Crate umap_rs 

Fast, parallel Rust implementation of the UMAP dimensionality reduction algorithm.

This library provides a streamlined implementation of UMAP (Uniform Manifold Approximation and Projection) focused on performance and correctness. It requires precomputed k-nearest neighbors and initialization, allowing you to use your preferred libraries for these steps.

Example

use umap::{Umap, UmapConfig};
use umap::EuclideanMetric;

// Configure UMAP
let config = UmapConfig::default();
let umap = Umap::new(config);

// Fit to data (requires precomputed KNN and initialization)
let model = umap.fit(
    data.view(),
    knn_indices.view(),
    knn_dists.view(),
    init.view(),
);

// Get the embedding
let embedding = model.embedding();

Features

• Parallel optimization: Rayon’s parallel SGD (Hogwild! algorithm)
• Extensible metrics: Custom distance functions via the Metric trait
• Zero-copy views: Efficient array handling with ndarray

Limitations

• Dense arrays only (no sparse matrix support)
• Fit only (transform for new points not yet implemented)
• Panics on invalid input (no Result-based error handling)
• Requires external KNN computation and initialization

Public API

The library exposes a minimal, well-defined API:

• Umap - Main algorithm struct
• FittedUmap - Fitted model with embeddings
• UmapConfig - Configuration parameters
• Metric - Distance metric trait
• EuclideanMetric - Euclidean distance implementation

Re-exports

pub use config::GraphParams;

pub use config::ManifoldParams;

pub use config::OptimizationParams;

pub use config::UmapConfig;

pub use manifold::LearnedManifold;

pub use metric::Metric;

pub use metric::MetricType;

pub use optimizer::Optimizer;

Modules

config

manifold

metric

optimizer

Structs

EuclideanMetric

Euclidean (L2) distance metric.

FittedUmap

A fitted UMAP model containing the learned manifold and embedding.

Umap

UMAP dimensionality reduction algorithm.

Type Aliases

SparseMat

Sparse matrix with u32 column indices to save memory (4 bytes vs 8 bytes per index). Uses usize for indptr since nnz can exceed u32::MAX for very large datasets. Valid for n_samples < 2^32 (~4 billion).