rustlearn

Machine learning crate for Rust

Programming language: Rust

License: Apache License 2.0

Tags: Machine-learning    

Latest version: v0.5.0

README

rustlearn

A machine learning package for Rust.

For full usage details, see the API documentation.

Introduction

This crate contains reasonably effective implementations of a number of common machine learning algorithms.

At the moment, rustlearn uses its own basic dense and sparse array types, but I will be happy to use something more robust once a clear winner in that space emerges.

Features

Matrix primitives

• dense matrices
• sparse matrices

Models

• logistic regression using stochastic gradient descent,
• support vector machines using the libsvm library,
• decision trees using the CART algorithm,
• random forests using CART decision trees, and
• factorization machines.

All the models support fitting and prediction on both dense and sparse data, and the implementations should be roughly competitive with Python sklearn implementations, both in accuracy and performance.

Cross-validation

• k-fold cross-validation
• shuffle split

Metrics

• accuracy
• ROC AUC score
• dcg_score
• ndcg_score
• mean absolute error
• mean squared error

Parallelization

A number of models support both parallel model fitting and prediction.

Model serialization

Model serialization is supported via serde.

Using rustlearn

Usage should be straightforward.

• import the prelude for all the linear algebra primitives and common traits:

use rustlearn::prelude::*;

• import individual models and utilities from submodules:

use rustlearn::prelude::*;

use rustlearn::linear_models::sgdclassifier::Hyperparameters;
// more imports

Examples

Logistic regression

use rustlearn::prelude::*;
use rustlearn::datasets::iris;
use rustlearn::cross_validation::CrossValidation;
use rustlearn::linear_models::sgdclassifier::Hyperparameters;
use rustlearn::metrics::accuracy_score;


let (X, y) = iris::load_data();

let num_splits = 10;
let num_epochs = 5;

let mut accuracy = 0.0;

for (train_idx, test_idx) in CrossValidation::new(X.rows(), num_splits) {

    let X_train = X.get_rows(&train_idx);
    let y_train = y.get_rows(&train_idx);
    let X_test = X.get_rows(&test_idx);
    let y_test = y.get_rows(&test_idx);

    let mut model = Hyperparameters::new(X.cols())
                                    .learning_rate(0.5)
                                    .l2_penalty(0.0)
                                    .l1_penalty(0.0)
                                    .one_vs_rest();

    for _ in 0..num_epochs {
        model.fit(&X_train, &y_train).unwrap();
    }

    let prediction = model.predict(&X_test).unwrap();
    accuracy += accuracy_score(&y_test, &prediction);
}

accuracy /= num_splits as f32;

Random forest

use rustlearn::prelude::*;

use rustlearn::ensemble::random_forest::Hyperparameters;
use rustlearn::datasets::iris;
use rustlearn::trees::decision_tree;

let (data, target) = iris::load_data();

let mut tree_params = decision_tree::Hyperparameters::new(data.cols());
tree_params.min_samples_split(10)
    .max_features(4);

let mut model = Hyperparameters::new(tree_params, 10)
    .one_vs_rest();

model.fit(&data, &target).unwrap();

// Optionally serialize and deserialize the model

// let encoded = bincode::serialize(&model).unwrap();
// let decoded: OneVsRestWrapper<RandomForest> = bincode::deserialize(&encoded).unwrap();

let prediction = model.predict(&data).unwrap();

Contributing

Pull requests are welcome.

To run basic tests, run cargo test.

Running cargo test --features "all_tests" --release runs all tests, including generated and slow tests. Running cargo bench --features bench (only on the nightly branch) runs benchmarks.