README.md

Neurotic is a work in progress machine-learning library for Rust.

crates.io docs.rs

Roadmap

List features to be implemented:

• GPU Acceleration support
• Autotuning
• Convolutional Layers
• Recurrent Layers
• Regularisation
• Performance metric output
• ONNX Support
• Reduced Precision Training

Installation

Add neurotic to your Cargo.toml dependencies:

cargo add neurotic

Examples

Defining the network architecture

Start by defining the amount of neurons in each layer, and the layers' activation functions.

use neurotic::{
    activation::ActivationFunction,
    core::NeuralNetwork,
    training::{CostFunction, NetworkTrainer},
};

let layer_sizes = &[2, 32, 16, 1]; // 2 neurons for the input layer, 32 and 16 for the hidden
layers, and 1 output neuron.
let activation_functions = &[
    ActivationFunction::Identity,
    ActivationFunction::ReLU,
    ActivationFunction::ReLU,
    ActivationFunction::Identity,
];
let cost_function = CostFunction::MeanSquaredError;

// Create a new instance of NeuralNetwork with the defined structure
let network = NeuralNetwork::new(layer_sizes, activation_functions, cost_function);

Preparing the training data

Load in or generate your training data. Here is a simple example that generates training data for a sum function.

use rand::Rng;

// This returns a vector of tuples. Each tuple is made up of inputs, and target outputs.
fn generate_sum_data(size: usize, range: f32) -> Vec<(Vec<f32>, Vec<f32>)> {
    let mut data = Vec::with_capacity(usize);
    for _ in 0..size {
        let a = rand::thread_rng.gen_range(0.0..range);
        let b = rand::thread_rng.gen_range(0.0..range);
        let output = a + b;
        data.push((vec![a, b], vec![output]));
    }
    data
}

// Store the generated training data in a variable
let training_data = generate_sum_data(1000, 10.0);

Training the network

Set up the training parameters, and train the network using a NetworkTrainer.

let learning_rate = 0.001; // Network's learning rate
let batch_size = 50; // Divide the training data into batches of this size
let epochs = 500; // Number of training iterations

let mut trainer: NetworkTrainer<NoOptimiser> = NetworkTrainer::new(network, learning_rate, batch_size, None);
trainer.train(training_data, epochs);

Saving or loading a network

Saving the trained network to a file.

trainer.get_network().save("path/to/file").expect("Failed to save network");

Loading a trained network from a file.

let network = NeuralNetwork::load("path/to/file").expect("Failed to load network");

Using Optimisers

Using Optimisers. Adam optimiser for this example. Refer to the docs.rs page for other optimisers' parameters.

// Defining optimiser parameters and creating new optimiser.
let layer_dimensions = layer_sizes
    .windows(2)
    .map(|w| (w[1], w[0]))
    .collect::<Vec<_>>();
let beta1 = 0.9;
let beta2 = 0.999;
let epsilon = 1e-8;
let adam = Adam::new(&layer_dims, beta1, beta2, epsilon);

// Creating a new trainer with the optimiser
let mut trainer: NetworkTrainer<Adam> =
    NetworkTrainer::new(network, learning_rate, batch_size, Some(adam));

Contributing

Pull requests are the best way to propose changes to the program.

1. Fork the repo and create your branch from main.
2. Make your changes.
3. If your change directly affects the program's functionality, update the documentation.
4. Issue a pull request

Any contributions you make will be under the MIT Software License

In short, when you submit code changes, your submissions are understood to be under the same MIT License that covers the project.

Report issues using Github's Issues tab.

I use GitHub issues to track public bugs. Report a bug by opening a new issue.

Issue Reports tend to have:

• A quick summary and/or background
• Steps to reproduce

License

This project is licensed under the MIT License. See the LICENSE file for more details.