//
public RProp(BackPropNet net, string patterns, string targets)
: base(net, patterns, targets)
{
try
{
// allocate mem for oldTotalDelta
oldTotalDelta = new double[net.Depth-1][];
totalDelta = new double[net.Depth-1][];
biasLearnRates = new double[net.Depth-1][];
oldTotalErrGrad = new double[net.Depth-1][,];
totalErrGrad = new double[net.Depth-1][,];
weightLearnRates = new double[net.Depth-1][,];
//
for(uint i=0; i< net.Depth-1; i++)
{
oldTotalDelta[i] = new double[net.GetLayer(i+1).GetSize()];
totalDelta[i] = new double[net.GetLayer(i+1).GetSize()];
biasLearnRates[i] = new double[net.GetLayer(i+1).GetSize()];
oldTotalErrGrad[i] = new double[errGrad[i].GetLength(0), errGrad[i].GetLength(1)];
totalErrGrad[i] = new double[errGrad[i].GetLength(0), errGrad[i].GetLength(1)];
weightLearnRates[i]= new double[errGrad[i].GetLength(0), errGrad[i].GetLength(1)];
}
//
// initialize all learning rates to 0.1
//
InitializeLearnRates();
}
catch(Exception ex)
{
Console.WriteLine(ex.Message);
throw;
}
}
//
public SimpleBPM(BackPropNet net, string patterns, string targets)
: base(net, patterns, targets)
{
//
try
{
// allocate memory for previous biases
pBiasUpdates = new double[net.Depth-1][];
//
for(uint i=0; i<net.Depth-1; i++)
pBiasUpdates[i] = new double[net.GetLayer(i+1).GetSize()];
//
// allocate memory for previous weights
pWeightUpdates = (double[][,])this.errGrad.Clone(); // it should be tested
//
}
catch(System.Exception ex)
{
Console.WriteLine(ex.Message);
throw;
}
}
protected uint vPeriod = 5; // validation period.
#endregion Fields
#region Constructors
//
//
public TrainingAlgorithm(BackPropNet net, string tPatterns, string tTargets)
{
uint depth = net.Depth;
//
try
{
// allocate mem for matrices of error gradients
errGrad = new double[depth-1][,]; // allocate mem for depth-1 vectors
//
for(uint i=0; i<depth-1; i++)
{
uint temp1 = net.GetLayer(i).GetSize();
uint temp2 = net.GetLayer(i+1).GetSize();
//
errGrad[i] = new double[temp1, temp2];
}
//
//
this.trPatterns = new DataSet(tPatterns);
this.trTargets = new DataSet(tTargets);
this.network = net;
//
if(trPatterns.GetNumberOfVectors() != trTargets.GetNumberOfVectors())
throw new Exception("Datasets have different number of vectors.");
if(trPatterns.GetDataValidation() != true || trTargets.GetDataValidation() != true)
throw new Exception("Invalid Data");
if(trPatterns.GetLengthOfVectors() != net.GetLayer(0).GetSize() ||
trTargets.GetLengthOfVectors() != net.GetLayer(net.Depth-1).GetSize())
throw new Exception("Incosistent Data/Inputs or Data/Outputs");
}
catch(Exception ex)
{
Console.WriteLine(ex.Message);
throw;
}
}
//
public TrainingAlgorithm(BackPropNet net, string tPatterns, string tTargets, string vPatterns, string vTargets)
: this(net, tPatterns, tTargets)
{
this.SetValidationData(vPatterns, vTargets);
}
//
public static void BinarySave(string filename, BackPropNet network)
{
try
{
FileStream fs = new FileStream(filename, FileMode.Create);
BinaryFormatter bf = new BinaryFormatter();
//
bf.Serialize(fs, network);
fs.Flush();
fs.Close();
}
catch(System.Exception e)
{
Console.WriteLine(e.Message);
throw;
}
}
//
public SimpleBPM(BackPropNet net, string tPatterns, string tTargets, string vPatterns, string vTargets)
: this(net, tPatterns, tTargets)
{
//
this.SetValidationData(vPatterns, vTargets);
}
private void button2_Click(object sender, System.EventArgs e)
{
try
{
if(of.ShowDialog() == DialogResult.OK && of.FileName != "")
{
network = BackPropNet.BinaryLoad(of.FileName);
rtb1.Text = "Network: "+network.GetStructure();
}
patterns = null;
targets = null;
}
catch(System.Exception ex)
{
MessageBox.Show(ex.Message, "Exception", MessageBoxButtons.OK, MessageBoxIcon.Error);
network = null;
}
//
}
//
private void button5_Click(object sender, System.EventArgs e)
{
string filename = null;
//
if(sf.ShowDialog() == DialogResult.OK && of.FileName != " ")
{
filename = sf.FileName;
//
try
{
if(patterns != null)
{
if(targets != null)
network.Simulate(patterns, targets, filename);
else
network.Simulate(patterns, filename);
}
//
MessageBox.Show("Simulation Finished", "Attention", MessageBoxButtons.OK, MessageBoxIcon.Exclamation);
}
catch(System.Exception ex)
{
MessageBox.Show(ex.Message, "Exception", MessageBoxButtons.OK, MessageBoxIcon.Error);
patterns = null;
patterns = null;
network = null;
this.button3.Enabled = false;
this.button4.Enabled = false;
this.button5.Enabled = false;
}
}
//
if(targets != null)
{
NeuralLib.DataSet dset = new NeuralLib.DataSet(filename);
//
double[] dummy;
//
chart1.Series.Clear();
chart1.OpenData(COD.Values, 0, dset.GetNumberOfVectors());
if(network.GetLayer(network.Depth-1).GetSize() == 1)
{
// network has one output neuron.
chart1.AxisY.Title.Text = "Real Output-Simulated Output";
double sse = 0.0;
for(int i=0; i<dset.GetNumberOfVectors(); i++)
{
dummy = dset.GetDataVector(i);
chart1.Value[0, i] = targets.GetDataVector(i)[0];
chart1.Value[1, i] = dummy[0];
sse += dummy[dummy.Length-1];
}
//
label1.Text = "SSE: "+sse.ToString("F6")+" MSE: "+(sse/dset.GetNumberOfVectors()).ToString("F6")+" RMSE: "+
Math.Sqrt(sse/dset.GetNumberOfVectors()).ToString("F6");
}
else
{
chart1.AxisY.Title.Text = "RMSE";
double sse = 0.0;
for(int i=0; i<dset.GetNumberOfVectors(); i++)
{
dummy = dset.GetDataVector(i);
chart1.Value[0, i] = Math.Sqrt(dummy[dummy.Length-1]);
sse += dummy[dummy.Length-1];
}
label1.Text = "SSE: "+sse.ToString("F6")+" MSE: "+(sse/dset.GetNumberOfVectors()).ToString("F6")+" RMSE: "+
Math.Sqrt(sse/dset.GetNumberOfVectors()).ToString("F6");
}
//
label1.Left = chart1.Left+chart1.Width/2-label1.Width/2;
chart1.CloseData(COD.Values);
chart1.RecalcScale();
chart1.Refresh();
}
}
//
public RProp(BackPropNet net, string tPatterns, string tTargets, string vPatterns, string vTargets)
: this(net, tPatterns, tTargets)
{
this.SetValidationData(vPatterns, vTargets);
}
private void bStart_Click(object sender, System.EventArgs e)
{
//
System.IO.StreamWriter sw = new System.IO.StreamWriter(results);
sw.AutoFlush = true;
keepTraining = true;
//
int numberOfTimesToTrain = Convert.ToInt32(cbTrainTimes.SelectedItem);
//
uint inputSize = Convert.ToUInt32(cbInputNeurons.SelectedItem);
uint outputSize = Convert.ToUInt32(cbOutputNeurons.SelectedItem);
//
uint minNeurons = Convert.ToUInt32(cbMinHiddenNeurons.SelectedItem);
uint maxNeurons = Convert.ToUInt32(cbMaxHiddenNeurons.SelectedItem);
//
tbResults.Text = "Hidden Neurons ---- Mean Train Error --- Mean Validation Error ------ Mean Test Error";
uint n = minNeurons-1;
while(keepTraining)
{
n++;
if(n > maxNeurons)
break;
//
double meanTrError = 0;
double meanVaError = 0;
double meanTeError = 0;
//
Application.DoEvents();
for(int i=0; i<numberOfTimesToTrain; i++)
{
nnet = new BackPropNet(inputSize, outputSize, n);
//
if(cbHiddenActFunction.SelectedItem.ToString() != "LogSig")
{
nnet.SetHiddenActivationFunction(ActivationFunction.TanSig);
}
//
if(cbOutputActFunction.SelectedItem.ToString() == "TanSig")
nnet.SetOutputActivationFunction(ActivationFunction.TanSig);
else if(cbOutputActFunction.SelectedItem.ToString() == "Linear")
nnet.SetOutputActivationFunction(ActivationFunction.Linear);
//
nnet.Randomize(-0.2, 0.2);
//
if(cbAlgorithm.Text == "RProp")
{
algorithm = new RProp(nnet, trainingInputs, trainingTargets);
}
else
{
algorithm = new SimpleBPM(nnet, trainingInputs, trainingTargets);
((SimpleBPM)algorithm).Momentum = Convert.ToDouble(this.tbMomentum.Text);
((SimpleBPM)algorithm).LearnRate = Convert.ToDouble(this.tbLearnRate.Text);
}
//
algorithm.Epochs = Convert.ToUInt32(this.tbEpochs.Text);
algorithm.MaxFailures = Convert.ToUInt32(this.tbMaxFails.Text);
//
if(valIsOk)
algorithm.SetValidationData(validationInputs, validationTargets);
//
if(cbAlgorithm.Text == "RProp")
{
((RProp)algorithm).Train();
}
else
{
((SimpleBPM)algorithm).Train();
}
//
double trError = algorithm.TrainingPerformance;
double vaError = Double.NaN;
double teError = Double.NaN;
//
if(valIsOk)
{
vaError = algorithm.ValidationPerformance;
meanVaError += vaError;
}
//
if(testIsOk)
{
teError = nnet.Simulate(new NeuralLib.DataSet(testInputs), new NeuralLib.DataSet(testTargets));
meanTeError += teError;
}
//
meanTrError += trError;
Application.DoEvents();
}
//
meanTrError = meanTrError/numberOfTimesToTrain;
//
if(valIsOk)
meanVaError = meanVaError/numberOfTimesToTrain;
//
if(testIsOk)
meanTeError = meanTeError/numberOfTimesToTrain;
//
tbResults.AppendText("\n"+"\t"+n.ToString()+"\t\t"+meanTrError.ToString("E5")+"\t\t"+
meanVaError.ToString("E6")+"\t\t"+meanTeError.ToString("E6"));
//
sw.WriteLine(nnet.GetStructure()+"\t\t"+meanTrError.ToString("E5")+"\t\t"+
meanVaError.ToString("E6")+"\t\t"+meanTeError.ToString("E6"));
Application.DoEvents();
}
//
if(cbUseSecondLayer.Checked == true)
{
uint k = 2;
while(keepTraining)
{
k++;
if(k > n)
break;
//
double meanTrError = 0;
double meanVaError = 0;
double meanTeError = 0;
//
Application.DoEvents();
for(int i=0; i<numberOfTimesToTrain; i++)
{
nnet = new BackPropNet(inputSize, outputSize, n-1, k-1);
//
if(cbHiddenActFunction.SelectedItem.ToString() != "LogSig")
{
nnet.SetHiddenActivationFunction(ActivationFunction.TanSig);
}
//
if(cbOutputActFunction.SelectedItem.ToString() == "TanSig")
nnet.SetOutputActivationFunction(ActivationFunction.TanSig);
else if(cbOutputActFunction.SelectedItem.ToString() == "Linear")
nnet.SetOutputActivationFunction(ActivationFunction.Linear);
//
nnet.Randomize(-0.2, 0.2);
//
if(cbAlgorithm.Text == "RProp")
{
algorithm = new RProp(nnet, trainingInputs, trainingTargets);
}
else
{
algorithm = new SimpleBPM(nnet, trainingInputs, trainingTargets);
((SimpleBPM)algorithm).Momentum = Convert.ToDouble(this.tbMomentum.Text);
((SimpleBPM)algorithm).LearnRate = Convert.ToDouble(this.tbLearnRate.Text);
}
//
algorithm.Epochs = Convert.ToUInt32(this.tbEpochs.Text);
algorithm.MaxFailures = Convert.ToUInt32(this.tbMaxFails.Text);
//
if(valIsOk)
algorithm.SetValidationData(validationInputs, validationTargets);
//
if(cbAlgorithm.Text == "RProp")
{
((RProp)algorithm).Train();
}
else
{
((SimpleBPM)algorithm).Train();
}
//
double trError = algorithm.TrainingPerformance;
double vaError = Double.NaN;
double teError = Double.NaN;
//
if(valIsOk)
{
vaError = algorithm.ValidationPerformance;
meanVaError += vaError;
}
//
if(testIsOk)
{
teError = nnet.Simulate(new NeuralLib.DataSet(testInputs), new NeuralLib.DataSet(testTargets));
meanTeError += teError;
}
//
meanTrError += trError;
Application.DoEvents();
}
//
meanTrError = meanTrError/numberOfTimesToTrain;
//
if(valIsOk)
meanVaError = meanVaError/numberOfTimesToTrain;
//
if(testIsOk)
meanTeError = meanTeError/numberOfTimesToTrain;
//
tbResults.AppendText("\n"+"\t"+nnet.GetStructure()+"\t\t"+meanTrError.ToString("E5")+"\t\t"+
meanVaError.ToString("E6")+"\t\t"+meanTeError.ToString("E6"));
//
sw.WriteLine(nnet.GetStructure()+"\t\t"+meanTrError.ToString("E5")+"\t\t"+
meanVaError.ToString("E6")+"\t\t"+meanTeError.ToString("E6"));
Application.DoEvents();
}
}
sw.Flush();
sw.Close();
}
private void bLoad_Click(object sender, System.EventArgs e)
{
//
if(of.ShowDialog() == DialogResult.OK)
{
try
{
Network = BackPropNet.BinaryLoad(of.FileName);
info.ShowDialog(this);
}
catch(System.Exception ex)
{
MessageBox.Show(ex.Message, "Exception", MessageBoxButtons.OK, MessageBoxIcon.Error);
Network = null;
}
}
}
