allow users to use `torch.cfloat` dtype input for MVDR module. It internally convert the spectrogram into `torch.cdouble` and output the tensor with the original dtype of the spectrogram.

Pull Request resolved: https://github.com/pytorch/audio/pull/2024

Reviewed By: carolineechen

Differential Revision: D32594051

Pulled By: nateanl

fbshipit-source-id: e32609ccdc881b36300d579c90daba41c9234b46

Summary:
Division first, multiplication second. This helps avoid the value overflow issue. It also helps the ``stv_evd`` solution pass the gradient check.

Pull Request resolved: https://github.com/pytorch/audio/pull/2004

Reviewed By: mthrok

Differential Revision: D32539827

Pulled By: nateanl

fbshipit-source-id: 70a386608324bb6e1b1c7238c78d403698590f22

Summary:
Adds streaming-capable recurrent neural network transducer (RNN-T) model that uses Emformer for its transcription network. Includes two factory functions — one that allows for building a custom model, and one that builds a preconfigured base model.

Pull Request resolved: https://github.com/pytorch/audio/pull/2003

Reviewed By: nateanl

Differential Revision: D32440879

Pulled By: hwangjeff

fbshipit-source-id: 601cb1de368427f25e3b7d120e185960595d2360

Co-authored-by: Facebook Community Bot <6422482+facebook-github-bot@users.noreply.github.com>

Summary:
Pull Request resolved: https://github.com/pytorch/audio/pull/2015

as titled

Reviewed By: hwangjeff, mthrok

Differential Revision: D32495691

fbshipit-source-id: 60d8a2337585e3147f24ca9f0b6518e30cd9134a

Following the plan #1337, this commit drops the support for pseudo complex type from `F.phase_vocoder` and `T.TimeStretch`.

Following the plan #1337, this commit drops the support for pseudo complex type from 
`F.spectrogram` and `T.Spectrogram`.

It also deprecates the use of `return_complex` argument.

- Make the test support other languages
- Fetch tetst asset on-the-fly

Move the computation of `#classes -> #bits` to the constructor of WaveRNN and attach it to the instance, so that it can be reused elsewhere.

This commit merges wav2vec2/hubert factory functions for pre-training and fine-tuning. In #1829, we added parameters to customize the models that are not part of architecture, and `aux_num_out` falls into this category, so it is no longer necessary to have separate functions. This concludes the wav2vec2/HuBERT API update in release 0.10.

The summary of BC-breaking changes on wav2vec2 APIs between 0.9 and 0.10 (when this commit is incorporated)
1. `Wav2Vec2Model.extract_features`
In 0.9, it was returning the output from `FeatureExtractor` module. In 0.10, it returns the list of outputs from the intermediate layers of `TransformerEncoder` block.
2. `wav2vec2_base(num_out: int)` -> `wav2vec2_base(<dropout_params:float>, aux_num_out: Optional[int]=None)`
    - `num_out` was renamed to `aux_num_out` and optional. If it is omitted, the resulting model does not have the linear layer for fine-tuning.
    - Added dropout parameters.

This commit makes the following changes
1. Make the factory function with full customizability public.
    i.e. `_get_model(...) -> wav2vec2_model(...)`.
2. Change the other architecture-specific factory functions so that they accept parameters not related to the model architecture (such as dropout).
    i.e. `wav2vec2_base() -> wav2vec2_base(encoder_projection_dropout, encoder_attention_dropout, encoder_ff_interm_dropout, ...)`

### Why?

While adding the pre-trained weight support, I realized that separating API for model construction and pre-trained support achieves simple code organization because of the good separation of concern. As mentioned in #1821, in this framework,
  1. Model implementation is responsible for computation logic,
  2. factory functions are responsible for customizability and model construction,
  3. and pre-trained weight API is responsible for constructing a model and loading pre-trained weights along with the complementary information (such as pre-processing and class labels).

(note: for simple models, combining 1 and 2 is also okay.)

This means that factory functions has to support all the customizability required by pre-trained weight API. The current implementation uses the internal function like `from .model import Wav2Vec2Model, _get_model`, which is a bit strange.

This PR rectifies it by making the mother factory function public.
This also clarifies the purpose of having the other factory functions as public API, which is just a syntax sugar for constructing un-trained model with specific architecture. So this commit also adds supplemental parameters to them.

Adds an implementation of Emformer, a memory-efficient transformer architecture 
introduced in https://ieeexplore.ieee.org/document/9414560 that targets low-latency 
streaming speech recognition applications.

Summary:
Pull Request resolved: https://github.com/pytorch/audio/pull/1817



This changes the imports in the `torchaudio` to include the new import locations.

```
codemod -d pytorch/audio --extensions py 'torch.quantization' 'torch.ao.quantization'
```

Reviewed By: mthrok

Differential Revision: D31302450

fbshipit-source-id: f31a0d4f453f840ea690edb688555a9d585787b5
Co-authored-by: Zafar Takhirov <zaf@fb.com>

1. Fix the HuBERT xlarge model config
2. In the 48 transformer layers of HuBERT xlarge model, very few elements deviate from the equivalent model of fairseq, and exceeds the default atol 1e-5. This commit relax it to 3e-5 for the specific test.

Writing scripted HuBERT XLarge models fail on Windows CI.

* Rename factory functions `wav2vec2_asr_ARCH` to `wav2vec2_ft_ARCH`

In #1783, we split the factory functions of wav2vec2 into ones for pretraining models
and ones for fine-tuning models (pretraining model + extra Linear module).

I picked the name scheme `wav2vec2_asr_ARCH` for factory functions of fine-tuning models,
but did not feel right, because the architecture code is more generic.
Even though the resulting model architecture was used for  ASR fine-tuning in the paper, 
it does not have to be ASR.
This became more evident as we add pre-trained parameters support, such as #1799.
It matters more for the weight files that for which task and on which dataset it was
trained on. For factory function, ASR task is not relevant.

Therefore renaming the functions by replacing `_asr_` to `_ft_` fine-tuning.

Note: Since the new functions are not release yet, this PR itself is not BC-breaking.

This commit adds the following HuBERT model architectures

 - `base` (pre-training)
 - `large` (pre-training / fine-tuning)
 - `xlarge` (pre-training / fine-tuning)

Since the internal components are same as `Wav2Vec2Model`, it reuses the existing modules..
With these models, it is possible to 
- import the pre-trained model published by `fairseq` and TorchScript it.
- fine-tune the existing model for downstream task.

* [BC-Breaking] Split pretraining and finetuning factory functions

Previously, factory functions of wav2vec2 only generated the architecture
for the fine-tuning architecture used in wav2ve2 paper for ASR task.
That is, pre-training architecture + Linear module, and it did not
provide a straightforward way to generate architectures for pre-training.

The goal of the original implementation was to allow the inference of
wav2vec2 in non-Python environment via TorchScript. Now we would like to
expand it to pre-training/fine-tuning and HuBERT model as well.

Therefore, we need to have factory functions for both pre-training and
fine-tuning. This commit introduces new factory functions and separate
functions for pre-training and fine-tuning.

1. New functions for ASR fine-tuning.

We introdcue `wav2vec2_asr_XXX` functions which generates the architecture
used for the fine-tuning task in wav2vec2 paper. *1

2. Re-purpse the old functions

The existing functions, `wav2vec2_XXX`, now generates the architecture with
pre-trainig module only. (no Linear module)

Note
*1 This architecture is just one way to define architecture for fine-tuning
and it is not universal definition. The new `wav2vec2_asr_XXX` functions are
designed to provide these specific fine-tuning configuration and they are not
meant to support generic architecture for downstream task.

Previously, the Linear module (called `readout`, which is used only for an ASR fine-tuning
task) was placed in encoder module. Conceptually, the encoder has nothing to
do with a module specific to fine-tuning / downstream task.

The problems here are that;
1. encoder can be also used in pre-training phase, in which such a module should
not present
2. The choice of Linear module is arbitral, and it is inconvenient for users
to have hard-coded module structure in encoder.

Therefore, this commit moves the Linear module out the encoder, and places it
as `aux` attribute of `Wav2Vec2Model`. (as a result `Wav2Vec2Model` has
`feature_extractor`, `encoder` and `aux` attributes.)

An alternative approach is to define another module and place `Wav2Vec2Model`
and aux module along each other. But that will introduce a new class we need
to maintain.
The expected use of `aux` is only  for 1. loading the pre-trained parameters 
published by `fairseq` (and it's variations from HF) and 2. creating the same model 
architectures for comparison experiment.
The newly introduced class will not be general enough for downstream adaptations, 
where there will be a bunch of different more complicated models. (i.e. s3prl)

Therefore, based on the minimalistic approach, we put them inside of `Wav2Vec2Model`.

* Fix HF model integration

Previously, when testing wav2vec models from HF transformers, all the model were
instantiated as `Wav2Vec2ForCTC` class, while some of them were supposed to be
`Wav2Vec2Model`.

Fixing this revealed that model importer cannot correctly handle `Wav2Vec2Model` import.

This PR fixes these issues.

Summary: Update fairseq reference from master to main elsewhere

Reviewed By: alexeib

Differential Revision: D30938472

fbshipit-source-id: 243b98550207f241c9d3265bf3d4060350aaf0a8
Co-authored-by: Diana Liskovich <dianaml@fb.com>

Apply tweak around the test names so that it's easier to see which tests are failing.

Before: `test_import_finetuned_model_2`
After:  `test_import_finetuned_model_2_wav2vec2_large_lv60k`

* [BC-Breaking] Update `extract_features` of Wav2Vec2Model

Originally, `extract_features` method was returning the result from
the convolutional feature extractor module.

The features commonly used in downstream tasks are outputs from intermediate
layers of transformer block in encoder.

This commit update the behavior of `extract_features` to allow selectively
retrieve such features.