GardenPy is a package that integrates an automatic differentiation package with machine learning algorithms to build machine learning from scratch.
It's built on NumPy, utilizing NumPy's computation to calculate linear algebra results. It utilizes similar syntax, meaning most forward-pass math in NumPy applies to GardenPy's Tensors, with the added benefit of automatically differentiating Tensors.
There are many machine learning packages, many of which are faster and easier to use than this one. This package's purpose isn't to compete with the speed or features of other packages; instead, the package's goal is to allow the creation of machine learning models using syntax as close to mathematical syntax as possible, to allow the model's creator to truly understand the mathematics behind machine learning.
GardenPy's purpose isn't to compete with other machine learning libraries regarding speed or features. Rather, GardenPy's mission is to allow the creation of machine learning models using syntax that is as similar to mathematical syntax as possible.
GardenPy automatically creates a computational graph to represent all calculations. When a line calls for gradient calculation, GardenPy uses tape-based automatic differentiation and a search tree to relate two Tensors.
GardenPy's automatic differentiation is designed to be easily expanded.
This library tries to do as many things as possible without relying on pre-defined methods. Below are the required prerequisites and what they're used for.
- Python 3.9 +
- Codebase.
- NumPy 1.26.1 +
- Linear algebra computations.
To install GardenPy, ...
GardenPy has pre-built model setups to quickly create models to train. These models utilize GardenPy's automatic differentiation and can be used as examples of how to best use GardenPy's automatic differentiation. Additionally, GardenPy has example applications on the MNIST dataset and a simple checkered or non-checkered dataset.
The Machine Learning from Scratch team also offers a website with mathematics for machine learning, detailed documentation, diagrams, videos, and an example of a dense neural network trained on the MNIST dataset. The website's repository can be found here.
GardenPy is licensed under the AGPL-3.0 License. See License for more.
For questions, recommendations, or feedback on GardenPy, contact:
- Christian SW Host-Madsen ([email protected])
- Doyoung Kim ([email protected])
- Mason YY Morales ([email protected])
- Isaac P Verbrugge ([email protected])
- Derek Yee ([email protected])
Report any issues to the issues page.
This project is a work in progress.
Tensor autograd – Done Tensor base operations – In progress ML algorithms – Done Tensor algorithm integration – Done Computational graph building – Done Base models – In progress Data visualization tools – Not started Example implementation – Not started Library setup, README, and website – In progress
