flowsom-rs 0.1.0

Rust-accelerated batch-learning SOM for FlowSOM

flowsom-rs

Rust SOM training for FlowSOM, via PyO3.

Implements the batch-learning SOM from Otsuka et al. 2025 with rayon parallelism. Unlike FlowSOM's online SOM, the batch algorithm processes all events simultaneously each epoch, so results are deterministic regardless of input event order.

Install

pip install flowsom-rs

Requires Python ≥ 3.10 and numpy ≥ 1.24.

Quick start

import numpy as np
import flowsom_rs
from scipy.spatial.distance import pdist, squareform

data = np.random.randn(500_000, 7)

# SOM grid setup (same format FlowSOM_Python uses)
grid = [(x, y) for x in range(10) for y in range(10)]
nhbrdist = squareform(pdist(grid, metric="chebyshev"))
radii = (np.quantile(nhbrdist, 0.67), 0.0)
codes = data[np.random.choice(len(data), 100, replace=False)]

# Train
codes, bmu_idx, bmu_dist = flowsom_rs.train_batch_som(
    data, codes, nhbrdist, radii, rlen=10
)

Functions

train_batch_som(data, codes, nhbrdist, radii, rlen, n_threads=None) Batch-learning SOM. Parallel BMU search + accumulation. Deterministic. Returns (codes, bmu_indices, bmu_distances).

train_online_som(data, codes, nhbrdist, alphas, radii, rlen, seed) Sequential Kohonen SOM, same algorithm as FlowSOM_Python's SOMEstimator.

train_replicas_som(data, codes, nhbrdist, alphas, radii, rlen, num_replicas=10, seed=42, n_threads=None) Parallel replicas merged via median, same approach as FlowSOM_Python's BatchSOMEstimator.

map_data_to_codes(data, codes) Parallel nearest-code assignment. Returns (indices, distances).

Benchmarks

10×10 grid, 10 epochs, Apple Silicon:

Events	Numba Online	Rust Batch	Speedup
50K	364 ms	44 ms	5.9×
100K	491 ms	85 ms	6.4×
500K	2,709 ms	391 ms	6.9×

Thread scaling at 500K events: 1→2→4→8 threads gives 1.0→2.0→3.3→4.6× speedup.

Build from source

pip install maturin
maturin develop --release

License

MIT