private void UnitTest1()
{
OutputText.Text = "";
NetworkDocument doc = new NetworkDocument();
InputNeuron i1 = new InputNeuron() {OutputValue = 0.7};
InputNeuron i2 = new InputNeuron() {OutputValue = 0.1};
InputNeuron i3 = new InputNeuron() {OutputValue = 0.3};
doc.Inputs.Add(i1);
doc.Inputs.Add(i2);
doc.Inputs.Add(i3);
OutputNeuron o = new OutputNeuron();
doc.Outputs.Add(new OutputNeuron());
NeuronHelper.ConnectNodes(i1, o, -0.3);
NeuronHelper.ConnectNodes(i2, o, 3.1);
NeuronHelper.ConnectNodes(i3, o, 0.5);
double expected = 0.25;
double answer = o.SumInputs(true);
OutputText.Text += "Summation test 3 inputs values 0.7, 0.1,0.3 respective weights 0.3,3.1,0.5" +
Environment.NewLine;
OutputText.Text += "Expected = " + expected + Environment.NewLine;
OutputText.Text += "Actual = " + answer + Environment.NewLine;
}
//
// apply momentum to help get out of dips
//
public void AdjustFromWeights(NetworkDocument doc)
{
foreach (Synapse s in BackwardConnections)
{
double last = s.Weight; //save the current weight for the momentum calculations
//http://www.cheshireeng.com/Neuralyst/nnbg.htm
//wij = w'ij + (1 - M).LR.ej.Xj+M.(w'ij - w''ij)
s.Weight = s.Weight +
((1 - doc.Momentum) * doc.LearningRate * ErrorValue * s.FromNode.OutputValue) +
(doc.Momentum * (s.Weight - s.LastWeight));
s.LastWeight = last;
Debug.WriteLine("Adjust from weights : " + s.Weight);
}
if (doc.UsesBias)
BiasValue = BiasValue + (doc.LearningRate * ErrorValue * ErrorValue) * 1;
}
public bool WithinTolerance(NetworkDocument doc)
{
bool ret = false;
double lowRange = 0 - doc.Tolerance;
double highRange = 0 + doc.Tolerance;
if (Math.Abs(ExpectedValue) < 0.1)
{
if (OutputValue >= lowRange && OutputValue <= highRange)
ret = true;
}
else
{
lowRange = ExpectedValue - (Math.Abs(ExpectedValue) * doc.Tolerance);
highRange = ExpectedValue + (Math.Abs(ExpectedValue) * doc.Tolerance);
if (OutputValue >= lowRange && OutputValue <= highRange)
ret = true;
}
return ret;
}
protected override void OnNavigatedTo(NavigationEventArgs e)
{
base.OnNavigatedTo(e);
CurrentDocument = (NetworkDocument)e.Parameter;
}
//
// The calculation for the input neuron is the same as the input - In many ways the code below only illustrates this serving little function
// but clarifies the use of activation identity
//
public override double CalcOutput(NetworkDocument doc)
{
OutputValue = (NeuronHelper.GetActivation(OutputValue, Enumerations.ActivationFunctions.ActivationIdentity,doc.Threshold));
return OutputValue;
}
//
// Using the error from the node adjust the weights before (going towards input layer)
//
public virtual void AdjustFromWeights1(NetworkDocument doc)
{
foreach (Synapse s in BackwardConnections)
{
//this has been unit tested and is correct
// DiagnosticData.addLogged(new DiagData("AdjustFromWeights", DiagData.DIAG_TYPE_INFO, "Weight before = " + String.valueOf(s.Weight)));
s.Weight = s.Weight + (doc.LearningRate*ErrorValue)*s.FromNode.OutputValue;
// DiagnosticData.addLogged(new DiagData("AdjustFromWeights", DiagData.DIAG_TYPE_INFO, "Weight after = " + String.valueOf(s.Weight)));
}
// DiagnosticData.addLogged(new DiagData("AdjustFromWeights", DiagData.DIAG_TYPE_INFO, "Bias before = " + String.valueOf(BiasValue)));
if (doc.UsesBias)
BiasValue = BiasValue + (doc.LearningRate * ErrorValue) * 1;
// DiagnosticData.addLogged(new DiagData("AdjustFromWeights", DiagData.DIAG_TYPE_INFO, "Bias after = " + String.valueOf(BiasValue)));
}
public virtual double CalcOutput(NetworkDocument doc)
{
double sum = SumInputs(doc.UsesBias);
OutputValue = (NeuronHelper.GetActivation(sum, doc.ActivationFunction,doc.Bias));
return OutputValue;
}
private void UnitTest3()
{
double output1;
double output2;
OutputText.Text = "";
NetworkDocument doc = new NetworkDocument() {Bias = -2.0,Momentum = 0.05,LearningRate = 0.5};
doc.ActivationFunction = Enumerations.ActivationFunctions.ActivationSigmoidal;
doc.UsesBias = true;
InputNeuron i0 = new InputNeuron() { OutputValue = 0.1 };
InputNeuron i1 = new InputNeuron() { OutputValue = 0.9 };
doc.Inputs.Add(i0);
doc.Inputs.Add(i1);
doc.NumLayers = 2;//this should build the layers
HiddenNeuron l00 = new HiddenNeuron(){BiasValue = 2.0};
HiddenNeuron l01 = new HiddenNeuron() { BiasValue = 2.0 };
HiddenNeuron l10 = new HiddenNeuron() { BiasValue = 3.0 };
HiddenNeuron l11 = new HiddenNeuron() { BiasValue = -2.0 };
doc.Hidden[0].Add(l00);
doc.Hidden[0].Add(l01);
doc.Hidden[1].Add(l10);
doc.Hidden[1].Add(l11);
OutputNeuron o = new OutputNeuron() {ExpectedValue = 0.9,BiasValue = -2};
doc.Outputs.Add(o);
//connect layer 1
NeuronHelper.ConnectNodes(i0, l00, -2);
NeuronHelper.ConnectNodes(i0, l01, 3);
NeuronHelper.ConnectNodes(i1, l00, -2);
NeuronHelper.ConnectNodes(i1, l01, 3);
//connect layer 2
NeuronHelper.ConnectNodes(l00, l10, -2);
NeuronHelper.ConnectNodes(l00, l11, 2);
NeuronHelper.ConnectNodes(l01, l10, -4);
NeuronHelper.ConnectNodes(l01, l11, 2);
//output layer
NeuronHelper.ConnectNodes(l10, o, 3);
NeuronHelper.ConnectNodes(l11, o, 1);
output1 = l00.SumInputs(true);
output2 = l01.SumInputs(true);
WriteOutput(0, output1);
WriteOutput(5, output2);
output1 = l00.CalcOutput(doc);
output2 = l01.CalcOutput(doc);
WriteOutput(0.5, output1);
WriteOutput(0.99, output2);
output1 = l10.SumInputs(doc.UsesBias);
output2 = l11.SumInputs(doc.UsesBias);
WriteOutput(-1.972, output1);
WriteOutput(0.98, output2);
output1 = l10.CalcOutput(doc);
output2 = l11.CalcOutput(doc);
WriteOutput(0.122, output1);
WriteOutput(0.728, output2);
output1 = doc.Outputs[0].SumInputs(doc.UsesBias);
output2 = doc.Outputs[0].CalcOutput(doc);
WriteOutput(-0.906, output1);
WriteOutput(0.288, output2);
//backward pass
output1 = doc.Outputs[0].CalcError();
WriteOutput(0.125, output1);
output1 = l10.CalcError();
WriteOutput(0.040, output1);
output1 = l11.CalcError();
WriteOutput(0.025, output1);
output1 = l00.CalcError();
WriteOutput(0, output1);
output1 = l01.CalcError();
WriteOutput(0, output1);
o.AdjustFromWeights(doc);
l10.AdjustFromWeights(doc);
l11.AdjustFromWeights(doc);
l00.AdjustFromWeights(doc);
l01.AdjustFromWeights(doc);
}
public async void LoadLastDocument()
{
if (ApplicationData.Current.LocalSettings.Values.ContainsKey("LastDocName"))
{
_lastDocName = (String)(ApplicationData.Current.LocalSettings.Values["LastDocName"]);
StorageFolder localFolder = ApplicationData.Current.LocalFolder;
try
{
StorageFile sampleFile = await localFolder.GetFileAsync(_lastDocName);
await FileIO.ReadTextAsync(sampleFile);
}
catch (Exception)
{
// No file to load or error loading it
}
}
else
{
CurrentDocument = new NetworkDocument
{
NumInput = Defaults.DefaultInputs,
NumOutput = Defaults.DefaultOutputs,
NumHidden = Defaults.DefaultHidden,
NumLayers = Defaults.DefaultLayers,
LearningRate = Defaults.DefaultLearningRate,
Momentum = Defaults.DefaultMomentum,
ActivationFunction = Enumerations.ActivationFunctions.ActivationSigmoidal,
Tolerance = Defaults.DefaultTolerance,
Threshold = Defaults.DefaultThreshold,
NetworkName = Defaults.DefaultName,
UsesBias = Defaults.UsesBias
};
}
SetDataContextsUi();
}
