A zero-dependency, portable, and efficient library for building, running, and training feed-forward and recurrent artificial neural networks.
Panann is adapted from the neural network component of a hobby project named PAN built way back in 2008. For the tenth anniversary of the project, panann was split out into its own library and re-implemented in c++17. Panann is the PAN artificial neural network (PanAnn) component.
Panann has a simple interface for constructing and running artificial neural networks.
// Make XOR 2-bit training set.
TrainingData data;
data.resize(4);
data[0]._input = { 0, 0 };
data[0]._output = { 0 };
data[1]._input = { 1, 1 };
data[1]._output = { 0 };
data[2]._input = { 1, 0 };
data[2]._output = { 1 };
data[3]._input = { 0, 1 };
data[3]._output = { 1 };
// Build a simple, multi-layer feed-forward network.
NeuralNetwork nn;
nn.SetInputNeuronCount(2);
nn.SetOutputNeuronCount(1);
nn.AddHiddenLayer(5);
nn.AddHiddenLayer(5);
nn.Construct();
nn.SetTrainingAlgorithmType(NeuralNetwork::TrainingAlgorithmType::SimulatedAnnealingResilientBackpropagation);
nn.InitializeWeightsRandom();
std::cout << "Error before training: " << nn.GetError(data) << std::endl;
nn.Train(&data, 100000);
std::cout << "Error after training for 100000 epochs: " << nn.GetError(&data) << std::endl;You can build panann on any platform with a compiler which supports c++17 language standards mode. The library is designed to be portable and easy to add to your project. We do not release binaries here, but panga compiles into a static library which can be added as a dependency. Add the panann cmake file to your build system and you should be ready to use panann.
Windows 10
- CMake 3.17.0
- Visual Studio 2019 16.11.23
Ubuntu 18.04
- CMake 3.16.3
- Clang 10.0.0
The library ships with a simple test program in the panann/test folder.
> git clone https://github.com/boingoing/panann/panann.git
> mkdir panann/build
> cd panann/build
> cmake ..
> make
> ./panann_test