C# DLL library from scratch with demo UI app for creating, training and validation neural network models
Applied as one of the methods in a PipeMonitor project for real-time object acoustic recognition with neural networks
- Building neural networks with different structure in object oriented style
- Training neural networks with a given learning rate, iterations, epochs, input signals
- Four demo learning modes included:
- Training a model to recognize images with rotation. MNIST handwritten digits database tested (png images)
- Training a model to recognize noisy sinusoidal signals
- Drawing with a mouse in interactive mode to train a model to recognize any kind of user paintings
- Signal approximation with neural network
- Validation with real-time monitoring of errors, weights values and neuron outputs on charts
- Saving pre-trained models to a file at any training state
- Loading pre-trained models from files
- DLL to use neural networks in your project
- Currently only a multilayer perceptron network is supported
Unpack "unpack_to_exe_folder.zip" with mnist images, user icons and pre-trained models to "Debug" and "Release" folders.
Launch "MyNeuralNetwork.sln" and compile.
-
Created in Visual Studio Community 2019
-
.Net Framework 4.7.2
-
MathNet.Numerics is used for weights generation with normal distribution
Code example of 800x400x200x30 multilayer perceptron training on 30 different noisy signals (1 to 31Hz) with 0.005 learning rate / 5 epochs x 1000 samples
Accuracy 99,17%, see results below
// - Add NeuralLibrary.dll and MathNet.Numerics.dll references to project in Visual Studio
// - Add "using NeuralLibrary;"
// - Add "using MathNet.Numerics;"
// To simplify the example, each network output value corresponds to [frequency value + 1 Hz] signal class
int epochs = 5;
int iterations = 1000;
int inputNeurons = 784; // Inputs
int[] hiddenLayers = { 400, 200 }; // Two hidden layers array with 400 and 200 neurons
int outputNeurons = 30; // Outputs
Random rndClass = new Random();
Random rndAmplitude = new Random();
network = new Network(inputNeurons, hiddenLayers, outputNeurons); // Create a Network object
signal = new Signal(inputNeurons); // Create a signal object
network.LearningRate = 0.005; // Default value: 0.01
for (int i = 0; i < epochs; i++)
{
for (int j = 0; j < iterations; j++)
{
int classToTrain = rndClass.Next(0, 30); // Randomize desired output (network performs badly without shuffling data)
signal.GenerateSinus(classToTrain + 1, rndAmplitude); // Generate noisy amplitudes with [classToTrain + 1 Hz] freq
network.SetTarget(classToTrain); // Set desired output and reset others
network.SendSignalsToInputLayer(signal.Amplitude); // Send signal to input layer
network.Pass(); // Forward propagation -> back propagation with stochastic gradient descent
}
}
Loading images and custom training data
signal.ImageFromFile(pathToImage, rotateAngle); // load image according to classToTrain value
network.SendSignalsToInputLayer(myDoubleArray); // using any double[] array with your training data
Instead of Pass() method, you can call separately
network.ForwardPropagation();
network.FindNetworkOutputError();
network.NeuronErrorDistribution();
network.RecalculateWeights();
network.CleanOldData();
Get current network output values and errors
double[] outputs = network.CurrentNetworkOutput;
double[] errors = network.CurrentNetworkOutputError;
Code example results: training/validation on 30 noisy signal classes with 800x400x200x30 network, 5 epochs x 1000 samples
