umapjax 0.0.1

UMAP, but optimized with jax.

umapjax

UMAP, but optimized with jax. (Experimental implementation)

umapjax inherits the API of umap-learn. The UmapJax class is a drop-in replacement for umap.UMAP, with a few key differences:

1. umapjax does not support densmap.
2. umapjax does not support output_metric other than euclidean.

Note: umapjax does not fully replicate umap-learn and care should be used when interpreting results.

This package is intended to be used in combination with accelerated hardware like GPUs and TPUs. There is no benefit to using umapjax on a CPU.

Getting started

import umapjax

model = umapjax.UmapJax(n_neighbors=15, n_epochs=200)
embedding = model.fit_transform(X)

Implementation details

The implementaion used in umapjax is very similar to the one used in umap-learn; however, rather than a single step updating one single point, we update a set of points in parallel using jax. The size of the set of points is controlled by the batch_size parameter.

Installation

You need to have Python 3.11 or newer installed on your system. If you don't have Python installed, we recommend installing uv.

There are several alternative options to install umapjax:

1. Install the latest release of umapjax from PyPI:

pip install umapjax

2. Install the latest development version:

pip install git+https://github.com/adamgayoso/umapjax.git@main

Release notes

See the changelog.

Contact

If you found a bug, please use the issue tracker.

Citation

t.b.a