Unda

General purpose neural network crate

Unda aims to bring the future of deep learning to the world of rust. With dynamic input traits, concurrent minibatch processing, and full Dense network support, Unda is quickly emerging and making neural network development easy and blazingly fast.

Installation

1. Identify the latest compatible versions of CUDA and cuDNN. Adapt these instructions to install the two version of CUDA and cuDNN together.

2. Install clang and libclang1.

3. Download and extract xla_extension to the root directory of your target project.

4. Make sure LD_LIBRARY_PATH includes /path/to/xla_extension/lib, and make sure the relevant CUDA paths are also visible to the system.

Use the package manager cargo to add unda to your rust project.

cargo add unda

or add the dependency directly in your cargo.toml file

[dependencies]
unda = "{version}"

Usage

use unda::core::network::Network;
use unda::core::layer::{methods::activations::Activations, layers::{LayerTypes, InputTypes}};
use unda::core::data::input::Input;
use unda::core::layer::{methods::errors::ErrorTypes};

fn main() {
    let inputs = vec![vec![0.0,0.0],vec![1.0,0.0],vec![0.0,1.0], vec![1.0,1.0]];
    let outputs = vec![vec![0.0],vec![1.0],vec![1.0], vec![0.0]];

    let mut new_net = Network::new(4);

    new_net.set_input(InputTypes::DENSE(2));
    new_net.add_layer(LayerTypes::DENSE(3, Activations::RELU, 0.001));
    new_net.add_layer(LayerTypes::DENSE(1, Activations::SIGMOID, 0.001));

    new_net.compile();

    new_net.fit(&inputs, &outputs, 2, ErrorTypes::MeanAbsolute);

    println!("1 and 0: {:?}", new_net.predict(&vec![1.0,0.0])[0]);
    println!("0 and 1: {:?}", new_net.predict(&vec![0.0,1.0])[0]);
    println!("1 and 1: {:?}", new_net.predict(&vec![1.0,1.0])[0]);
    println!("0 and 0: {:?}", new_net.predict(&vec![0.0,0.0])[0]);

    new_net.save("best_network.json");
}

Examples

The unda repository hosts a plethora of example ML models to compute a series of common problems. These examples can be found in the /examples folder and can be run by entering:

cargo run --release --example {example_name}

where example_name is the name of the file/folder you wish to run, omitting the .rs

Currently, Unda has example implementations for XoR, MNIST and a breast cancer model from Kaggle

Important! When using running the MNIST example, please make sure to put the appropriate ubyte files into the /src/util/mnist directory of this repository. We are currently working on using reqwest to automatically build the dataset, but for now it must be done manually

Implications for the future of ML

Using the built in Input trait, practically any data type can be mapped to an input for a neural network without the need for cutting corners, and the inner trait for layers allows for a plug and play style to neural network development. Currently, Unda has full support for Dense layers, Adam Optimization for Backprop, Activation functions (Sigmoid, TanH, ReLU and LeakyReLU), and even loss analysis per model and per layer.

Gradient descent currently can happen both syncronously as stochastic gradient descent or asynchronously through minibatch gradient descent.

If open source development is your thing, we at Unda would love additional work on anything that can be implemented, please contact [email protected] if you'd like to help out!

License

Licensed under the Apache License, Version 2.0 http://www.apache.org/licenses/LICENSE-2.0 or the MIT license http://opensource.org/licenses/MIT, at your option. This file may not be copied, modified, or distributed except according to those terms.