pyfit 0.1.2

A minimalist neural networks library built on a tiny autograd engine

pyfit

pyfit is a minimalist neural networks library written from scratch in Python for educational purposes.

Overview

This project aims to:

• demonstrate automatic differentiation, a core concept of modern Deep Learning frameworks like PyTorch and TensorFlow;
• define a simple API for training neural nets, somehow mimicking PyTorch Ignite and Keras;
• follow good coding practices, including type annotations and unit tests.

This material is used in the Machine Learning course taught at ENSC. ENSEIRB-MATMECA and IOGS. See also Acknowledgments.

Demonstration

The demo notebook showcases what pyfit is all about.

Features

• Autograd engine [ source | tests ]
• Neural networks API [ source | tests ]
• Metrics [ source | tests ]
• Optimizers [ source | tests ]
• Data utilities [ source | tests ]
• Training API [ source | tests ]

Development Notes

Checking the code

pyfit uses the following tools:

• black for code formatting.
• pylint and mypy for linting.
• pytest for testing.

Run the following commands in project root folder to check the codebase.

> python -m pylint ./pyfit # linting (including type checks)
> python -m mypy .         # type checks only
> python -m pytest         # test suite

Uploading the package to PyPI

> python setup.py sdist bdist_wheel
> python -m twine upload dist/* --skip-existing