protected const int MAX_EPOCHS = 1000; //Limiting the number of iterations in the process.
/*
* CONSTRUCTORS
*/
public NeuralNetwork(DataSetReader dataSet, double[] target, double bias, double eta)
{
this.target = target;
this.data = dataSet.data;
this.featureMask = VectorTools.sequence(dataSet.features);
this.classMask = VectorTools.sequence(dataSet.classes);
this.weight = VectorTools.ones(this.featureMask.Length + 1); //+1 for bias.
this.samples = dataSet.samples;
this.eta = eta;
this.bias = bias;
}
public NeuralNetwork(DataSetReader dataSet, double[] target, double[] weight, double eta, double bias)
{
this.target = target;
this.data = dataSet.data;
this.featureMask = VectorTools.sequence(dataSet.features);
this.classMask = VectorTools.sequence(dataSet.classes);
this.samples = dataSet.samples;
this.weight = weight;
this.eta = eta;
this.bias = bias;
}
public override double classify(double[] input)
{
if (input.Length > this.featureMask.Length)
{
throw new ArgumentOutOfRangeException("Input features is bigger than features in feature mask");
}
input = VectorTools.prepend(input, this.bias); //Add the bias.
double net = VectorTools.multiply(this.weight, input);
return(getvalue(net));
}
public override void train(int trainCount)
{
int epochs = 0;
bool moreError = true;
double minError = 1E-2;
double[][] error = new double[this.classMask.Length][];
double[] mse = new double[MAX_EPOCHS]; //Mean square error.
bool weightChanged = true;
int classIndex;
double[] lineData;
/*REAL WORK*/
while (weightChanged && epochs < NeuralNetwork.MAX_EPOCHS && moreError)
{
weightChanged = false;
for (int i = 0; i < this.classMask.Length; i++) //Class index.
{
error[i] = new double[trainCount];
classIndex = (int)classMask[i] - 1;
for (int j = 0; j < trainCount; j++) //Data index.
{
lineData = VectorTools.trim(this.data[classIndex][j], this.featureMask);
lineData = VectorTools.prepend(lineData, this.bias); //Prepend the bias.
double net = VectorTools.multiply(this.weight, lineData);
net = getvalue(net);
if (net != this.target[i])
{
weightChanged = true;
lineData = VectorTools.trim(this.data[classIndex][j], this.featureMask);
lineData = VectorTools.prepend(lineData, this.bias);
error[i][j] = this.target[i] - net;
double[] mulOut = VectorTools.multiply(lineData, error[i][j] * this.eta);
error[i][j] = 0.5 * error[i][j] * error[i][j];
this.weight = VectorTools.sum(this.weight, mulOut);
}
}
} //End of inner for.
double[] temp = VectorTools.get1D(error);
mse[epochs] = VectorTools.mean(temp);
if (mse[epochs] < minError)
{
moreError = false;
}
epochs++;
} //End of outer while.
}
public NeuralNetwork(DataSetReader dataSet, double[] target, double eta, double bias, double[] featureMask, double[] classMask)
{
if (featureMask.Length > dataSet.features || featureMask.Length < 1)
{
throw new ArgumentOutOfRangeException("Number of features specified is not applicable");
}
this.target = target;
this.data = dataSet.data;
this.featureMask = featureMask;
this.classMask = classMask;
this.weight = VectorTools.ones(this.featureMask.Length + 1); //+1 for bias.
this.samples = dataSet.samples;
this.eta = eta;
this.bias = bias;
}
public void drawLine(NeuralNetwork network)
{
//Checks and getting ready:
double[] classMask = network.getClassMask(),
featureMask = network.getFeatureMask(),
weight = network.getWeight();
double bias = network.getBias(),
weightX = weight[0],
weightY = weight[1];
if (classMask.Length != 2)
{
throw new ArgumentOutOfRangeException("Mask must contain exactly two classes");
}
classMask[0]--; classMask[1]--; //For using as indices.
int classOne = (int)classMask[0],
classTwo = (int)classMask[1];
//Real work:
System.Collections.Generic.List <double[]>[] set =
dataSet.normalized ? dataSet.dataNorm : dataSet.data;
double maxX = VectorTools.max(set[classOne], featureMask),
minX = VectorTools.min(set[classOne], featureMask);
double xOne = (minX + weightX + bias),
xTwo = (maxX + weightY + bias),
yOne = ((-xOne * weightX) / weightY) - (bias / weightY), //Derived from: x*w1 + y*w2 = 0.
yTwo = ((-xTwo * weightX) / weightY) - (bias / weightY);
if (dataSet.normalized)
{
double[] newDataOne = dataSet.unNorm(new double[] { xOne, yOne }, featureMask);
double[] newDataTwo = dataSet.unNorm(new double[] { xTwo, yTwo }, featureMask);
xOne = newDataOne[0]; yOne = newDataOne[1];
xTwo = newDataTwo[0]; yTwo = newDataTwo[1];
}
Bitmap bitmap = this.pictureBox.Image as Bitmap;
Graphics g = Graphics.FromImage(bitmap);
Pen pen = new Pen(Color.Black);
g.DrawLine(pen, (float)xOne * factorX, (float)yOne * factorY, (float)xTwo * factorX, (float)yTwo * factorY);
}
/// <summary>Train the machine using the perceptron algorithm.</summary>
/// <param name="trainCount">Number of data set smaples to use in training.</param>
public override void train(int trainCount)
{
int epochs = 0;
bool weightChanged = true;
int classIndex;
double[] lineData;
/*REAL WORK*/
while (weightChanged && epochs < NeuralNetwork.MAX_EPOCHS)
{
weightChanged = false;
for (int i = 0; i < this.classMask.Length; i++) //Class index.
{
classIndex = (int)classMask[i] - 1;
for (int j = 0; j < trainCount; j++) //Data index.
{
lineData = VectorTools.trim(this.data[classIndex][j], this.featureMask);
lineData = VectorTools.prepend(lineData, this.bias); //Prepend the bias.
double net = VectorTools.multiply(this.weight, lineData);
int sgnOut = ActivationFunctions.signum(net);
if (sgnOut != this.target[i])
{
weightChanged = true;
lineData = VectorTools.trim(this.data[classIndex][j], this.featureMask);
lineData = VectorTools.prepend(lineData, this.bias);
double error = this.target[i] - sgnOut;
double[] mulOut = VectorTools.multiply(lineData, error * this.eta);
this.weight = VectorTools.sum(this.weight, mulOut);
}
}
} //End of inner for.
epochs++;
} //End of outer while.
}
