Release V1.2.0

[AnFreTh]

Release v1.2.0

This update enhances the preprocessing and embedding layers in the mambular package, introducing several key improvements:

• Feature-Specific Preprocessing: The Preprocessor class now includes a feature preprocessing dictionary, enabling different preprocessing strategies for each feature.
• Support for Unstructured Data: The model can now handle a combination of tabular features and unstructured data, such as images and text.
• Latent Representation Generation: It is now possible to generate latent representations of the input data, improving downstream modeling and interpretability.

These changes enhance flexibility and extend mambular's capabilities to more diverse data modalities.

Preprocessing improvements:

• mambular/preprocessing/preprocessor.py: Added feature_preprocessing parameter to allow custom preprocessing techniques for individual columns. Updated the fit method to use this parameter for both numerical and categorical features. [1] [2] [3] [4] [5]

Embedding layer updates:

• mambular/arch_utils/layer_utils/embedding_layer.py: Modified the forward method to handle different dimensions of categorical embeddings and ensure they are properly processed. [1] [2]

Allow unstructured data as inputs:

• mambular/arch_utils/layer_utils/embedding_layer.py: Modified the forward method to handle num_features, cat_features and pre-embedded unstructured data. [1] [2]

Get latent representation of tables

• mambular/base_models/basemodel.py: Updated the encode method to accept a single data parameter instead of separate num_features and cat_features parameters. [1] [2]

Documentation Updates:

• Added a new "What's New" section to highlight recent features and improvements, such as individual preprocessing for each feature and the use of embeddings as inputs (README.md).
• Enhanced the preprocessing section to include details about specifying individual preprocessing methods for each feature and using pre-trained encoding for categorical features (README.md) [1] [2].
• Updated the usage examples to demonstrate how to get latent representations for each feature and use unstructured data (README.md).
• Removed the "Custom Training" section and integrated custom model implementation details into other sections (README.md) [1] [2].

Workflow Automation:

• Introduced a new GitHub Actions workflow (.github/workflows/pr-tests.yml) to run unit tests on pull requests targeting the develop and master branches. This workflow sets up the environment, installs dependencies, runs the tests, and ensures that pull requests cannot be merged if tests fail.

Codebase Improvements:

• Updated the version number from 1.1.0 to 1.2.0 in mambular/__version__.py to reflect the new release.
• Removed the rotary_embedding_torch dependency and associated code from the attention mechanism (mambular/arch_utils/layer_utils/attention_utils.py) to simplify the implementation [1] [2] [3] [4] [5].
• Enhanced the EmbeddingLayer class to support embedding projections and handle additional embedding types (mambular/arch_utils/layer_utils/embedding_layer.py) [1] [2] [3] [4] [5] [6].