/* In this software sample, a neural network (from Drew Wilson's NeuralNetwork library)
* is trained using training data generated from sin(x) with x range from 0 to "maxX"
* and y range normalized between 0 and 1.
*
* The software creates a "data.csv" file of the neural network's approximation of the
* sine function within the given range.
*/
public static void Main(string[] args)
{
// initialize fields
DataIO.DataExport DE = new DataIO.DataExport();
MathClasses.Probability Prob = MathClasses.Probability.Instance;
MathClasses.LinearAlgebra LA = new MathClasses.LinearAlgebra();
NeuralNetworks.BackpropNetwork network = new NeuralNetworks.BackpropNetwork();
int trainDataSize = 10000;
int maxX = 10;
int testDataSize = 10000;
List <double[]> allInputs = new List <double[]>();
List <double[]> allOutputs = new List <double[]>();
double[] tempInput;
double[] tempOutput;
List <double[]> plotData = new List <double[]>();
// create training data from sin(x) function
for (int i = 0; i < trainDataSize; i++)
{
tempInput = new double[1];
tempOutput = new double[1];
tempInput[0] = Prob.NextDouble() * maxX;
tempOutput[0] = Math.Sin(tempInput[0]);
tempOutput[0] = (tempOutput[0] + 1) / 2; //normalize between 0 and 1
allInputs.Add(tempInput);
allOutputs.Add(tempOutput);
}
// set up network parameters (adjustable)
network.NumInputs = 1; // number of simulator inputs
network.NumHidden = 5; // number of hidden nodes per network in simulator ensemble
network.NumOutputs = 1; // number of outputs of simulator
network.WeightInitSTD = .75; // std for random weight initialization
network.Eta = 0.1; // learning rate for backprop
network.Episodes = 2000; // episodes for backprop
network.Momentum = 0.5; // momentum for backprop
network.Shuffle = NeuralNetworks.ToggleShuffle.no; // do we shuffle training data during backprop?
// give training data to network
network.Inputs = allInputs;
network.Outputs = allOutputs;
network.ValidationInputs = new List <double[]>();
network.ValidationOutputs = new List <double[]>();
network.TrainNetwork();
// test the network
tempInput = new double[1];
tempOutput = new double[1];
for (int i = 0; i < testDataSize; i++)
{
tempInput[0] = (i + 1) * (double)maxX / (double)testDataSize;
tempOutput = network.NeuralNet.ForwardPass(tempInput);
plotData.Add(LA.Join(tempInput, tempOutput));
}
DE.ExportData(plotData, "data");
}
