Crate fann

A Rust wrapper for the Fast Artificial Neural Network library.

A new neural network with random weights can be created with the Fann::new method, or, for different network topologies, with its variants Fann::new_sparse and Fann::new_shortcut. Existing neural networks can be saved to and loaded from files.

Similarly, training data sets can be loaded from and saved to human-readable files, or training data can be provided directly to the network as slices of floating point numbers.

Example:

extern crate fann;
use fann::{ActivationFunc, Fann, TrainAlgorithm, QuickpropParams};

fn main() {
   // Create a new network with two input neurons, a hidden layer with three neurons, and one
   // output neuron.
   let mut fann = Fann::new(&[2, 3, 1]).unwrap();
   // Configure the activation functions for the hidden and output neurons.
   fann.set_activation_func_hidden(ActivationFunc::SigmoidSymmetric);
   fann.set_activation_func_output(ActivationFunc::SigmoidSymmetric);
   // Use the Quickprop learning algorithm, with default parameters.
   // (Otherwise, Rprop would be used.)
   fann.set_train_algorithm(TrainAlgorithm::Quickprop(Default::default()));
   // Train for up to 500000 epochs, displaying progress information after intervals of 1000
   // epochs. Stop when the network's error on the training data drops to 0.001.
   let max_epochs = 500000;
   let epochs_between_reports = 1000;
   let desired_error = 0.001;
   // Train directly on data loaded from the file "xor.data".
   fann.on_file("test_files/xor.data")
       .with_reports(epochs_between_reports)
       .train(max_epochs, desired_error).unwrap();
   // The network now approximates the XOR problem:
   assert!(fann.run(&[-1.0,  1.0]).unwrap()[0] > 0.9);
   assert!(fann.run(&[ 1.0, -1.0]).unwrap()[0] > 0.9);
   assert!(fann.run(&[ 1.0,  1.0]).unwrap()[0] < 0.1);
   assert!(fann.run(&[-1.0, -1.0]).unwrap()[0] < 0.1);
}

FANN also supports cascade training, where the network’s topology is changed during training by adding additional neurons:

extern crate fann;
use fann::{ActivationFunc, CascadeParams, Fann};

fn main() {
   // Create a new network with two input neurons and one output neuron.
   let mut fann = Fann::new_shortcut(&[2, 1]).unwrap();
   // Use the default cascade training parameters, but a higher weight multiplier:
   fann.set_cascade_params(&CascadeParams {
                                weight_multiplier: 0.6,
                                ..CascadeParams::default()
                            });
   // Add up to 50 neurons, displaying progress information after each.
   // Stop when the network's error on the training data drops to 0.001.
   let max_neurons = 50;
   let neurons_between_reports = 1;
   let desired_error = 0.001;
   // Train directly on data loaded from the file "xor.data".
   fann.on_file("test_files/xor.data")
       .with_reports(neurons_between_reports)
       .cascade()
       .train(max_neurons, desired_error).unwrap();
   // The network now approximates the XOR problem:
   assert!(fann.run(&[-1.0,  1.0]).unwrap()[0] > 0.9);
   assert!(fann.run(&[ 1.0, -1.0]).unwrap()[0] > 0.9);
   assert!(fann.run(&[ 1.0,  1.0]).unwrap()[0] < 0.1);
   assert!(fann.run(&[-1.0, -1.0]).unwrap()[0] < 0.1);
}

Structs

BatchParams

CascadeParams

Parameters for cascade training.

Fann

FannError

FannTrainer

A training configuration. Create this with Fann::on_data or Fann::on_file and run the training with train.

IncrementalParams

QuickpropParams

RpropParams

TrainData

Enums

ActivationFunc

The activation functions used for the neurons during training. They can either be set for a group of neurons using set_activation_func_hidden and set_activation_func_output, or for a single neuron using set_activation_func.

CallbackResult

ErrorFunc

Error function used during training.

FannErrorType

NetType

Network types

StopFunc

Stop critieria for training.

TrainAlgorithm

The Training algorithms used when training on fann_train_data with functions like fann_train_on_data or fann_train_on_file. The incremental training alters the weights after each time it is presented an input pattern, while batch only alters the weights once after it has been presented to all the patterns.

Type Aliases

Connection

FannResult

FannType

The type of weights, inputs and outputs in a neural network. It is defined as c_float by default, and as c_double if the double feature is configured.