v1.3.1

Removed

• Dependency pins for python version, should work for python 3.9-3.12

v1.3.0

Fixed

• Custom encode functions that return dictionaries of tensors/arrays not working.

Removed

• Dependency pins, all basic functionality tested with latest dependency versions.

v1.2.1

Added

• Profiler script.

Changed

• Performance enhancements by using batch operations in hugging face and torch.
  All interaction functions now need to support accepting an array of texts.
  (The encoding function has changed as a result.)

Fixed

• Corrected bug in how summation is occuring in sort_salience_map() and added unit test.

v1.1.0

Added

• Suggested dev_env.yml environment for contributors.
• API update to devices where "mps" is now checked in addition to "cuda".

Changed

• Stopwords will be from tx2 init rather than nltk download.
• Bumped package versions.
• Specified package versions in requirements and setup.py.
• Updating example jupyter notebooks to use new versions of packages.
• Datasources in jupyter example notebooks.

Fixed

• Updated to patched numpy version 1.22.
• Potential issue in calc.frequent_words_in_cluster() where clusters of empty
  string values would stop computation.

v1.0.2

Fixed

• Wrapper function still expecting pandas series instead of numpy array.
• missing nltk.download("stopwords")

v1.0.1

Added

• Example notebook demonstrating using TX2 with a huggingface model with
  sequence classification head, rather than a custom torch implementation.
• Pre-commit hooks.

Changed

• Add support for huggingface sequence classification head to default
  interaction functions.

Fixed

• Code formatting to fix flake8-indicated issues.

v1.0.0

We're considering this library's API relatively stable at this point, so we're making this our 1.0 release and will be following semver for any future updates!

Notable changes from pre-1.0 version:

• The wrapper class now takes numpy arrays and/or pandas series rather than dataframes and column names for the training and testing data. (See the wrapper docs and the second code block in basic usage#default-approach for an example)
• We have unit tests now! They can be run with pytest from the project root
• The cuda torch device can be overridden for the default encoding handler, this is passed to the wrapper constructor with cuda_device
• Version specifications are added for several of the required dependencies
• Many various bug fixes