Squeeze

Explore structure. Measure the tradeoffs.

Squeeze is a Python library for CPU dimensionality reduction, combining a Python/Numba UMAP implementation with ten Rust algorithms. Fit a linear baseline, compare neighborhood layouts, and inspect the quality you keep when reducing data to two dimensions.

Install Squeeze Compare the algorithms

One project, several ways to reduce dimensions

Start here	What you get
Quick start	A runnable Digits embedding with a quality measurement
Algorithm guide	Eleven methods, their actual APIs, and their constraints
Benchmark results	Digits and Fashion-MNIST heatmaps, timing and raw data
New graph methods	NeighborMap and SpectralMap, experimental Rust alternatives
API reference	Constructors, supported methods, and composition utilities

What is implemented

UMAP offers the richest estimator interface, including out-of-sample transforms. Rust PCA provides fit and transform; the other Rust methods primarily expose fit_transform. They do not all implement the full scikit-learn estimator contract. Check the capability table before building a pipeline.

This is an alpha research library. The benchmarks document particular datasets, parameters, and machines; they are not a blanket speed or quality guarantee. Squeeze targets CPU execution. No CUDA installation is needed for the core methods.

Built on previous work

Squeeze derives its UMAP implementation from umap-learn. The project keeps that lineage explicit while developing its own Rust methods and comparison tooling. See credits and licensing for attribution.