hnjm/machine.academy

Neural Network training library in C# with GPU acceleration

machine.academy

Neural network training library written in C# (.net Core 3.1) as a learning project.

Demos:

Handwritten digit recognizer web app

Car racing with Evolutional algorithm

Features:

• GPU accelerated Backpropagation (using OpenCL)
• Cost functions:
  • Mean-Squared
  • Cross-entropy
• Activation functions:
  • Sigmoid
• Optional L1 and L2 regularization
• Xavier weight initialization
• Stochastic gradient descent with shuffled mini-batches

Sample code:

Initializing a random network:

var layerConfig = new int[]{ 784, 32, 32, 10 };
var network = Network.CreateNetworkInitRandom( layerConfig, new SigmoidActivation() );

Initializing from JSON:

string jsonData = "{ ... }";
var network = Network.CreateNetworkFromJSON( jsonData ); 

Exporting to JSON:

string json = network.ExportToJSON();

Calculating an output to a given input:

float[] input = { ... };
ComputeDevice computeDevice = ComputeDeviceFactory.GetFirstOpenCLDevice(); //Get an OpenCL device
float[] results = network.Compute( input, computeDevice );

Training

ComputeDevice computeDevice = ComputeDeviceFactory.GetFirstOpenCLDevice(); //Get an OpenCL device
List<TrainingSuite.TrainingData> trainingData = new List<TrainingSuite.TrainingData>();

//Set up training examples
for (int i = 0; i < 10000; ++i)
{
    float[] trainingInput = ...;
    float[] desiredOutput = ...;
    trainingData.Add( new TrainingSuite.TrainingData( trainingInput, desiredOutput ) );
}

TrainingSuite trainingSuite = new TrainingSuite( trainingData );

//Set up training configuration
trainingSuite.config.epochs = 100;
trainingSuite.config.shuffleTrainingData = true;
trainingSuite.config.miniBatchSize = 50;
trainingSuite.config.learningRate = 0.005f;
trainingSuite.config.costFunction = new CrossEntropyErrorFunction();
trainingSuite.config.regularization = TrainingSuite.TrainingConfig.Regularization.L2;
trainingSuite.config.regularizationLambda = 0.5f;

var trainingPromise = network.Train( trainingSuite, computeDevice );

trainingPromise.Await();

Projects in the repo

Macademy

The neural network library

OpenCl.DotNetCore.Interop

Part of the following library: https://github.com/lecode-official/opencl-dotnet. Used for OpenCL access from .net Core

Testing

Tests for the library

TestConsole

A console command line for testing

NumberRecognizer (uses WinForms)

A sample app that can load the MNIST digit dataset and trains a network.

Sandbox (uses WinForms)

A test app for development with no specific purpose but to try our some features using a ui.

docs/WebApps/NumberRecognizer

An HTML web application that uses an already trained network to recognize a drawn number.

docs/WebApps/Evo

An HTML web application that uses neural networks and an evolutional algorithm for training racecars to learn driving on various racetracks.

Dependencies

The following dependencies are set up in nuget:

• Newtonsoft.Json

This project also includes parts of opencl-dotnet: https://github.com/lecode-official/opencl-dotnet

Future plans:

• Optimize memory layout (to reduce number of buffer copies)
• New features: Dropout, Softmax layers
• More activation functions
• Convolutional layer
• LSTM networks

Resources

These are some nice resources for learning about Neural Networks:

• Michael A. Nielsen, Neural Networks and Deep Learning
• 3Blue1Brown's But what is a neural network?