Skip to content

GitLab

Switch branch/tag
  1. 21 Oct, 2022 1 commit
  3. 20 Oct, 2022 1 commit
  5. 19 Oct, 2022 7 commits
  7. 18 Oct, 2022 1 commit
  9. 17 Oct, 2022 1 commit
  11. 14 Oct, 2022 2 commits
  13. 13 Oct, 2022 4 commits
  15. 12 Oct, 2022 4 commits
    • Nikita Shulga's avatar
      Fix typos in tacotron2 tutorial (#2761) · 7aabcbd4
      Nikita Shulga authored 
      Summary:
`publishe`->`published`

Also, not sure if it should be `pre-trained weight is published` or `pre-trained weights are published`

Pull Request resolved: https://github.com/pytorch/audio/pull/2761

Reviewed By: carolineechen

Differential Revision: D40313042

Pulled By: malfet

fbshipit-source-id: c22085ca0b1125a06aa04bf38231d0a9fbfed00b
      7aabcbd4
    • Zhaoheng Ni's avatar
      Improve hubert recipe for pre-training and fine-tuning (#2744) · 27433050
      Zhaoheng Ni authored 
      Summary:
following pr https://github.com/pytorch/audio/issues/2716
- For preprocessing
  - The HuBERT feature takes lots of memory which may not fit some machines. Enable to use a subset of feature for training a k-means model.

- For pre-training
  - Normalize the loss based on the total number of masked frames across all GPUs.
  - Use mixed precision training. fp16 is not well supported in pytorch_lightning.
  - Log accuracies of masked/unmasked frames during training.
  - Clip the gradients with norm `10.0`.

- For ASR fine-tuning
  - Normalize the loss based on the total number of batches across all GPUs, same as in the conformer recipe of TorchAudio.
  - Use mixed precision training.
  - Add "|" after the end of transcription to capture the silence/word termination, same as in fairseq recipe.

- Update the WER results on LibriSpeech dev and test sets.

|                   | WER% (Viterbi)|  WER% (KenLM) |
|:-----------------:|--------------:|--------------:|
| dev-clean         |       10.9    |       4.2     |
| dev-other         |       17.5    |       9.4     |
| test-clean        |       10.9    |       4.4     |
| test-other        |       17.8    |       9.5     |

Pull Request resolved: https://github.com/pytorch/audio/pull/2744

Reviewed By: carolineechen

Differential Revision: D40282322

Pulled By: nateanl

fbshipit-source-id: 4723584c912e70e8970149fe09de005385eaab90
      27433050
    • Zhaoheng Ni's avatar
      Improve wav2vec2/hubert model for pre-training (#2716) · c5bd93b6
      Zhaoheng Ni authored 
      Summary:
This PR improves the Wav2Vec2/HuBERT model regarding model pre-training.

- The model initialization of positional embedding and transformer module is essential to model pre-training. The accuracy of unmasked frames should be higher than masked frames, as it is an easier task. but without the initialization, the accuracy of masked frames is higher than unmasked frames.
  Compared the performance after two epochs with 16 GPUs.
  - With model initialization, the accuracies of masked/unmasked frames are 0.08/0.11.
  - Without model initialization, the accuracies of masked/unmasked frames are 0.06/0.04.
- After adding the model initialization, the gradient is easy to overflow (aka `nan` gradient). In paper [Self-Supervised Learning for speech recognition with Intermediate layer supervision](https://arxiv.org/abs/2112.08778) the authors propose a simple but effective method to mitigate the overflow issue, by scaling down the multiplication of query and key and subtracting the maximum value from it (subtracting a constant value won't change the output of softmax). Then it guarantees the value won't be overflowed.
- In the original fairseq, the mask indices are generated by `numpy.random.choice`. Here replace `torch.multinomial` with `torch.randperm`. (cc carolineechen).

Other improvements within training scripts will be included in a separate PR.

Pull Request resolved: https://github.com/pytorch/audio/pull/2716

Reviewed By: xiaohui-zhang

Differential Revision: D39832189

Pulled By: nateanl

fbshipit-source-id: f4d2a473a79ad63add2dd16624bd155d5ce4de27
      c5bd93b6
    • Caroline Chen's avatar
      Skip hubert xlarge torchscript test (#2758) · c2ea6898
      Caroline Chen authored 
      Summary:
a couple of circleci unittests are failing during hubert xlarge torchscript test, which has been known to fail on Windows in the past (#65776). this PR disables this test on circleci

cc atalman

Pull Request resolved: https://github.com/pytorch/audio/pull/2758

Reviewed By: mthrok

Differential Revision: D40290535

Pulled By: carolineechen

fbshipit-source-id: 5c5fb43434a517b6c439a8cb8e853015d1550a57
      c2ea6898
  17. 11 Oct, 2022 4 commits
  19. 10 Oct, 2022 2 commits
    • Zhaoheng Ni's avatar
      Add unit test for LibriMix dataset (#2659) · c5b8e585
      Zhaoheng Ni authored 
      Summary:
Besides the unit test, the PR also addresses these issues:
- The original `LibriMix` dataset only supports "min" mode, which means the audio length is the minimum of all clean sources. It is default for source separation task. Users may also want to use "max" mode which allows for end-to-end separation and recognition. The PR adds ``mode`` argument to let users decide which dataset they want to use.
- If the task is ``"enh_both"``, the target is the audios in ``mix_clean`` instead of separate clean sources. The PR fixes it to use ``mix_clean`` as target.

Pull Request resolved: https://github.com/pytorch/audio/pull/2659

Reviewed By: carolineechen

Differential Revision: D40229227

Pulled By: nateanl

fbshipit-source-id: fc07e0d88a245e1367656d3767cf98168a799235
      c5b8e585
    • Zhaoheng Ni's avatar
      Fix HuBERT docstring (#2746) · be938e7e
      Zhaoheng Ni authored 
      Summary:
The docstring of `wav2vec2` argument is wrong. Fix it in this PR.

Pull Request resolved: https://github.com/pytorch/audio/pull/2746

Reviewed By: carolineechen

Differential Revision: D40225995

Pulled By: nateanl

fbshipit-source-id: 770e9c928ebebd7b6307e181601eb64625d668da
      be938e7e
  21. 09 Oct, 2022 1 commit
  23. 08 Oct, 2022 1 commit
  25. 07 Oct, 2022 3 commits
  27. 06 Oct, 2022 3 commits
  29. 05 Oct, 2022 1 commit
  31. 03 Oct, 2022 3 commits
  33. 01 Oct, 2022 1 commit