BrabeNetz

BrabeNetz is a supervised neural network written in C++, aiming to be as fast as possible¹ by using C arrays and only raw values instead of objects (May be hard to read because of that)

BrabeNetz has no bounds- or error-checking for performance reasons, be careful what you feed it.

I've currently trained it to solve XOR.

You can save the network's state to disk with the network::save(string) function, and load it on next startup. An example of this can be found here.

Benchmarks

It's pretty fast (TODO)

Goals

• Fast Feed-Forward algorithm
• Fast Backwards-Propagation algorithm
• Easy to use (Inputs, outputs)
• C arrays > std::vector
• Fast network state saving via state.nn file (Weights, Biases, Sizes)
• Multithreaded if worth the spawn-overhead (std::thread or NVIDIA CUDA)
• Scalability (Neuron size, Layer count) - only limited by hardware

Specs

• Randomly generated values to begin with
• Easily save/load with network::save(string)/network::load(string)
• Sigmoid squashing function (TODO: ReLU?)
• Biases for each neuron
• network_topology helper objects for loading/saving state (user friendly)

Usage

1. Constructors

   • network(initializer_list<int>): Create a new neural network with the given topology vector and fill it with random numbers ({ 2, 3, 4, 1} = 2 Input, 3 Hidden, 4 Hidden, 1 Output)
   • network(network_topology&): Create a new neural network with the given network topology and load it's values
   • network(string): Create a new neural network with the given path to the sate.nn file and load it. (See network_topology::load(string) for more info)

2. Functions

   • double train(double* input_values, int length, int& out_length): Feed the network input_values and return an array of output values (size of output layer in topology) (where out_length will be set to the length of returned output values)
   • double train(double* input_values, int length, double* expected_output): Feed the network input_values and compare predicted output with expected_output, Backwards-Propagate (adjust weights/biases) if needed. Returns the total error of the output layer.
   • void save(string path): Save the current network state (topology, weights, biases) to disk (with the given path or default: state.nn)
   • void set_learnrate(double value): Set the learn rate of the network (used by train(..) function). Should always be 1 / (total train times + 1)

¹: I'm new to C/C++ so; as fast as I can do it