Skip to content

GitLab

Switch branch/tag
  1. 10 Aug, 2022 1 commit
  3. 05 Aug, 2022 1 commit
    • Caroline Chen's avatar
      Add note for lexicon free decoder output (#2603) · 33485b8c
      Caroline Chen authored 
      Summary:
``words`` field of CTCHypothesis is empty if no lexicon is provided, which produces confusing output (see issue https://github.com/pytorch/audio/issues/2584) when following our tutorial example with lexicon free usage. This PR adds a note in both docs and tutorial.

Followup: determine if we want to modify the behavior of ``words`` in the lexicon free case. One option is to merge and then split the generated tokens by the input silent token to populate the words field, but this is tricky since the meaning of a "word" in the lexicon free case can be vague and not all languages have whitespaces between words, etc

Pull Request resolved: https://github.com/pytorch/audio/pull/2603

Reviewed By: mthrok

Differential Revision: D38459709

Pulled By: carolineechen

fbshipit-source-id: d64ff186df4633f00e94c64afeaa6a50cebf2934
      33485b8c
  5. 01 Aug, 2022 1 commit
  7. 29 Jul, 2022 2 commits
    • moto's avatar
      Update forced alignment tutorial (#2544) · c26b38b2
      moto authored 
      Summary:
1. Fix initialization.
Previously, the SOS token score was initialized to 0 across the time axis.
This was biasing the alignment to delay the start.
The proper way to delay the SOS is via blank token.
The new initilization takes the cumulated sum of blank scores.
2. Fill the end of trellis with Inf
Similar to the start, at the end where there remaining time frame is less
than the number of tokens, it is no longer possible to align the text, thus
we fill with Inf for better visualization.
3. Clean up asset management code.

Pull Request resolved: https://github.com/pytorch/audio/pull/2544

Reviewed By: nateanl

Differential Revision: D38276478

Pulled By: mthrok

fbshipit-source-id: 6d934cc850a0790b8c463a4f69f8f1143633d299
      c26b38b2
    • Zhaoheng Ni's avatar
      Improve speech enhancement tutorial (#2527) · d6267031
      Zhaoheng Ni authored 
      Summary:
- The "speech + noise" mixture still has a high SNR, which can't show the effectiveness of MVDR beamforming. To make the task more challenging, amplify the noise waveform to reduce the SNR of mixture speech.
- Show the Si-SNR score of mixture speech when visualizing the mixture spectrogram.
- FIx the figure in `rtf_power` subsection.
    - The description of enhanced spectrogram by `rtf_power` is wrong. Correct it to `rtf_power`.
- Print PESQ, STOI, and SDR metric scores.

Pull Request resolved: https://github.com/pytorch/audio/pull/2527

Reviewed By: mthrok

Differential Revision: D38190218

Pulled By: nateanl

fbshipit-source-id: 39562850a67f58a16e0a2866ed95f78c3f4dc7de
      d6267031
  9. 28 Jul, 2022 2 commits
  11. 11 Jul, 2022 1 commit
  13. 23 Jun, 2022 1 commit
  15. 17 Jun, 2022 1 commit
  17. 08 Jun, 2022 2 commits
  19. 07 Jun, 2022 3 commits
  21. 04 Jun, 2022 1 commit
  23. 03 Jun, 2022 3 commits
  25. 02 Jun, 2022 1 commit
    • Zhaoheng Ni's avatar
      Update MVDR beamforming tutorial (#2398) · d01f5891
      Zhaoheng Ni authored 
      Summary:
- Use `download_asset` to download audios.
- Replace `MVDR` module with new-added `SoudenMVDR` and `RTFMVDR` modules.
- Benchmark performances of `F.rtf_evd` and `F.rtf_power` for RTF computation.
- Visualize the spectrograms and masks.

Pull Request resolved: https://github.com/pytorch/audio/pull/2398

Reviewed By: carolineechen

Differential Revision: D36549402

Pulled By: nateanl

fbshipit-source-id: dfd6754e6c33246e6991ccc51c4603b12502a1b5
      d01f5891
  27. 01 Jun, 2022 1 commit
    • Caroline Chen's avatar
      Move CTC beam search decoder to beta (#2410) · 93024ace
      Caroline Chen authored 
      Summary:
Move CTC beam search decoder out of prototype to new `torchaudio.models.decoder` module.

hwangjeff mthrok any thoughts on the new module + naming, and if we should move rnnt beam search here as well??

Pull Request resolved: https://github.com/pytorch/audio/pull/2410

Reviewed By: mthrok

Differential Revision: D36784521

Pulled By: carolineechen

fbshipit-source-id: a2ec52f86bba66e03327a9af0c5df8bbefcd67ed
      93024ace
  29. 26 May, 2022 1 commit
  31. 23 May, 2022 1 commit
  33. 21 May, 2022 1 commit
    • moto's avatar
      Add file-like object support to Streaming API (#2400) · a984872d
      moto authored 
      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
      a984872d
  35. 15 May, 2022 1 commit
    • John Reese's avatar
      [codemod][usort] apply import merging for fbcode (8 of 11) · d62875cc
      John Reese authored 
      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
      d62875cc
  37. 13 May, 2022 1 commit
    • moto's avatar
      Move Streamer API out of prototype (#2378) · 72b712a1
      moto authored 
      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
      72b712a1
  39. 12 May, 2022 2 commits
    • Zhaoheng Ni's avatar
      Fix CollateFn in HuBERT pre-training recipe (#2296) · 09639680
      Zhaoheng Ni authored 
      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
      09639680
    • John Reese's avatar
      [black][codemod] formatting changes from black 22.3.0 · 595dc5d3
      John Reese authored 
      Summary:
Applies the black-fbsource codemod with the new build of pyfmt.

paintitblack

Reviewed By: lisroach

Differential Revision: D36324783

fbshipit-source-id: 280c09e88257e5e569ab729691165d8dedd767bc
      595dc5d3
  41. 11 May, 2022 1 commit
    • hwangjeff's avatar
      Refactor LibriSpeech Conformer RNN-T recipe (#2366) · 69467ea5
      hwangjeff authored 
      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
      69467ea5
  43. 28 Apr, 2022 1 commit
  45. 26 Apr, 2022 1 commit
  47. 22 Apr, 2022 1 commit
    • Zhaoheng Ni's avatar
      Introduce DistributedBatchSampler (#2299) · 6411c9ad
      Zhaoheng Ni authored 
      Summary:
When using customized `batch_sampler`, pytorch_lightning can't wrap the distributed sampler onto it. Hence we provide a `DistributedBatchSampler` that supports `BucketizeBatchSampler` in `ddp` mode.

The `DistributedBatchSampler` assumes `BucketizeBatchSampler.iter_list` is a list of lists, where each sub-list contains a batch of indices. Setting `shuffle` to `True` will shuffle the lists based on `seed` and current `epoch`.

The `shuffle` only happens in the initialization, and won't be changed if user don't reset it. The reason is shuffling `BucketizeBatchSampler` may have a different length than before, do shuffling in ``__iter__`` may result in mismatch between ``__len__`` and the real length value.
Hence users need to set `reload_dataloaders_every_n_epochs=1` in pytorch_lightning's Trainer. Then the value of ``__len__``  and the real length is the same.

Pull Request resolved: https://github.com/pytorch/audio/pull/2299

Reviewed By: hwangjeff

Differential Revision: D35781538

Pulled By: nateanl

fbshipit-source-id: 6e8396615497f1aeddab1ee5678830c0445c2b2a
      6411c9ad
  49. 21 Apr, 2022 1 commit
    • hwangjeff's avatar
      Change underlying implementation of RNN-T hypothesis to tuple (#2339) · 6b242c29
      hwangjeff authored 
      Summary:
PyTorch Lite, which is becoming a standard for mobile PyTorch usage, does not support containers containing custom classes. Consequently, because TorchAudio's RNN-T decoder currently returns and accepts lists of `Hypothesis` namedtuples, it is not compatible with PyTorch Lite. This PR resolves said incompatibility by changing the underlying implementation of `Hypothesis` to tuple.

Pull Request resolved: https://github.com/pytorch/audio/pull/2339

Reviewed By: nateanl

Differential Revision: D35806529

Pulled By: hwangjeff

fbshipit-source-id: 9cbae5504722390511d35e7f9966af2519ccede5
      6b242c29
  51. 13 Apr, 2022 2 commits
    • hwangjeff's avatar
      Add Conformer RNN-T LibriSpeech training recipe (#2329) · c262758b
      hwangjeff authored 
      Summary:
Adds Conformer RNN-T LibriSpeech training recipe to examples directory.

Produces 30M-parameter model that achieves the following WER:

|                     |          WER |
|:-------------------:|-------------:|
| test-clean          |       0.0310 |
| test-other          |       0.0805 |
| dev-clean           |       0.0314 |
| dev-other           |       0.0827 |

Pull Request resolved: https://github.com/pytorch/audio/pull/2329

Reviewed By: xiaohui-zhang

Differential Revision: D35578727

Pulled By: hwangjeff

fbshipit-source-id: afa9146c5b647727b8605d104d928110a1d3976d
      c262758b
    • hwangjeff's avatar
      Add nightly build installation code snippet to prototype feature tutorials (#2325) · fb51cecc
      hwangjeff authored 
      Summary:
Tutorial notebooks that leverage TorchAudio prototype features don't run as-is on Google Colab due to its runtime's not having nightly builds pre-installed. To make it easier for users to run said notebooks in Colab, this PR adds a code block that installs nightly Pytorch and TorchAudio builds as a comment that users can copy and run locally.

Pull Request resolved: https://github.com/pytorch/audio/pull/2325

Reviewed By: xiaohui-zhang

Differential Revision: D35597753

Pulled By: hwangjeff

fbshipit-source-id: 59914e492ad72e31c0136a48cd88d697e8ea5f6c
      fb51cecc
  53. 05 Apr, 2022 1 commit
  55. 04 Apr, 2022 2 commits
  57. 01 Apr, 2022 1 commit
  59. 25 Mar, 2022 1 commit