Skigen

High-performance machine learning for modern C++ and Eigen.

Skigen is a header-only C++ template library for machine learning. It brings the scikit-learn API — fit(), transform(), predict() — to the Eigen ecosystem.

Overview

• Skigen is versatile.
  • Preprocessing, linear models, decomposition, clustering, trees, neighbors, pipelines, and metrics — covering the core scikit-learn surface with a consistent API.
  • All estimators are templatized on Scalar (default: double). Switch to float for 2× SIMD throughput on the same hardware.
• Skigen is fast.
  • Eigen's expression templates eliminate temporaries and enable lazy evaluation.
  • Explicit SIMD vectorization via Eigen for SSE, AVX, AVX-512, and ARM NEON.
  • Compile-time polymorphism via CRTP — zero vtable overhead, zero runtime dispatch.
  • No interpreter, no garbage collector, no GIL.
• Skigen is reliable.
  • Bit-level parity with scikit-learn for identical inputs and default parameters, verified via cross-language parity tests.
  • CI-tested on GCC, Clang, and MSVC across Linux, macOS, and Windows.
• Skigen is elegant.
  • Header-only — drop Skigen/ next to Eigen/ and #include. No compiled libraries, no linker flags.
  • Include patterns mirror Eigen: #include <Skigen/Dense> feels like #include <Eigen/Dense>.
  • The same API you know from scikit-learn, native to C++.

Design

Principle	Implementation
Eigen-native	Headers, namespaces, and include patterns mirror Eigen conventions.
Header-only	No compiled libraries, no link-time dependencies.
Templatized	Scalar template parameter — double default, float for dense SIMD.
Zero-copy	Inputs use Eigen::Ref<const MatrixType> — sub-blocks and memory-mapped data without copying.
CRTP	Static polymorphism via the Curiously Recurring Template Pattern.
Bit-level parity	Results match scikit-learn for identical inputs and default parameters.

Quick Example

#include <Eigen/Dense>
#include <Skigen/Preprocessing>
#include <iostream>

int main() {
    Eigen::MatrixXd X(4, 2);
    X << 1, 10,
         2, 20,
         3, 30,
         4, 40;

    Skigen::StandardScaler scaler;
    Eigen::MatrixXd Z = scaler.fit_transform(X);

    std::cout << "Standardized:\n" << Z << "\n";
    return 0;
}