2 releases

0.3.1 Jun 5, 2019
0.3.0 Jun 5, 2019

#521 in Machine learning

Custom license

34KB
807 lines

R.A.I.L: A Rust Artificial Intelligence Library

RAIL is designed to be a library for easily creating and training Neural Networks, akin to the Keras API. It aims to be fast and easy to use.

Dependencies

RAIL depends on arrayfire-rust, so before using RAIL make sure you have arrayfire installed.

A Simple XOR Problem

Solving the XOR Problem with Mold is super easy! Simply add the crate to your Cargo.toml:

rail = { git = "https://github.com/nlsnightmare/rail" }

Then add this to your code

use rail::model::Model;
use rail::layers::dense::Dense;
use rail::layers::activations::Activation;

pub fn main() {
    let mut model = Model::new()
        .learning_rate(0.01)
        .input_size(2)
        .layer(Dense::new(2).activation(Activation::Tanh))
        .layer(Dense::new(1).activation(Activation::Tanh))
        .build(true)
        .unwrap();

    let tranining_data = vec![
        (vec![0., 0.], vec![0.]),
        (vec![0., 1.], vec![1.]),
        (vec![1., 0.], vec![1.]),
        (vec![1., 1.], vec![0.]),
    ];

    // Train with a batch of 2 for 4000 epochs
    model.train(&tranining_data, 2, 4000);

    println!("[0, 0] -> {}", model.predict(vec![0., 0.])[0]); // should be close to 0
    println!("[0, 1] -> {}", model.predict(vec![0., 1.])[0]); // should be close to 1
    println!("[1, 0] -> {}", model.predict(vec![1., 0.])[0]); // should be close to 1
    println!("[1, 1] -> {}", model.predict(vec![1., 1.])[0]); // should be close to 0
}

Plans

As of now, RAIL is in a very early state, and under heavy development. The API will change a lot.
So far, only Dense (aka fully connected) layers are supported, and batched SGD is the only way of training the network. However, there are plans to support:

  • Convolutional Layers
  • RNN Cells
  • LSTM Cells
  • Genetic Crossover
  • ADAM optimizer
  • More Activation functions
  • More Error functions
  • Documentation

Dependencies

~2.5MB
~40K SLoC