Features include:
- Ergonomic APIs for building and training your networks.
- Predefined optimizer methods and activation functions, as well as the ability to add your own via implementing a trait.
- Fully customizable GUI via egui for seeing your network during training, including some prebuilt GUIs.
For more detailed examples, look at the examples directory.
- Examples directory. Featured examples:
- adder_gui -- a simple example of creating a custom training GUI to train a neural network to add numbers
- learn_image -- a more involved example for training a neural network to upscale and interpolate images
Note: Running in release makes training drastically faster, i.e. do cargo run --release
.
Here is how you'd create a model that learns to behave like an XOR gate.
let mut net = NeuralNet::new(2, 2, Activations::Arctan);
net.add_layer(1, Activations::Sigmoid);
let inputs = Matrix2::from_array([[0, 0], [0, 1], [1, 0], [1, 1]]).into();
let targets = Matrix2::from_array([[0], [1], [1], [0]]).into();
let learning_rate = 10e-0;
let epochs = 10_000;
// Initialize the optimizer to log the error at each epoch
let optim = Optimizer::new(OptimizerMethod::Backprop, learning_rate, epochs).with_log(Some(1));
optim.train(&mut net, &inputs, &targets);
let fin = net.mean_squared_error(&inputs, &targets).unwrap();
println!("------------------");
let res = net.run_batch(&inputs).unwrap();
for (inp, out) in inputs.to_vec().into_iter().zip(res.to_vec()) {
println!("{:?} -> {}", inp.to_vec(), out[0])
}
To make the optimizer display graph the error at each epoch, you can do this:
optim.train_gui<NNGui>(&mut net, &inputs, &targets);
Here, NNGui
is a predefined struct implementing the Visualizer
trait that allows the optimizer to spawn and asynchronously update a GUI (meaning performance shouldn't be affected).