sklears (crate)

Latest release: 0.2.0 (July 14, 2026). See the workspace release notes for highlights and upgrade guidance.

Overview

This crate exposes the top-level sklears API that bundles all subcrates into a cohesive, scikit-learn compatible experience. It re-exports sklears-core/sklears-utils at the crate root and exposes each optional algorithm subcrate as a feature-gated module.

Key Features

• Per-Module Re-exports: Each enabled feature exposes its subcrate as a named module (sklears::linear, sklears::tree, sklears::neighbors, ...) — there is no single prelude glob import.
• Feature Flags: Enable only the modules you need (linear, ensemble, gpu, etc.) to keep builds lightweight.
• Rust + Python: Designed to work with both native Rust projects and the sklears-python bindings.
• Documentation Hub: Acts as the canonical entry point for examples, tutorials, and integration guides.

Quick Start

[dependencies]
sklears = { version = "0.2.0", features = ["linear", "ensemble", "gpu"] }

use sklears::tree::RandomForestClassifier;
use sklears::traits::{Fit, Predict};
use scirs2_core::ndarray::{array, Array1};

let x = array![
    [0.0, 1.0],
    [1.0, 0.0],
    [1.0, 1.0],
];
let y = Array1::from(vec![0, 1, 1]);

let model = RandomForestClassifier::new()
    .n_estimators(200)
    .fit(&x, &y)?;

let predictions = model.predict(&x)?;

Note: this crate is a facade — it has no prelude module. Each enabled feature re-exports its subcrate under its own name (sklears::linear, sklears::tree, sklears::ensemble, ...); import types from the specific module for the algorithms you enabled.

Status

• Serves as the umbrella facade crate, re-exporting sklears-core/sklears-utils plus one module per optional algorithm-category feature (linear, clustering, ensemble, tree, neighbors, naive-bayes, multiclass, semi-supervised, and more — see Cargo.toml's [features] for the full list).
• GPU acceleration is available behind the gpu feature (OxiCUDA/CUDA-backed via sklears-core::gpu), forwarding to each enabled subcrate's own gpu feature with an honest CPU fallback when no device is detected.
• This crate's own integration/property tests: 32 passing (13 skipped) with --all-features for 0.2.0; each subcrate carries its own, much larger, unit-test suite documented in its own README.
• The preprocessing feature and its sklears-preprocessing optional dependency (previously commented out workspace-wide) are restored in 0.2.0, along with the 8 algorithm-showcase examples and the tree_ensemble_benchmarks bench target that require it (linear_models_showcase, lasso_regression, kmeans_clustering, dbscan_clustering, hierarchical_clustering, mean_shift_clustering, spectral_clustering, gmm_clustering, plus performance_comparison_comprehensive).
• Re-export map kept in sync with individual module READMEs and documentation.
• Further enhancements (module-level doc consolidation, feature flag audits) tracked in TODO.md.