GitLab

Summary:
This commit adds file-like object support to Streaming API.

## Features
- File-like objects are expected to implement `read(self, n)`.
- Additionally `seek(self, offset, whence)` is used if available.
- Without `seek` method, some formats cannot be decoded properly.
  - To work around this, one can use the existing `decoder` option to tell what decoder it should use.
  - The set of `decoder` and `decoder_option` arguments were added to `add_basic_[audio|video]_stream` method, similar to `add_[audio|video]_stream`.
  - So as to have the arguments common to both audio and video in front of the rest of the arguments, the order of the arguments are changed.
  - Also `dtype` and `format` arguments were changed to make them consistent across audio/video methods.

## Code structure

The approach is very similar to how file-like object is supported in sox-based I/O.
In Streaming API if the input src is string, it is passed to the implementation bound with TorchBind,
if the src has `read` attribute, it is passed to the same implementation bound via PyBind 11.

![Untitled drawing](https://user-images.githubusercontent.com/855818/169098391-6116afee-7b29-460d-b50d-1037bb8a359d.png)

## Refactoring involved
- Extracted to https://github.com/pytorch/audio/issues/2402
  - Some implementation in the original TorchBind surface layer is converted to Wrapper class so that they can be re-used from PyBind11 bindings. The wrapper class serves to simplify the binding.
  - `add_basic_[audio|video]_stream` methods were removed from C++ layer as it was just constructing string and passing it to `add_[audio|video]_stream` method, which is simpler to do in Python.
  - The original core Streamer implementation kept the use of types in `c10` namespace minimum. All the `c10::optional` and `c10::Dict` were converted to the equivalents of `std` at binding layer. But since they work fine with PyBind11, Streamer core methods deal them directly.

## TODO:
- [x] Check if it is possible to stream MP4 (yuv420p) from S3 and directly decode (with/without HW decoding).

Pull Request resolved: https://github.com/pytorch/audio/pull/2400

Reviewed By: carolineechen

Differential Revision: D36520073

Pulled By: mthrok

fbshipit-source-id: a11d981bbe99b1ff0cc356e46264ac8e76614bc6

Summary:
After https://github.com/pytorch/audio/issues/2395, build_doc job is exceeding default no-output-timeout
threshould (10m).

This commit updates the timeout threshold to 30m.
Also it moves the installation of tools to the previous step.

Pull Request resolved: https://github.com/pytorch/audio/pull/2399

Reviewed By: carolineechen

Differential Revision: D36539022

Pulled By: mthrok

fbshipit-source-id: 391764a0fb5bf87cfb2beaab401a90dcb56493e5

Summary:
Pull Request resolved: https://github.com/pytorch/audio/pull/2392

Refactors LibriSpeech tests to accommodate different dataset classes

Reviewed By: xiaohui-zhang

Differential Revision: D36387835

fbshipit-source-id: 73b4e7565b4a077b25f036f4bd854ac7f2194b28

Summary:
This commit adds tutorial to enable/use NVDEC with Stream API.

https://output.circle-artifacts.com/output/job/19e66a4b-1819-4804-8834-d38e6c80c4fd/artifacts/0/docs/hw_acceleration_tutorial.html

Because the use of NVDEC requires build / install FFmpeg from source,
this tutorial was authored on Google Colab, tailored to its environment.

The tutorial here is the result of the notebook execution, with
the link to the publicly accessible Google Colab notebook.

Pull Request resolved: https://github.com/pytorch/audio/pull/2393

Reviewed By: hwangjeff

Differential Revision: D36404408

Pulled By: mthrok

fbshipit-source-id: 9c820d3db4d06c5b343ecad0708489125ca06948

Summary:
To resolve nightly / general build issues relating to OpenMP not being found, see https://hud.pytorch.org/pytorch/audio/commit/c6a376cc5679c1940e49fc3e0ba22eaead6c2467



```
-- Found Torch: /Users/distiller/miniconda3/envs/env3.10/lib/python3.10/site-packages/torch/lib/libtorch.dylib
CMake Error at /Users/distiller/miniconda3/envs/env3.10/lib/python3.10/site-packages/cmake/data/CMake.app/Contents/share/cmake-3.22/Modules/FindPackageHandleStandardArgs.cmake:230 (message):
  Could NOT find OpenMP_C (missing: OpenMP_C_FLAGS OpenMP_C_LIB_NAMES)
Call Stack (most recent call first):
  /Users/distiller/miniconda3/envs/env3.10/lib/python3.10/site-packages/cmake/data/CMake.app/Contents/share/cmake-3.22/Modules/FindPackageHandleStandardArgs.cmake:594 (_FPHSA_FAILURE_MESSAGE)
  /Users/distiller/miniconda3/envs/env3.10/lib/python3.10/site-packages/cmake/data/CMake.app/Contents/share/cmake-3.22/Modules/FindOpenMP.cmake:544 (find_package_handle_standard_args)
  CMakeLists.txt:131 (find_package)

-- Configuring incomplete, errors occurred!
```
Signed-off-by: Eli Uriegas <eliuriegas@fb.com>

Pull Request resolved: https://github.com/pytorch/audio/pull/2404

Reviewed By: atalman

Differential Revision: D36495791

Pulled By: seemethere

fbshipit-source-id: 7b6fa2a62fda6fc468cfcbdf8d2163e6b9c327b0

Summary:
* Move the helper wrapping code in TorchBind layer to proper wrapper class for so that it will be re-used in PyBind11.
* Move `add_basic_[audio|video]_stream` methods from C++ to Python, as they are just string manipulation. This will make PyBind11-based binding simpler as it needs not to deal with dtype.
* Move `add_[audio|video]_stream` wrapper signature to Streamer core, so that Streamer directly deals with `c10::optional`.†

† Related to this, there is a slight change in how the empty filter expression is stored. Originally, if an empty filter expression was given to `add_[audio|video]_stream` method, the `StreamReaderOutputStream` was showing it as empty string `""`, even though internally it was using `"anull"` or `"null"`. Now `StreamReaderOutputStream` shows the corresponding filter expression that is actually being used.

Ref https://github.com/pytorch/audio/issues/2400

Pull Request resolved: https://github.com/pytorch/audio/pull/2402

Reviewed By: nateanl

Differential Revision: D36488808

Pulled By: mthrok

fbshipit-source-id: 877ca731364d10fc0cb9d97e75d55df9180f2047

Summary:
In Wav2Vec2 and HuBERT model training, the convolutional feature extraction layers use `group_norm` for normalization in `Base` model, while they use `layer_norm` in `Large` and `XLarge` models. For `Base` model, the gradients of feature extraction layers will be unstable in pre-training, thus we need to scale down the gradient by multiplying 0.1.

In this PR, we add such argument to `HuBERTPretrainModel` to control the gradient of feature extractor layers. We also put the argument in the factory functions (`hubert_pretrain_base`, `hubert_pretrain_large`, and `hubert_pretrain_xlarge`. The reason is in finetuning, the feature extractor's parameters are fixed, we can multiply the gradient with 0.0 to avoid back propagating gradients.

Pull Request resolved: https://github.com/pytorch/audio/pull/2335

Reviewed By: xiaohui-zhang, mthrok

Differential Revision: D35646928

Pulled By: nateanl

fbshipit-source-id: 6a9563e227aac6e3127b634357946d860f26c994

Summary:
This commit updates the `window.sideMenus.handleRightMenu`, so that
subsections are expanded on tutorials by default.

https://output.circle-artifacts.com/output/job/98508917-87df-4666-9958-c70683b3245d/artifacts/0/docs/tutorials/audio_io_tutorial.html

Tutorial subsections are important because they have anchors so
allow us to get the link to the specific figures / audio samples.

When responding issues/questions and when there is a corresponding
code snippet in tutorial, it is often easy to answer with links to
the tutorial.

However, by default the tutorial page collapses right side bar, and
I have to click the small "+" symbols to navigate to the subsection,
and the state of expansion does not persist across the page refresh.

This has been a pain point since we updated the Sphinx version to 3
in https://github.com/pytorch/audio/pull/1685.

Pull Request resolved: https://github.com/pytorch/audio/pull/2397

Reviewed By: xiaohui-zhang

Differential Revision: D36429745

Pulled By: mthrok

fbshipit-source-id: 97a5ae9270e68f8e88f0bca766d5a2c1839634e3

Summary:
This commit moves `build_doc` job to run on top of Conda binary
build job.

The motivation is that Conda provides easy access to third party
tools that are required to build complex documentation.

Specifically in https://github.com/pytorch/audio/pull/2393,
ipynb-style tutorial is being added, which requires `nbsphinx`.

`nbsphinx` requires `pandoc` package and there was some issue
with the version from PyPI. A workaround is to use the one from
Conda package.

Pull Request resolved: https://github.com/pytorch/audio/pull/2395

Reviewed By: carolineechen, nateanl

Differential Revision: D36404407

Pulled By: mthrok

fbshipit-source-id: 26ec5ebfd5be795384306a9f24817a2eb3ec96c1

Summary:
Applies new import merging and sorting from µsort v1.0.

When merging imports, µsort will make a best-effort to move associated
comments to match merged elements, but there are known limitations due to
the diynamic nature of Python and developer tooling. These changes should
not produce any dangerous runtime changes, but may require touch-ups to
satisfy linters and other tooling.

Note that µsort uses case-insensitive, lexicographical sorting, which
results in a different ordering compared to isort. This provides a more
consistent sorting order, matching the case-insensitive order used when
sorting import statements by module name, and ensures that "frog", "FROG",
and "Frog" always sort next to each other.

For details on µsort's sorting and merging semantics, see the user guide:
https://usort.readthedocs.io/en/stable/guide.html#sorting

Reviewed By: lisroach

Differential Revision: D36402214

fbshipit-source-id: b641bfa9d46242188524d4ae2c44998922a62b4c

Summary:
Refactors `librispeech.py` to clarify its logic.

Pull Request resolved: https://github.com/pytorch/audio/pull/2387

Reviewed By: nateanl

Differential Revision: D36359176

Pulled By: hwangjeff

fbshipit-source-id: 595dd1421738279896348448dd72ca57bfe7cef2

Summary:
This commit moves the Streaming API out of prototype module.

* The related classes are renamed as following

  - `Streamer` -> `StreamReader`.
  - `SourceStream` -> `StreamReaderSourceStream`
  - `SourceAudioStream` -> `StreamReaderSourceAudioStream`
  - `SourceVideoStream` -> `StreamReaderSourceVideoStream`
  - `OutputStream` -> `StreamReaderOutputStream`

This change is preemptive measurement for the possibility to add
`StreamWriter` API.

* Replace BUILD_FFMPEG build arg with USE_FFMPEG

We are not building FFmpeg, so USE_FFMPEG is more appropriate

 ---

After https://github.com/pytorch/audio/issues/2377

Remaining TODOs: (different PRs)
- [ ] Introduce `is_ffmpeg_binding_available` function.
- [ ] Refactor C++ code:
   - Rename `Streamer` to `StreamReader`.
   - Rename `streamer.[h|cpp]` to `stream_reader.[h|cpp]`.
   - Rename `prototype.cpp` to `stream_reader_binding.cpp`.
   - Introduce `stream_reader` directory.
- [x] Enable FFmpeg in smoke test (https://github.com/pytorch/audio/issues/2381)

Pull Request resolved: https://github.com/pytorch/audio/pull/2378

Reviewed By: carolineechen

Differential Revision: D36359299

Pulled By: mthrok

fbshipit-source-id: 6a57b702996af871e577fb7addbf3522081c1328

Summary:
This commit updates the lazy module initialization logic for
`torchaudio.prototype.io` and `torchaudio.prototype.ctc_decoder`.

- The modules are importable regarless of optional dependencies.
i.e. `import torchaudio.prototype.io` does not trigger the check for
optional dependencies.

- Optional dependencies are checked when the actual
API is imported for the first time.
i.e. `from torchaudio.prototype.io import Streamer` triggers the check
for optional dependencies.

The downside is that;

- `import torchaudio.prototype.io.Streamer` no longer works.

## Details:

Starting from Python 3.7, modules can bave `__getattr__` function,
which serves as a fallback if the import mechanism cannot find the
attribute.

This can be used to implement lazy import.

```python
def __getattr__(name):
    global pi
    if name == 'pi':
        import math
        pi = math.pi
        return pi
    raise AttributeError(...)
```

Ref: https://twitter.com/raymondh/status/1094686528440168453

The implementation performs lazy import for the APIs that work with
external/optional dependencies. In addition, it also check if the
binding is initialized only once.

## Why is this preferable approach?

Previously, the optional dependencies were checked at the tiem module
is imported;

https://github.com/pytorch/audio/blob/2f4eb4ac2f48a597825d3631a840afd855fe6b39/torchaudio/prototype/io/__init__.py#L1-L5

As long as this module is in `prototype`, which we ask users to import
explictly, users had control whether they want/do not want to install
the optional dependencies.

This approach only works for one optional dependencies per one module.
Say, we add different I/O library as an optional dependency, we need to
put all the APIs in dedicated submodule. This prevents us from having
flat namespace.
i.e. the I/O modules with multiple optional dependencies would look like

```python
# Client code
from torchaudio.io.foo import FooFeature
from torchaudio.io.bar import BarFeature
```

where the new approach would allow

```python
#client code
from torchaudio.io import FooFeature, BarFeature
```

Pull Request resolved: https://github.com/pytorch/audio/pull/2377

Reviewed By: nateanl

Differential Revision: D36305603

Pulled By: mthrok

fbshipit-source-id: c1eb6cac203f6dd0026d99f9a1de1af590a535ae

Summary:
- Use `apply_beamforming`, `rtf_evd`, `rtf_power`, `mvdr_weights_souden`, `mvdr_weights_rtf` methods under `torchaudio.functional` to replace the class methods.
- Refactor docstrings in `PSD` and `MVDR`.
- Put `_get_mvdr_vector` outside of `MVDR` class as it doesn't call self methods inside.
- Since MVDR uses einsum for matrix operations, packing and unpacking batches are not necessary. It can be tested by the [batch_consistency_test](https://github.com/pytorch/audio/blob/main/test/torchaudio_unittest/transforms/batch_consistency_test.py#L202). Removed it from the code.

Pull Request resolved: https://github.com/pytorch/audio/pull/2383

Reviewed By: carolineechen, mthrok

Differential Revision: D36338373

Pulled By: nateanl

fbshipit-source-id: a48a6ae2825657e5967a19656245596cdf037c5f

Summary:
- When cropping the waveform and corresponding label, we use the formula `torch.div(audio_start - kernel_size * sample_rate, stride * sample_rate, rounding_mode="floor")` to align the audio start and label start indices. However, sometimes the value can be negative, which result in an empty label. The training example will hurt the performance after zero-padding (i.e., the labels are all zero for the input waveform).
This PR fixes the bug by checking if `label_start` is negative, and change it to zero if so.
- If `pad` is True, the `length` should be the length of each waveform instead of the max length. Fix it to make the model ignore the padding component in pre-training.

Pull Request resolved: https://github.com/pytorch/audio/pull/2296

Reviewed By: mthrok

Differential Revision: D36323217

Pulled By: nateanl

fbshipit-source-id: 1ffa71e39bbc0e8dee55c3b829911bc2e785b423

Summary:
Applies the black-fbsource codemod with the new build of pyfmt.

paintitblack

Reviewed By: lisroach

Differential Revision: D36324783

fbshipit-source-id: 280c09e88257e5e569ab729691165d8dedd767bc

Summary:
Conda package build performs simple smoke test, which is different
from smoke_test jobs we define on our CI jobs.

Currently Conda packaging smoke test verifies the imporatability of
`torchaudio.prototype.io`, which requires FFmpeg 4.

1. We list FFmpeg 4 as runtime requirements, but this means that
conda's dependency resolver takes FFmpeg 4 into consideration.
FFmpeg 5 was release this year, and we can expect that user base
will move to FFmpeg gradually. If user environment has some constraint
on FFmpeg, torchaudio will have conflict and it will prevent users
from install torchaudio.

2. In #2377 the way optional dependency is checked/initialized is changed,
so this Conda smoke test will no longer check the integrity with FFmpeg libraries.

To solve the issues above, this commit moves the part that tests integrity with
FFmpeg libraries to the smoke test we define on CircleCI.

Pull Request resolved: https://github.com/pytorch/audio/pull/2381

Reviewed By: carolineechen

Differential Revision: D36323706

Pulled By: mthrok

fbshipit-source-id: 57ca816e0f3ad8e16d21e56062f6ed8a09ab93a3

Summary:
The modules include:
- PSD
- MVDR
- RTFMVDR
- SoudenMVDR

Pull Request resolved: https://github.com/pytorch/audio/pull/2382

Reviewed By: carolineechen

Differential Revision: D36314096

Pulled By: nateanl

fbshipit-source-id: 9d7d962b1c70cdc435a579191ad88838dd6fc0ba

Summary:
Since a while ago, CodeQL is always emitting red signal, but the team
does not know what this is / how to fix this. At this point, it is
purely noise while not providing a valuable signal.

Ref https://github.com/pytorch/audio/issues/2314

Pull Request resolved: https://github.com/pytorch/audio/pull/2380

Reviewed By: carolineechen

Differential Revision: D36305599

Pulled By: mthrok

fbshipit-source-id: 27ece58730066543600f3873397b9a239e54beb0

Summary:
On CircleCI, Windows unittests are failing for Python 3.7 with
`PermissionError` at the end of test when it cleans up temporary
directory.

According to the discussion https://github.com/python/cpython/issues/74168,
this is caused by a known issue with `shutil.rmtree`.

In the above thread it is advised to simply ignore the error as it
is not guaranteed that temp directories are cleaned up.

This commit follows the same path and simply ignore the error
so that our CI gets back to green.

Pull Request resolved: https://github.com/pytorch/audio/pull/2379

Reviewed By: carolineechen

Differential Revision: D36305595

Pulled By: mthrok

fbshipit-source-id: d9049c2ee3447712119786311f639a1f9f8911c5

Summary:
Modifies the example LibriSpeech Conformer RNN-T recipe as follows:
- Moves data loading and transforms logic from lightning module to data module (improves generalizability and reusability of lightning module and data module).
- Moves transforms logic from dataloader collator function to dataset (resolves dataloader multiprocessing issues on certain platforms).
- Replaces lambda functions with `partial` equivalents (resolves pickling issues in certain runtime environments).
- Modifies training script to allow for specifying path model checkpoint to restart training from.

Pull Request resolved: https://github.com/pytorch/audio/pull/2366

Reviewed By: mthrok

Differential Revision: D36305028

Pulled By: hwangjeff

fbshipit-source-id: 0b768da5d5909136c55418bf0a3c2ddd0c5683ba

Summary:
This commit refactor the constructor of wrapper classes so that
wrapper classes are only responsible for deallocation of underlying
FFmpeg custom structures.

The responsibility of custom initialization is moved to helper functions.

Context:

FFmpeg API uses bunch of raw pointers, which require dedicated allocater
and deallcoator. In torchaudio we wrap these pointers with
`std::unique_ptr<>` to adopt RAII semantics.

Currently all of the customization logics required for `Streamer` are
handled by the constructor of wrapper class. Like the following;

```
AVFormatContextPtr(
      const std::string& src,
      const std::string& device,
      const std::map<std::string, std::string>& option);
```

This constructor allocates the raw `AVFormatContext*` pointer,
while initializing it with the given option, then it parses the
input media.

As we consider the write/encode features, which require different way
of initializing the `AVFormatContext*`, making it the responsibility
of constructors of `AVFormatContextPtr` reduce the flexibility.

Thus this commit moves the customization to helper factory function.

- `AVFormatContextPtr(...)` -> `get_input_format_context(...)`
- `AVCodecContextPtr(...)` -> `get_decode_context(...)`

Pull Request resolved: https://github.com/pytorch/audio/pull/2373

Reviewed By: hwangjeff

Differential Revision: D36230148

Pulled By: mthrok

fbshipit-source-id: 202d57d549223904ee958193f3b386ef5a9cda3a

Summary:
Adds an implementation of the convolution-augmented streaming transformer (effectively Emformer with convolution block) described in https://arxiv.org/abs/2110.05241.

Continuation of https://github.com/pytorch/audio/issues/2324.

Pull Request resolved: https://github.com/pytorch/audio/pull/2358

Reviewed By: nateanl, xiaohui-zhang

Differential Revision: D36137992

Pulled By: hwangjeff

fbshipit-source-id: 9c7a7c233944fe9ef15b9ba397d7f0809da1f063

Summary:
Address https://github.com/pytorch/audio/issues/2375
The MVDR module internally transforms the dtype of complex tensors to `torch.complex128` for computation and transforms it back to the original dtype before returning the Tensor. However, it didn't convert back successfully due to `specgram_enhanced.to(dtype)`, which should be `specgram_enhanced = specgram_enhanced.to(dtype)`. Fix it to make the output dtype consistent with original input.

Pull Request resolved: https://github.com/pytorch/audio/pull/2376

Reviewed By: hwangjeff

Differential Revision: D36280851

Pulled By: nateanl

fbshipit-source-id: 553d1b98f899547209a4e3ebc59920c7ef1f3112

Summary:
previous update for rocm: https://github.com/pytorch/audio/pull/2186

Pull Request resolved: https://github.com/pytorch/audio/pull/2362

Reviewed By: seemethere

Differential Revision: D36283672

Pulled By: mthrok

fbshipit-source-id: bfd38940d027c8ccd72ab48991e5ab7f84b0e9c0

Summary:
Add a new design of MVDR module.
The RTFMVDR module supports the method based on the relative transfer function (RTF) and power spectral density (PSD) matrix of noise.
The input arguments are:
- multi-channel spectrum.
- RTF vector of the target speech
- PSD matrix of noise.
- reference channel in the microphone array.
- diagonal_loading option to enable or disable diagonal loading in matrix inverse computation.
- diag_eps for computing the inverse of the matrix.
- eps for computing the beamforming weight.
The output of the module is the single-channel complex-valued spectrum for the enhanced speech.

Pull Request resolved: https://github.com/pytorch/audio/pull/2368

Reviewed By: carolineechen

Differential Revision: D36214940

Pulled By: nateanl

fbshipit-source-id: 5f29f778663c96591e1b520b15f7876d07116937

Summary:
When computing the MVDR beamforming weights using the power iteration method, the PSD matrix of noise can be applied with diagonal loading to improve the robustness. This is also applicable to computing the RTF matrix (See https://github.com/espnet/espnet/blob/master/espnet2/enh/layers/beamformer.py#L614 as an example). This also aligns with current `torchaudio.transforms.MVDR` module to keep the consistency.

This PR adds the `diagonal_loading` argument with `True` as default value to `torchaudio.functional.rtf_power`.

Pull Request resolved: https://github.com/pytorch/audio/pull/2369

Reviewed By: carolineechen

Differential Revision: D36204130

Pulled By: nateanl

fbshipit-source-id: 93a58d5c2107841a16c4e32f0c16ab0d6b2d9420

Summary:
Add a new design of MVDR module.
The `SoudenMVDR` module supports the method proposed by [Souden et, al.](http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.725.673&rep=rep1&type=pdf).
The input arguments are:
- multi-channel spectrum.
- PSD matrix of target speech.
- PSD matrix of noise.
- reference channel in the microphone array.
- diagonal_loading option to enable or disable diagonal loading in matrix inverse computation.
- diag_eps for computing the inverse of the matrix.
- eps for computing the beamforming weight.

The output of the module is the single-channel complex-valued spectrum for the enhanced speech.

Pull Request resolved: https://github.com/pytorch/audio/pull/2367

Reviewed By: hwangjeff

Differential Revision: D36198015

Pulled By: nateanl

fbshipit-source-id: 4027f4752a84aaef730ef3ea8c625e801cc35527

Summary:
This commits add `hw_accel` option to `Streamer::add_video_stream` method.
Specifying `hw_accel="cuda"` allows to create the chunk Tensor directly from CUDA,
when the following conditions are met.
1. the video format is H264,
2. underlying ffmpeg is compiled with NVENC, and
3. the client code specifies `decoder="h264_cuvid"`.

A simple benchmark yields x7 improvement in the decoding speed.

<details>

```python
import time

from torchaudio.prototype.io import Streamer

srcs = [
    "https://download.pytorch.org/torchaudio/tutorial-assets/stream-api/NASAs_Most_Scientifically_Complex_Space_Observatory_Requires_Precision-MP4.mp4",
    "./NASAs_Most_Scientifically_Complex_Space_Observatory_Requires_Precision-MP4.mp4",  # offline version
]

patterns = [
    ("h264_cuvid", None, "cuda:0"),  # NVDEC on CUDA:0 -> CUDA:0
    ("h264_cuvid", None, "cuda:1"),  # NVDEC on CUDA:1 -> CUDA:1
    ("h264_cuvid", None, None),  # NVDEC -> CPU
    (None, None, None),  # CPU
]

for src in srcs:
    print(src, flush=True)
    for (decoder, decoder_options, hw_accel) in patterns:
        s = Streamer(src)
        s.add_video_stream(5, decoder=decoder, decoder_options=decoder_options, hw_accel=hw_accel)

        t0 = time.monotonic()
        num_frames = 0
	for i, (chunk, ) in enumerate(s.stream()):
	    num_frames += chunk.shape[0]
        t1 = time.monotonic()
        print(chunk.dtype, chunk.shape, chunk.device)
        print(time.monotonic() - t0, num_frames, flush=True)
```
</details>

```
https://download.pytorch.org/torchaudio/tutorial-assets/stream-api/NASAs_Most_Scientifically_Complex_Space_Observatory_Requires_Precision-MP4.mp4
torch.uint8 torch.Size([5, 3, 1080, 1920]) cuda:0
10.781158386962488 6175
torch.uint8 torch.Size([5, 3, 1080, 1920]) cuda:1
10.771313901990652 6175
torch.uint8 torch.Size([5, 3, 1080, 1920]) cpu
27.88662809302332 6175
torch.uint8 torch.Size([5, 3, 1080, 1920]) cpu
83.22728440898936 6175
./NASAs_Most_Scientifically_Complex_Space_Observatory_Requires_Precision-MP4.mp4
torch.uint8 torch.Size([5, 3, 1080, 1920]) cuda:0
12.945253834011964 6175
torch.uint8 torch.Size([5, 3, 1080, 1920]) cuda:1
12.870224556012545 6175
torch.uint8 torch.Size([5, 3, 1080, 1920]) cpu
28.03406483103754 6175
torch.uint8 torch.Size([5, 3, 1080, 1920]) cpu
82.6120332319988 6175
```

With HW resizing

<details>

```python
import time

from torchaudio.prototype.io import Streamer

srcs = [
    "./NASAs_Most_Scientifically_Complex_Space_Observatory_Requires_Precision-MP4.mp4",
    "https://download.pytorch.org/torchaudio/tutorial-assets/stream-api/NASAs_Most_Scientifically_Complex_Space_Observatory_Requires_Precision-MP4.mp4",
]

patterns = [
    # Decode with NVDEC, CUDA HW scaling -> CUDA:0
    ("h264_cuvid", {"resize": "960x540"}, "", "cuda:0"),
    # Decoded with NVDEC, CUDA HW scaling -> CPU
    ("h264_cuvid", {"resize": "960x540"}, "", None),
    # CPU decoding, CPU scaling
    (None, None, "scale=width=960:height=540", None),
]

for src in srcs:
    print(src, flush=True)
    for (decoder, decoder_options, filter_desc, hw_accel) in patterns:
        s = Streamer(src)
        s.add_video_stream(
            5,
            decoder=decoder,
            decoder_options=decoder_options,
            filter_desc=filter_desc,
            hw_accel=hw_accel,
        )

        t0 = time.monotonic()
        num_frames = 0
        for i, (chunk, ) in enumerate(s.stream()):
            num_frames += chunk.shape[0]
        t1 = time.monotonic()
        print(chunk.dtype, chunk.shape, chunk.device)
        print(time.monotonic() - t0, num_frames, flush=True)
```

</details>

```
./NASAs_Most_Scientifically_Complex_Space_Observatory_Requires_Precision-MP4.mp4
torch.uint8 torch.Size([5, 3, 540, 960]) cuda:0
12.890056837990414 6175
torch.uint8 torch.Size([5, 3, 540, 960]) cpu
10.697489063022658 6175
torch.uint8 torch.Size([5, 3, 540, 960]) cpu
85.19899423001334 6175

https://download.pytorch.org/torchaudio/tutorial-assets/stream-api/NASAs_Most_Scientifically_Complex_Space_Observatory_Requires_Precision-MP4.mp4
torch.uint8 torch.Size([5, 3, 540, 960]) cuda:0
10.712715593050234 6175
torch.uint8 torch.Size([5, 3, 540, 960]) cpu
11.030170071986504 6175
torch.uint8 torch.Size([5, 3, 540, 960]) cpu
84.8515750519582 6175
```

Pull Request resolved: https://github.com/pytorch/audio/pull/2331

Reviewed By: hwangjeff

Differential Revision: D36217169

Pulled By: mthrok

fbshipit-source-id: 7979570b083cfc238ad4735b44305d8649f0607b

Summary: Pull Request resolved: https://github.com/pytorch/audio/pull/2371

Reviewed By: xiaohui-zhang

Differential Revision: D36246167

Pulled By: carolineechen

fbshipit-source-id: 23042a1c393711864a18c9815d248c18d1d258b4

Summary:
Cleanup old version of cuda115 and other legacy versions

Pull Request resolved: https://github.com/pytorch/audio/pull/2374

Reviewed By: nateanl, mthrok

Differential Revision: D36250955

Pulled By: atalman

fbshipit-source-id: 6b7f0e2926eeb688991c939901c980428cf8e7ef

Summary:
This commit changes the way torchaudio binary distributions are built.

* For all the binary distributions (conda/pip on Linux/macOS/Windnows), build custom FFmpeg libraries.
* The custom FFmpeg libraries do not use `--use-gpl` nor `--use-nonfree`, so that they stay LGPL.
* The custom FFmpeg libraries employ rpath so that the torchaudio binary distributions look for the corresponding FFmpeg libraries installed in the runtime environment.
* The torchaudio binary build process will use them to bootstrap its build process.
* The custom FFmpeg libraries are NOT shipped.

This commit also add disclaimer about FFmpeg in README.

Pull Request resolved: https://github.com/pytorch/audio/pull/2355

Reviewed By: nateanl

Differential Revision: D36202087

Pulled By: mthrok

fbshipit-source-id: c30e5222ba190106c897e42f567cac9152dbd8ef

Summary:
The smoke test jobs simply perform `import torchaudio` to check
if the package artifacts are sane.

Originally, the CI was executing it in the root directory.
This was fine unless the source code is checked out.
When source code is checked out, performing `import torchaudio` in
root directory would import source torchaudio directory, instead of the
installed package.

This error is difficult to notice, so this commit introduces common script to
perform the smoke test, while moving out of root directory.

Pull Request resolved: https://github.com/pytorch/audio/pull/2365

Reviewed By: carolineechen

Differential Revision: D36202069

Pulled By: mthrok

fbshipit-source-id: 4396f85fec5c54869ada4c08f51304539f1b05cf

Summary:
Currently smoke tests are only executed on nightly jobs.
This is inconvenient as PRs that changes build process do not get
the signal naturally.

This commit changes it by always executing smoke tests.

Pull Request resolved: https://github.com/pytorch/audio/pull/2364

Reviewed By: atalman

Differential Revision: D36171267

Pulled By: mthrok

fbshipit-source-id: e549965ba139b5992177b7a094d87c9ef4432a7f

Summary:
This PR fixes Windows Smoke tests

Tested via  circleci :
https://app.circleci.com/pipelines/github/pytorch/audio/10572/workflows/970fd791-25cc-4af4-8183-a7835e1891bf/jobs/637607

Pull Request resolved: https://github.com/pytorch/audio/pull/2361

Reviewed By: nateanl, mthrok

Differential Revision: D36167317

Pulled By: atalman

fbshipit-source-id: 1418ebffd74614cc1110dc032d16ee9502a7d571

Summary:
libmad integration should be enabled only from source-build

Pull Request resolved: https://github.com/pytorch/audio/pull/2354

Reviewed By: nateanl

Differential Revision: D36012035

Pulled By: mthrok

fbshipit-source-id: adeda8cbfd418f96245909cae6862b648a6915a7

Summary:
Fix audio win smoke test to use GPU hosts for CUDA builds

Pull Request resolved: https://github.com/pytorch/audio/pull/2353

Reviewed By: mthrok

Differential Revision: D36006928

Pulled By: atalman

fbshipit-source-id: a27c4cc34093810c8cc08e01188e09b474478001

Summary:
This PR amends `RNNTBeamSearch`'s streaming decoding method to correctly unsqueeze `length` when its dimension is 0.

Original comment: Is "input.dim() == 0" unreachable as it could only be 2 or 3 in assertion of Line 329?

Pull Request resolved: https://github.com/pytorch/audio/pull/2344

Reviewed By: carolineechen, nateanl

Differential Revision: D35899740

Pulled By: hwangjeff

fbshipit-source-id: 84c1692b8cc9e5d35798d87f4a1bd052d94af9fb

Summary:
Add support for lexicon free decoding based on [fairseq's](https://github.com/pytorch/fairseq/blob/main/examples/speech_recognition/new/decoders/flashlight_decoder.py#L53) implementation. Reached numerical parity with fairseq's decoder in offline experimentation

Follow ups
- Add pretrained LM support for lex free decoding
- Add example in tutorial
- Replace flashlight C++ source code with flashlight text submodule
- [optional] fairseq compatibility test

Pull Request resolved: https://github.com/pytorch/audio/pull/2342

Reviewed By: nateanl

Differential Revision: D35856104

Pulled By: carolineechen

fbshipit-source-id: b64286550984df906ebb747e82f6fb1f21948ac7

Summary:
In different pre-training and fine-tuning settings, the `mask_prob`, `mask_channel_prob`, and `mask_channel_length` are different. For example, the settings in [pre-training](https://github.com/pytorch/fairseq/blob/main/examples/hubert/config/pretrain/hubert_base_librispeech.yaml#L70) and [fine-tuning](https://github.com/pytorch/fairseq/blob/main/examples/hubert/config/finetune/base_10h.yaml#L69-L73) are different. The motivation is to avoid overfitting when fine-tuning on a small dataset (example: [fine-tune on 10 minutes of audio](https://github.com/pytorch/fairseq/blob/main/examples/wav2vec/config/finetuning/vox_10m.yaml#L57-L59)).
This PR adds the required arguments in the factory functions to make them tunable for pre-training and fine-tuning. `mask_length` is set to `10` by default for all cases, hence it's not included in the factory function.

Pull Request resolved: https://github.com/pytorch/audio/pull/2345

Reviewed By: carolineechen, xiaohui-zhang

Differential Revision: D35845117

Pulled By: nateanl

fbshipit-source-id: 0cbb74d09535d189b8258aa8ee0f88779bdb77e7