Getting started#

This package serves as a skeleton package aiding at developing compatible scikit-learn contribution.

Creating your own scikit-learn contribution package#

Download and setup your repository#

To create your package, you need to clone the project-template repository:

git clone https://github.com/scikit-learn-contrib/project-template.git

Before to reinitialize your git repository, you need to make the following changes. Replace all occurrences of skltemplate, sklearn-template, or project-template with the name of you own project. You can find all the occurrences using the following command:

git grep skltemplate
git grep sklearn-template
git grep project-template

To remove the history of the template package, you need to remove the .git directory:

rm -rf .git

Then, you need to initialize your new git repository:

git init
git add .
git commit -m 'Initial commit'

Finally, you create an online repository on GitHub and push your code online:

git remote add origin https://github.com/your_remote/your_contribution.git
git push origin main

Develop your own scikit-learn estimators#

You can modify the source files as you want. However, your custom estimators need to pass the check_estimator test to be scikit-learn compatible. We provide a file called test_common.py where we run the checks on our custom estimators.

You can refer to the User Guide to help you create a compatible scikit-learn estimator.

In any case, developers should endeavor to adhere to scikit-learn’s Contributor’s Guide which promotes the use of:

algorithm-specific unit tests, in addition to check_estimator’s common tests;
PEP8-compliant code;
a clearly documented API using NumpyDoc and PEP257-compliant docstrings;
references to relevant scientific literature in standard citation formats;
doctests to provide succinct usage examples;
standalone examples to illustrate the usage, model visualisation, and benefits/benchmarks of particular algorithms;
efficient code when the need for optimization is supported by benchmarks.

Managing your local and continuous integration environment#

Here, we set up for you an repository that uses pixi. The pixi.toml file defines the packages and tasks to be run that we will present below. You can refer to the following documentation link to install pixi: https://pixi.sh/latest/#installation

Once done, you can refer to the documentation to get started but we provide the command below to interact with the main task requested to develop your package.

Edit the documentation#

The documentation is created using Sphinx. In addition, the examples are created using sphinx-gallery. Therefore, to generate locally the documentation, you can leverage the following pixi task:

pixi run build-doc

The documentation is made of:

a home page, doc/index.rst;
an API documentation, doc/api.rst in which you should add all public objects for which the docstring should be exposed publicly.
a User Guide documentation, doc/user_guide.rst, containing the narrative documentation of your package, to give as much intuition as possible to your users.
examples which are created in the examples/ folder. Each example illustrates some usage of the package. the example file name should start by plot_*.py.

Local testing#

To run the tests locally, you can use the following command:

pixi run test

It will use pytest under the hood to run the package tests.

In addition, you have a linter task to check the code consistency in terms of style:

pixi run lint

Setup the continuous integration#

The project template already contains configuration files of the continuous integration system. It leverage the above pixi commands and run on GitHub Actions. In short, it will:

run the tests on the different platforms (Linux, MacOS, Windows) and upload the coverage report to codecov.io;
check the code style (linter);
build the documentation and deploy it automatically on GitHub Pages.

Publish your package#

You can make your package available through PyPi and conda-forge. Refer to the associated documentation to be able to upload your packages such that it will be installable with pip and conda.