[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

Releases

1.1.1

Released on 12 August 2024 • View on Gitlab

Breaking changes

  • The nnutils library is no longer maintained and is only compatible with Python 3.6, 3.7, 3.8. As such its dependency has been removed. The crnn.use_masks parameter has been removed. It is still supported to keep the compatibility with older training configuration but will be ignored.

Feature

  • The number of worker processes created in dataloaders is now exposed through the data.num_workers parameter.
  • There is a new command to run basic checks and compute statistics on your training dataset. Learn more about it in the documentation.
  • Pretraining is now available. Load the weights of a previous checkpoint using the train.pretrain parameter when fine-tuning a model on a new dataset. Learn more about it in the documentation.

  • When training on a small dataset, freezing some of the layers can help with model convergence. The train.freeze_layers parameter supports freezing:

    • convolutional layers,
    • recurrent layers,
    • linear layers.

  • Proper support for right-to-left (RTL) languages is now available. Enable it using the data.reading_order argument both during training and decoding.

Dependencies

Misc

  • A deprecation warning from jsonargparse was fixed.
  • The package's metadata are now stored in pyproject.toml as per PEP-0621.
  • PyLaia now uses ruff for linting and formatting.

1.1.0

Released on 22 December 2023 • View on Gitlab

Breaking changes

  • Official support for Python3.8 has been dropped. This doesn't mean that the current code doesn't run on python3.8, we simply do not test that compatibility anymore. This decision was made since active support of python 3.8 has stopped for a while now and many libraries in the ML world have stopped supporting it as well.

Feature

  • A Docker image with the needed code to use this library is now built on every tag.
  • The coverage of our tests suite is displayed again as a GitLab badge on the repository as well as in the README.md file.

Documentation

Dependencies

  • Bumped pytorch-lightning to version 1.3.0

  • Some dependencies were pinned to a version to avoid breakage:

    • natsort was pinned to version 8.4.0,
    • textdistance was pinned to version 4.6.0,
    • scipy was pinned to version 1.11.3,
    • matplotlib was pinned to version 3.8.2,
    • numpy direct dependency was removed since it's installed through scipy and matplotlib.

  • PyLaia dropped support for python 3.8 so the dataclasses dependency was dropped.

Misc

  • The torch.testing.assert_allclose has been replaced by torch.testing.assert_close since it became deprecated in PyTorch 1.12.0.

1.0.7

Released on 18 October 2023 • View on Gitlab

Feature

  • When using a language model, a confidence score is now returned based on the log-likelyhood of the hypothesis.

Documentation

A public documentation is now available on https://atr.pages.teklia.com/pylaia/. It's still under construction but next releases will add more and more content.

Dependencies

Development

  • Releases are now built more easily through a Makefile.
  • The documentation is also redeployed after each push on master branch.
  • Fixed a test that behaved differently locally and during CI.

1.0.6

Released on 12 September 2023 • View on Github

Feature

  • During training, too small images are now padded to be able to pass the multiple convolution layers.

Documentation

  • Fixed typos.

Dependencies

Development

  • Pre-commit hooks were updated.

1.0.5

Released on 29 March 2023 • View on Github

Dependencies

  • Requires torch version 1.13.0 or 1.13.1.
  • Requires torchvision version 0.14.0 or 0.14.1 (depending on torch version).

1.0.4

Released on 4 January 2023 • View on Github

Dependencies

  • Requires torch version 1.13.0.

1.0.3

Released on 12 December 2022 • View on Github

Feature

  • Now able to decode using a trained Language model through beam search decoding.
  • Exposes torch Dataloaders's num_workers parameter on the Python training function to limit resource usage when needed.

Dependencies

  • Added dependency to torchaudio version 0.13.0.

Development

  • Package version is now tracked through the VERSION file.

1.0.2

Released on 7 December 2022 • View on Github

Dependencies

  • Pinned dependency to pytorch-lightning to version 1.1.0.

1.0.1

Released on 7 December 2022 • View on Github

1.0.0

Released on 2 December 2020 • View on Github

Added

  • Support distributed training
  • Scripts can now be configured using yaml configuration files
  • Add support for the SGD and Adam optimizers
  • Support color images
  • Log the installed version of each module when scripts are called from shell
  • Add char/word segmentation to the decode script
  • Add several badges to the README
  • Support using a ReduceLROnPlateau scheduler during training
  • A CSV file (metrics.csv) is now created with the results obtained during training
  • Add CONTRIBUTING file
  • Training now can include GPU stats in the progress bar
  • Add isort to pre-commit to keep consistent imports throughout the codebase
  • Users can run the PyLaia scripts using Python now
  • Support half-precision training for fixed height models.
  • Add script to visualize the segmentation output
  • Use Codecov to produce test coverage reports
  • Code is now analyzed using CodeFactor

Changed

  • Make Python 3.6 the minimum supported version
  • Make PyTorch 1.4.0 the minimum supported version
  • Remove ImageToTensor in favor of vision transform ToImageTensor
  • Remove all of the internal logic (engine, actions, hooks, etc) in favor of pytorch-lightning's constructs
  • Change Travis CI for GitHub actions
  • Greatly improve the progress bar. It is used now in all scripts
  • The entire shell API has changed for the better (thanks to jsonargparse). Arguments are now separated into groups and help messages are clearer.
  • Drastically improve our test suite, we now have a 91% coverage

Removed

  • Remove egs directory. These live now at https://github.com/carmocca/PyLaia-examples
  • Remove Baidu's CTC loss in favor of PyTorch's
  • Remove PHOC code. Please open an issue if you were using it
  • Remove Dortmund code. Please open an issue if you were using it
  • Remove CTCLatticeGenerator. Please open an issue if you were using it
  • We no longer support saving checkpoints for more than one metric. Will be added back in a future version

Fixed

  • Fix WER calculation when long delimiters are used
  • Exit training if a delimiter is not present in the vocabulary
  • Hundreds of other minor fixes and refactors to improve the code quality!