sklears-feature-extraction

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

Overview

sklears-feature-extraction contains text, signal, and image feature transformers designed to mirror scikit-learn’s feature extraction API with Rust-first performance.

Key Features

• Text Processing: CountVectorizer, TfidfVectorizer, HashingVectorizer, N-gram analyzers, character models.
• Image Features: Patch extraction, HOG descriptors, a full SIFT keypoint pipeline, and SIMD-accelerated pixel/statistical operations; image::simd_accelerated genuinely delegates to scirs2_core::simd_ops::SimdUnifiedOps for vectorized execution (previously several of its functions were plain scalar loops mislabeled as SIMD). There is no GPU backend in this crate.
• Signal Features: Windowed statistics, spectrograms, wavelet transforms, and FFT-based descriptors.
• Pipeline Support: Integrates with sklears preprocessing, selection, and model selection crates.

Quick Start

use sklears_feature_extraction::text::TfidfVectorizer;

let docs = vec![
    "Rust brings fearless concurrency".to_string(),
    "Machine learning in Rust is fast".to_string(),
];

let mut vectorizer = TfidfVectorizer::new()
    .ngram_range((1, 2))
    .min_df(1)
    .max_features(Some(4096));

let tfidf = vectorizer.fit_transform(&docs)?;

Status

• Extensively tested; 416 passing crate tests in 0.2.0.
• Offers >99% parity with scikit-learn’s feature extraction module, with SIMD-accelerated hot paths (no GPU backend).
• This session fixed image::simd_accelerated: 7 of its 9 public functions were labeled "SIMD-accelerated" (with fabricated speedup figures up to "8.7x" in their own doc comments) despite being plain scalar loops — e.g. simd_array_subtraction was literally a - b; all 7 now genuinely delegate to scirs2_core::simd_ops::SimdUnifiedOps, verified numerically equivalent to the prior behavior to 1e-9 tolerance. This audit was scoped to that one module — other image:: files (patch_extraction, wavelet_features, surf_features) still carry legacy "SIMD-accelerated"/speedup doc comments that were not verified against their implementations in this pass.
• Additional work (streaming text ingestion, audio-specific transforms) documented in TODO.md.