umapjax

UMAP, but accelerated. (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 implements the following backends (despite being named umapjax):

1. torch (PyTorch)
2. mx (MLX)
3. jax (JAX)

Getting started

import umapjax

layout_backend: Literal["jax", "mx", "torch"] = "jax"
spectral_backend: Literal["jax", "scipy", "torch"] = "scipy"
batch_size: int | None = None # Defaults to X.shape[0]

model = umapjax.UmapJax(
    n_neighbors=15,
    layout_backend=layout_backend,
    spectral_backend=spectral_backend
)
embedding = model.fit_transform(X)

If the optimization is slow, try increasing the batch size as a multiple of X.shape[0]. All backends will automatically use accelerated hardware if available.

If using "torch", you can set umapjax.layouts.torch.TORCH_DEVICE and umapjax.spectral.torch.TORCH_DEVICE to control the default device used for the layout and spectral embedding, respectively.

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 gradients of the points are weighted by edge weights, which control sampling frequencies in the original algorithm. If results look strange, try changing n_epochs or batch_size. The batch_size argument can also be used to control acceleration on GPUs/TPUs.

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 with a preferred backend:

pip install "umapjax[jax,mlx,torch]"

2. Install the latest development version:

pip install "umapjax[jax,mlx,torch] @ git+https://github.com/adamgayoso/umapjax.git@main"

Contact

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

Citation

t.b.a