#machine-learning #tensor #matrix #complex #iron #learn #gradient

bin+lib iron_learn

A pure Rust Machine Learning Library with Generic Tensor and a Gradient Descent Optimization Function

7 unstable releases (3 breaking)

new 0.4.0 Jun 6, 2024
0.3.2 May 19, 2024
0.2.0 May 13, 2024
0.1.1 May 13, 2024
0.0.1 Apr 13, 2024

#90 in Machine learning

Download history 183/week @ 2024-04-13 2/week @ 2024-04-20 616/week @ 2024-05-11 187/week @ 2024-05-18 7/week @ 2024-05-25

810 downloads per month

MIT license

67KB
764 lines

iron_learn

A pure Rust Machine Learning Library

Status

Version 0.4.0 released with Logistic Regression Algorithm. Now you can directly use linear_regression or logistic_regression functions instead of directly using gradient_descent function, which has been marked as deprecated and is going to introduce breaking changes and in future may be removed. Under active development for further implementation support.

N.B: This library was built as a product of Rust Learning Efforts. So, "School Grade" algorithm is used for Matrix Multiplication. I have plans for improving on the performance part though in future.

Overview

This library is designed to facilitate machine learning tasks with a focus on linear algebra operations. Currently, the library supports matrix addition, subtraction, multiplication, transpose and scaling by a scalar providing a robust foundation for building more complex machine learning algorithms.

Modules

tensor

At the core of the library, the tensor module supports multi-dimensional Tensor data structure. It includes methods for tensor instantiation and defines the + operator for tensor addition, the - operator for subtraction and the * operator for tensor multiplication. These operators however, take ownership of the variables(both rhs and lhs). So, the variables cannot be used later. To facilitate performing these operations without the ownership issue, it also provides add, sub and mul methods to perform addition, subtraction and division by borrowing the variables for operation.

Additionally, it features a method t for transpose and a multiply method for the Hadamard product, an element-wise multiplication operation. It also pose scale method for scaling by a scalar value.

These operations on Tensor can fail due to multiple reasons and hence, it returns a result object. The library assumes a better error handling by the user rather than causing panic.

N.B: This Module is limited to only two dimensional Matrix Support. This is a temporary restriction and I plan to remove this restriction in future.

complex

The complex module is experimental and provides a representation of complex numbers, which are fundamental in various machine learning computations, especially in handling operations involving complex-valued data. The Complex type supports all arithmatic operations.

numeric

This module defines all supported numeric types necessary for machine learning operations, including integer, unsigned, and floating-point variants, as well as the custom type Complex.

gradient_descent

The gradient_descent function performs a single step of the gradient descent optimization algorithm. It can work for both Linear Regression and Logistic Regression, configured by a boolean parameter.

matrix (Use 2-Dimensional Tensor instead)

The matrix module provides the Matrix structure which serves as a wrapper for the Tensor object of two dimensions, enabling matrix operations. It defines the + and * operators for matrix addition and multiplication, respectively, mirroring the behavior of tensors. It also enables multiply method to support hadamard product of two matrices.

vector (Use 1-Dimensional Tensor instead)

The vector module provides Vector type which is a specialized wrapper for the Tensor object, tailored for vector operations. It defines the + operator for vector addition and the * operator for computing the dot product between two vectors.

Usage

To use the library, include the following in your project:

use iron_learn::Complex;
use iron_learn::Matrix;
use iron_learn::Tensor;
use iron_learn::Vector;

Examples

Here are some examples of how to use the library for basic operations:

Tensor Addition

let a = Tensor::new(vec![2, 2], vec![1, 2, 3, 4]).unwrap(); // Define matrix `a`
let b = Tensor::new(vec![2, 2], vec![5, 6, 7, 8]).unwrap(); // Define matrix `b`
let c = (a + b).unwrap(); // Perform matrix addition. Causes a move. Cannot use a or b later.

// No move syntax
let a = Tensor::new(vec![2, 2], vec![1, 2, 3, 4]).unwrap(); // Define matrix `a`
let b = Tensor::new(vec![2, 2], vec![5, 6, 7, 8]).unwrap(); // Define matrix `b`
let c = a.add(&b).unwrap(); // Perform matrix addition without taking ownership..

Tensor Subtraction

let a = Tensor::new(vec![2, 2], vec![1, 2, 3, 4]).unwrap(); // Define matrix `a`
let b = Tensor::new(vec![2, 2], vec![5, 6, 7, 8]).unwrap(); // Define matrix `b`
let c = (a - b).unwrap(); // Perform matrix addition. Causes a move. Cannot use a or b later.

// No move syntax
let a = Tensor::new(vec![2, 2], vec![1, 2, 3, 4]).unwrap(); // Define matrix `a`
let b = Tensor::new(vec![2, 2], vec![5, 6, 7, 8]).unwrap(); // Define matrix `b`
let c = a.sub(&b).unwrap(); // Perform matrix addition without taking ownership..

Tensor Multiplication

let a = Tensor::new(vec![2, 2], vec![1, 2, 3, 4]).unwrap(); // Define matrix `a`
let b = Tensor::new(vec![2, 2], vec![5, 6, 7, 8]).unwrap(); // Define matrix `b`
let c = (a * b).unwrap(); // Perform matrix multiplication. Causes a move. Cannot use a or b later.

// No move syntax
let a = Tensor::new(vec![2, 2], vec![1, 2, 3, 4]).unwrap(); // Define matrix `a`
let b = Tensor::new(vec![2, 2], vec![5, 6, 7, 8]).unwrap(); // Define matrix `b`
let c = a.add(&b).unwrap(); // Perform matrix multiplication without taking ownership.

Tensor Hadamard Product

let a = Tensor::new(vec![2, 2], vec![1, 2, 3, 4]).unwrap(); // Define matrix `a`
let b = Tensor::new(vec![2, 2], vec![5, 6, 7, 8]).unwrap(); // Define matrix `b`
let c = a.multiply(b).unwrap(); // Perform matrix hadamard product

Tensor Transpose

let m = Tensor::new(vec![6], vec![1, 2, 3, 4, 5, 6]).unwrap();
 m.t().unwrap();

Tensor Scaling

let m = Tensor::new(vec![6], vec![1, 2, 3, 4, 5, 6]).unwrap();
let r = Tensor::new(vec![6], vec![5, 10, 15, 20, 25, 30]).unwrap();
 assert-eq!(m.scale(5).unwrap(), r);

Gradient Descent Optimization

use iron_learn::Tensor;
use iron_learn::gradient_descent::gradient_descent;

let learning_rate: f64 = 0.01;
let w = Tensor::new(vec![2, 1], vec![3.0, 4.0]).unwrap();
let x = Tensor::new(vec![1, 2], vec![3.0, 4.0]).unwrap();
let y = Tensor::new(vec![1, 1], vec![5.0]).unwrap();
let w = gradient_descent(&x, &y, &w, learning_rate, true);

Complex Number Arithmatic

let a = Complex::new(1.0, 2.0);
let b = Complex::new(3.0, 4.0);

let c = a + b;
let c = a - b;
let c = a * b;
let c = a / b;

Complex Number Tensor Addition & Multiplication

let a = Complex::new(1.0, 2.0);
let b = Complex::new(3.0, 4.0);
let c = Complex::new(5.0, 6.0);
let d = Complex::new(7.0, 8.0);

let m1 = Tensor::new(vec![2, 2], vec![a, b, c, d]).unwrap();

let m2 = Tensor::new(vec![2, 2], vec![a, c, b, d]).unwrap();

let result = m1.add(&m2).unwrap();
let result = m1.mul(&m2).unwrap();

Matrix Addition (Deprecated, use 2-Dimensional Tensor instead)

let a = Matrix::new(vec![2, 2], vec![1, 2, 3, 4]).unwrap(); // Define matrix `a`
let b = Matrix::new(vec![2, 2], vec![5, 6, 7, 8]).unwrap(); // Define matrix `b`
let c = (a + b).unwrap(); // Perform matrix addition

Matrix Multiplication (Deprecated, use 2-Dimensional Tensor instead)

let a = Matrix::new(vec![2, 2], vec![1, 2, 3, 4]).unwrap(); // Define matrix `a`
let b = Matrix::new(vec![2, 2], vec![5, 6, 7, 8]).unwrap(); // Define matrix `b`
let c = (a * b).unwrap(); // Perform matrix multiplication

Matrix Hadamard Product (Deprecated, use 2-Dimensional Tensor instead)

let a = Matrix::new(vec![2, 2], vec![1, 2, 3, 4]).unwrap(); // Define matrix `a`
let b = Matrix::new(vec![2, 2], vec![5, 6, 7, 8]).unwrap(); // Define matrix `b`
let c = a.multiply(b).unwrap(); // Perform matrix hadamard product

Vector Addition (Deprecated, use 1-Dimensional Tensor instead)

let a = Vector::new(vec![2], vec![1, 2]).unwrap(); // Define matrix `a`
let b = Vector::new(vec![2], vec![3, 4]).unwrap(); // Define matrix `b`
let c = (a + b).unwrap(); // Perform matrix addition

Vector Dot Product (Deprecated, use 1-Dimensional Tensor instead)

let a = Vector::new(vec![2], vec![1, 2]).unwrap(); // Define matrix `a`
let b = Vector::new(vec![2], vec![3, 4]).unwrap(); // Define matrix `b`
let c = (a * b).unwrap(); // Perform matrix addition

Dependencies

~0.7–1.5MB
~33K SLoC