sklears-neural

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

Overview

sklears-neural delivers multilayer perceptrons and neural utility blocks that align with scikit-learn’s neural-network module while embracing Rust’s performance story.

Key Features

• Models: MLPClassifier, MLPRegressor, RBMs, autoencoders (including self-supervised/contrastive variants).
• Optimizers: SGD, Adam, AdamW, Nadam, RMSprop, L-BFGS, and adaptive learning-rate schedules.
• Hardware Acceleration: SIMD kernels and CUDA execution (oxicuda-backed gpu feature) — real on-device FP16 tensor-core GEMM (tensor_core_gemm_f16/mixed_precision_gemm) and a real oxicuda-dnn conv2d forward pass (tensor_core_conv2d), plus a pooled GPU memory allocator (gpu_pool module) with real hit/miss/allocation telemetry.
• Integration: Works with sklears pipelines, calibration, inspection, and export utilities.

Quick Start

use sklears_neural::{Activation, MLPClassifier, Solver};
use scirs2_core::ndarray::array;

let x = array![
    [0.0, 0.0],
    [0.0, 1.0],
    [1.0, 0.0],
    [1.0, 1.0],
];
let y: Vec<usize> = vec![0, 1, 1, 0];

let mlp = MLPClassifier::new()
    .hidden_layer_sizes(&[16, 16])
    .activation(Activation::Relu)
    .solver(Solver::Adam)
    .max_iter(1000)
    .random_state(42);

let fitted = mlp.fit(&x, &y)?;
let probs = fitted.predict_proba(&x)?;

Status

• Exercised via 449 passing crate tests in 0.2.0 (86 skipped).
• Verified against scikit-learn parity tests for convergence and scoring APIs.
• Knowledge distillation ships today (knowledge_distillation module); ONNX export is not yet implemented.