1 unstable release

0.1.0 Dec 31, 2022

#439 in Machine learning

21 downloads per month

GPL-3.0 license

150KB
3K SLoC

Machine Learning in Rust

Learn the Rust programming language through implementing classic machine learning algorithms. This project is self-completed without relying on any third-party libraries, serving as a bootstrap machine learning library.

❗❗❗:Actively seeking code reviews and welcome suggestions on fixing bugs or code refactoring. Please feel free to share your ideas. Happy to accept advice!

Basics

  1. NdArray Module, just as the name. It has implemented broadcast, matrix operations, permute and etc. in arbitrary dimension. SIMD is used in matrix multiplication thanks to auto vectorizing by Rust.
  2. Dataset Module, supporting customized loading data, re-format, normalize, shuffle and Dataloader. Several popular dataset pre-processing recipes are available.

Algorithms

  1. Decision Tree, supporting both classification and regression tasks. Info gains like gini or entropy are provided.
  2. Logistic Regression, supporting regularization (Lasso, Ridge and L-inf)
  3. Linear Regression, same as logistic regression, but for regression tasks.
  4. Naive Bayes, free to handle discrete or continuous feature values.
  5. SVM, with linear kernel using SGD and Hinge Loss to optimize.
  6. nn Module, containing linear(MLP) and some activation functions which could be freely stacked and optimized by gradient back propagations.
  7. KNN, supporting both KdTree and vanilla BruteForceSearch.
  8. K-Means, clustering data with an unsupervised learning approach

Start

Let's use KNN algorithm to solve a classification task. More examples can be found in examples directory.

  1. create some synthetic data for tests

    use std::collections::HashMap;
    
    let features = vec![
        vec![0.6, 0.7, 0.8],
        vec![0.7, 0.8, 0.9],
        vec![0.1, 0.2, 0.3],
    ];
    let labels = vec![0, 0, 1];
    // so it is a binary classifiction task, 0 is for the large label, 1 is for the small label
    let mut label_map = HashMap::new();
    label_map.insert(0, "large".to_string());
    label_map.insert(1, "small".to_string());
    
  2. convert the data to the dataset

    use mlinrust::dataset::Dataset;
    
    let dataset = Dataset::new(features, labels, Some(label_map));
    
  3. split the dataset into train and valid sets and normalize them by Standard normalization

    let mut temp =  dataset.split_dataset(vec![2.0, 1.0], 0); // [2.0, 1.0] is the split fraction, 0 is the seed
    let (mut train_dataset, mut valid_dataset) = (temp.remove(0), temp.remove(0));
    
    use mlinrust::dataset::utils::{normalize_dataset, ScalerType};
    
    normalize_dataset(&mut train_dataset, ScalerType::Standard);
    normalize_dataset(&mut valid_dataset, ScalerType::Standard);
    
  4. build and train our KNN model using KdTree

    use mlinrust::model::knn::{KNNAlg, KNNModel, KNNWeighting};
    
    // KdTree is one implementation of KNN; 1 defines the k of neighbours; Weighting decides the way of ensemble prediction; train_dataset is for training KNN; Some(2) is the param of minkowski distance
    let model = KNNModel::new(KNNAlg::KdTree, 1, Some(KNNWeighting::Distance), train_dataset, Some(2));
    
  5. evaluate the model

    use mlinrust::utils::evaluate;
    
    let (correct, acc) = evaluate(&valid_dataset, &model);
    println!("evaluate results\ncorrect {correct} / total {}, acc = {acc:.5}", test_dataset.len());
    

Todo

  1. model weights serialization for saving and loading
  2. Boosting/bagging
  3. matrix multiplication with multi threads
  4. refactor codes, sincerely request for comments from senior developers

Reference

  1. scikit-learn
  2. The book, 机器学习西瓜书 by Prof. Zhihua Zhou

Thanks

The rust community. I received many help from rust-lang Discord.

License

Under GPL-v3 license. And commercial use is strictly prohibited.

No runtime deps