umap-rstr

UMAP (Uniform Manifold Approximation and Projection) for Clojure, built on raster — typed multiple dispatch with devirtualizing bytecode compilation. The -rstr suffix in the repo/artifact name (umap-rstr) marks the raster substrate, à la Julia's .jl repo suffix; the namespace you require is just umap.

A faithful port of umap-learn (Python+numba is the gold standard): cosine/ euclidean kNN → fuzzy simplicial set → spectral/random init → negative-sampled SGD layout. The numeric kernels are deftm functions that JIT-compile to primitive-speed JVM bytecode; the orchestration is plain Clojure.

Status

Validated against umap-learn on MNIST-70k and Fashion-MNIST-70k (cosine):

dataset	trust (rstr)	trust (umap-learn)	fit (rstr)	fit (umap-learn)
MNIST-70k	0.951	0.950	83 s	109 s
Fashion-70k	0.901	0.906	67 s	181 s

Trustworthiness matches; wall-clock is faster (incl. cold JVM/JIT). See dev/ for the reproducible comparison harness.

Embeddings side-by-side (left: umap-rstr, right: umap-learn), coloured by the true class label:

Installation

raster depends on a typedclojure fork via git, and git deps resolve transitively only through deps.edn (not a Maven POM), so pin this lib as a git dependency:

io.github.replikativ/umap-rstr
{:git/url "https://github.com/replikativ/umap-rstr"
 :git/sha "<sha>"}

(Once typedclojure ships its fixes in a Maven release upstream, this moves to a org.replikativ/umap-rstr {:mvn/version "..."} Clojars coordinate.)

Usage

(require '[umap :as umap])

;; X: flat row-major double[] (or float[]) of n*dim
(def result (umap/fit X n dim :k 15 :metric :cosine :init :auto :seed 42))
(:emb result)   ;; => double[n*2] embedding

Options: :k (neighbors, 15), :out-dim (2), :n-epochs (auto 500/200), :neg-rate (5.0), :gamma (1.0), :init (:auto/:spectral/:random), :metric (:cosine/:euclidean), :seed (42).

Namespaces

• umap — public fit orchestrator
• umap.layout — negative-sampled SGD (the hot kernel)
• umap.graph — fuzzy simplicial set (smooth-knn-dist, membership, symmetrize)
• umap.spectral — spectral init (matrix-free Lanczos, disconnected-graph handling)

kNN / RP-trees / Tausworthe RNG live in raster (raster.knn, raster.spatial.*, raster.tausworthe) since they're shared with clustering.

Requirements

Runs on the Valhalla JDK (raster's deftm kernels use preview features). See raster's README for JDK setup and run with the :valhalla alias.

clojure -M:valhalla:test          # run tests

License

BSD 3-Clause (see LICENSE and NOTICE).

umap-rstr is a derivative work — a Clojure / raster port of umap-learn (© 2017 Leland McInnes, BSD 3-Clause). It follows umap-learn's algorithm and numerical behaviour but is an independent reimplementation, not endorsed by or affiliated with the original authors. The port is © 2026 Christian Weilbach, released under the same license.

McInnes, L., Healy, J., & Melville, J. (2018). UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. arXiv:1802.03426.