sklearn-morpho

Scikit-learn classifiers toolbox for morphological perceptrons.

Current features

• Scikit-learn classifiers:
  • Linear Dilation-Erosion Perceptron (l-DEP)
  • Dilation-Erosion Perceptron (DEP)
  • Simple Dilation and Erosion Morphological Perceptrons
• Modular wrapper for DCCP optimization tasks with cvxpy

File tree:

• src/sklearn_morpho: contains the source code and a testsuite in its tests subdirectory.
• tests: pytest testsuite.
• testing: standalone files that use this library. They are used internally for testing, but they are not tests nor part of the library.
• MREs: standalone jupyter notebooks to showcase some of this library's features.

Getting started

Take a look at the Jupyter code examples in the MREs directory.

Running the project

Install Python 3 and hatch. Then run one of these commands:

• hatch run jupyter lab to run the Jupyter notebooks
• hatch run pytest for tests
• hatch shell to run testing files like testing/display_boundary.py in the right environment.

Special note for the estimators comparison testing files: they are split in two files to avoid training the estimators every time one wants to view the results.

Contributing

As the project is open source, any help is greatly appreciated!

To keep code clean, this repository uses a CI/CD pipeline with tests and an enforced coding style, which can all be found in the project's configuration files.

Please adhere to these rules when contributing.

Regarding AI pull requests, they are generally discouraged as you should be able to understand and help maintaining any features you add to this repo.

For Arch users

The Python package is available in the Arch User Repository as python-sklearn-morpho, as are all its dependencies not already in the official packages, except for dccp.

Since the latter is an optional dependency, you can either:

• Download dccp through pip or use pip install sklearn-morpho[dccp] in a Python virtual environment.
• Download python-sklearn-morpho from the AUR and use pip install dccp --user.

What are morphological perceptrons?

Perceptrons are the fundamental building blocks of many machine learning constructs, namely the well-known neural networks.

They fulfill this role because their simple yet effective activation function ($y = w \cdot x + b$) allows for the emergence of arbitrarily complex behavior, for example when such perceptrons are wired together in networks.

In the equation above, $w$ and $x$ are row and column matrices respectively, meaning their multiplication equates to a set of scalar multiplications, followed by a summation of all these multiplication results.

"Morphological" perceptrons are similar to the aforedescribed "classical" perceptrons, but differ in the fact that they generally do not have a bias $b$, and most importantly that the operations performed are different:

While, as established above, a classical perceptron computes several multiplications followed by a summation, a morphological perceptron performs multiple additions, followed by a $\min$ or $\max$.

This simple shift into a different semiring creates a more complex decision boundary, making these constructs a natural choice for designing binary classifiers.

This repository explores the behavior of such perceptrons, namely standalone ones, but also more complex structures, such as DEP and l-DEP.