#clustering #cluster #means #simd-vector

nightly kmeans

Small and fast library for k-means clustering calculations

5 releases (1 stable)

1.1.0 Nov 17, 2024
1.0.0 Nov 16, 2024
0.11.0 Oct 8, 2024
0.10.0 Jun 21, 2024
0.1.0 Jul 27, 2019

#222 in Algorithms

Download history 3/week @ 2024-09-11 5/week @ 2024-09-18 11/week @ 2024-09-25 117/week @ 2024-10-02 95/week @ 2024-10-09 18/week @ 2024-10-16 34/week @ 2024-10-23 35/week @ 2024-10-30 83/week @ 2024-11-06 251/week @ 2024-11-13 78/week @ 2024-11-20

459 downloads per month
Used in 2 crates

Apache-2.0

110KB
1.5K SLoC

kmeans

Current Crates.io Version docs

kmeans is a small and fast library for k-means clustering calculations. It requires a nightly compiler with the portable_simd feature to work.

Here is a small example, using kmean++ as initialization method and lloyd as k-means variant:

use kmeans::*;

fn main() {
    let (sample_cnt, sample_dims, k, max_iter) = (20000, 200, 4, 100);

    // Generate some random data
    let mut samples = vec![0.0f64;sample_cnt * sample_dims];
    samples.iter_mut().for_each(|v| *v = rand::random());

    // Calculate kmeans, using kmean++ as initialization-method
    // KMeans<_, 8> specifies to use f64 SIMD vectors with 8 lanes (e.g. AVX512)
    let kmean: KMeans<f64, 8, _> = KMeans::new(samples, sample_cnt, sample_dims, EuclideanDistance);
    let result = kmean.kmeans_lloyd(k, max_iter, KMeans::init_kmeanplusplus, &KMeansConfig::default());

    println!("Centroids: {:?}", result.centroids);
    println!("Cluster-Assignments: {:?}", result.assignments);
    println!("Error: {}", result.distsum);
}

Datastructures

For performance-reasons, all calculations are done on bare vectors, using hand-written SIMD intrinsics from the packed_simd crate. All vectors are stored row-major, so each sample is stored in a consecutive block of memory.

Supported variants / algorithms

  • lloyd (standard kmeans)
  • minibatch

Supported centroid initialization methods

  • KMean++
  • random partition
  • random sample

Supported distance functions

  • Euclidean distance
  • Histogram distance

Dependencies

~2.2–3MB
~61K SLoC