v0.1.1

0.1.1 - 2024-08-19

Performance Improvements

• Made the ternary array saving more memory efficient. (#3)

Miscellaneous Changes

• Bumped dependency versions in poetry.lock.
• Removed pytest-xdist dependency for the tests Poetry group.
• Bumped sigstore/gh-action-sigstore-python from 2.1.1 to 3.0.0 in the github-actions group. (#4)

v0.1.0

0.1.0 - 2024-07-22

Initial stable release of Keras-MatMulLess.

New Features

• Added a matmul-less version of the Dense layer in Keras.
• Added a Root Mean Square Normalization (RMSNorm) layer.
• Added a matmul-less version of a Gated Linear Unit (GLU), along with its variants by using different activation functions.
• Added a matmul-less Gated Recurrent Unit (GRU) layer.
• Added a matmul-less Linear Recurrent Unit (LRU) layer.
• Added a matmul-less transformer block layer, along with its associated attention and embedding layers.

---

Changelog From Last Prerelease: https://github.com/PhotonicGluon/Keras-MatMulLess/compare/v0.1.0rc2..v0.1.0

---

New Contributors

• @dependabot made their first contribution in #1

v0.1.0rc2

0.1.0rc2 - 2024-07-19

Documentation Changes

• Added missing call conventions for GRU and LRU.

Miscellaneous Changes

• Updated the versions of the dependencies listed in poetry.lock.

v0.1.0rc1

0.1.0rc1 - 2024-07-18

Miscellaneous Changes

• Cleaned up the install.py script.
• Fixed Input 'repository_url' has been deprecated with message: The inputs have been normalized to use kebab-case. warning in GitHub actions.
• Removed useless setup code for Ruff linting action.

v0.1.0b1

0.1.0b1 - 2024-07-11

Changes

• Changed linter from PyLint to Ruff.
• Loosened dependency requirements to allow for more compatibility with other packages.
• Reorganized poetry dependency groups.

Fixes

• Added missing scikit-learn dependency for the notebook Poetry group.
• Fixed missing type annotation for _quantize_kernel() in BaseDenseMML.

Miscellaneous Changes

• Bump GitHub actions' versions. (#1)
• Added a security policy, code of conduct, and contribution guidelines.
• Added new training tests to GitHub actions to ensure that the changes does not affect performance and results of the models.
• Configured dependabot to look out for dependency updates.
• Made devcontainer settings become VSCode workspace settings.
• Renamed and redid some GitHub actions.

---

New Contributors

• @dependabot made their first contribution in #1

v0.1.0a3

0.1.0a3 - 2024-07-10

Removals

• Removed Python 3.12 support (to be in line with Keras' Python supported versions)

New Features

• Added stricter call annotation using the jaxtyping package.

Changes

• Downgraded minimum required version of NumPy from 1.26.4 to 1.23.5.

Documentation Changes

• Added new code example on vision transformers.
• Added new style for call convention.

Miscellaneous Changes

• Fixed stable-build.yml GitHub action.
• Added more package information (e.g., classifiers, keywords) into pyproject.toml.

v0.1.0a2

0.1.0a2 - 2024-07-05

New Features

• Added more parameters for GRUMML.
• Added more parameters for LRUMML.
• Created Patches and PatchEmbedding layers for vision transformers (ViT).

Changes

• Improved coding standards to better match PyLint standards.
• Moved TokenEmbedding to the core package, to align with the Keras organization of the layers.
• Split tests to better run in parallel.

Performance Improvements

• Combined several internal layers in GRUMML into one kernel layer, reducing computational load.

Documentation Changes

• Added explanation for how the matmul-less recurrent units work.
• Added hover tips to documentation.
• Fixed scrollbar ugliness with the math block rendering on the documentation site.
• Made a new look for the homepage.
• Removed spurious "defaults to" in some layers' documentation.

v0.1.0a1

0.1.0a1 - 2024-07-03

Initial alpha release of Keras-MML.

New Features

• Added a matmul-less version of the Dense layer in Keras.
• Added a Root Mean Square Normalization (RMSNorm) layer.
• Added a matmul-less version of a Gated Linear Unit (GLU), along with its variants by using different activation functions.
• Added a matmul-less Gated Recurrent Unit (GRU) layer.
• Added a matmul-less Linear Recurrent Unit (LRU) layer.
• Added a matmul-less transformer block layer, along with its associated attention and embedding layers.