sklears-calibration

Latest release: 0.1.0-beta.1 (January 1, 2026). See the workspace release notes for highlights and upgrade guidance.

Overview

sklears-calibration provides probability calibration tools, matching scikit-learn’s calibration module with additional Rust-centric performance improvements.

Key Features

• CalibratedClassifierCV: Platt scaling, isotonic regression, and temperature scaling strategies.
• Probability Tools: Reliability diagrams, Brier score decomposition, and calibration curve generation.
• Integration: Works with sklears pipelines, model selection, and inspection modules.
• GPU Support: Optional CUDA/WebGPU acceleration for large-scale calibration workloads.

Quick Start

use sklears_calibration::CalibratedClassifierCV;
use sklears_ensemble::RandomForestClassifier;

let base = RandomForestClassifier::builder()
    .n_estimators(200)
    .n_jobs(-1)
    .build();

let calibrated = CalibratedClassifierCV::builder()
    .base_estimator(base)
    .method("sigmoid")
    .cv(5)
    .build();

let fitted = calibrated.fit(&x_train, &y_train)?;
let probas = fitted.predict_proba(&x_test)?;

Status

• Covered by the 11,292 passing workspace tests in 0.1.0-beta.1.
• API parity with scikit-learn 1.5, including multi-class calibration.
• Future work (Bayesian calibration, streaming reliability) tracked in this crate’s TODO.md.