private void StartButtonClick(object sender, RoutedEventArgs e)
{
var layers = LayersTextBox.Text.Split(new string[] { ",", " ", "-", "_", "." }, StringSplitOptions.RemoveEmptyEntries);
List <int> layersVal = new List <int>();
foreach (var layer in layers)
{
layersVal.Add(int.Parse(layer));
}
YesNo bias = (YesNo)BiasCombobox.SelectedItem;
ActivationFunction activation = (ActivationFunction)ActivationCombobox.SelectedItem;
int iterations = int.Parse(Iterations.Text);
double learningRate = LearningRateSlider.Value;
double momentum = MomentumSlider.Value;
BackpropagationRunMode runMode = (BackpropagationRunMode)RunCombobox.SelectedItem;
ProblemType problem = (ProblemType)ProblemCombobox.SelectedItem;
CasesData trainingCases;
CasesData testCases;
YesNo appendTestFile = (YesNo)AppendTestCombobox.SelectedItem;
int input = testData.First().Count();
int output = trainingData.First().Count() - input;
CasesData.InitializeAllData(trainingData, testData, problem, out trainingCases, out testCases);
layersVal.Insert(0, input);
layersVal.Add((problem == ProblemType.Regression) ? output : trainingCases.ClassCount);
NeuralNetwork network = new NeuralNetwork(
(activation == ActivationFunction.Bipolar) ? (IActivation) new BipolarTanhActivation() : new UnipolarSigmoidActivation(),
bias == YesNo.Yes, layersVal.ToArray());
LearningResult learningResult = BackpropagationManager.Run(network, trainingCases, testCases,
runMode, iterations, learningRate, momentum);
if (appendTestFile == YesNo.Yes)
{
AppendCSVile(testDataPath, testCases);
}
ShowNetworkErrorWindow(learningResult);
if (problem == ProblemType.Regression)
{
ShowRegressionWindow(network, trainingCases, testCases);
}
else if (problem == ProblemType.Classification)
{
ShowClassificationWindow(network, trainingCases, testCases);
}
}
public void TestDirectionGuessed20PercentIdealToNetFalling()
{
List <Pattern> patterns = new List <Pattern>();
for (int i = 0; i < 101; i++)
{
patterns.Add(new Pattern(i));
patterns[i].IdealOutput = new DenseVector(new double[] { (double)-i });
patterns[i].NetworkAnswer = new DenseVector(new double[] { (double)-(i - 10) - (i % 5) * 2.26 });
}
double directionGuessed = BackpropagationManager.CalculateDirIdealToNet(patterns);
Assert.AreEqual(0.2, directionGuessed, 0.0001);
}
public void TestDirectionGuessedPerfectlyNetToNet()
{
List <Pattern> patterns = new List <Pattern>();
for (int i = 0; i < 20; i++)
{
patterns.Add(new Pattern(i));
patterns[i].IdealOutput = new DenseVector(new double[] { (double)i });
patterns[i].NetworkAnswer = new DenseVector(new double[] { (double)i + 20 }); //(i - 10) + (i % 5) * 100
}
double directionGuessed = BackpropagationManager.CalculateDirNetToNet(patterns);
Assert.AreEqual(1.0, directionGuessed, 0.0001);
}
public EngineResult Run()
{
foreach (LearningSettings learningSettings in eid.SettingsToRun)
{
resultsBySettings[learningSettings] = new List <SingleRunReport>();
for (int i = 0; i < eid.RunsPerSettings; i++)
{
runCounter++;
List <int> layersVal = BuildLayersVal(learningSettings);
BuildDataSet(layersVal);
INetwork network = CreateNetwork(learningSettings, layersVal);
NormalizeData(network, trainSet, testSet);
CheckIfPerformPCA(network);
learningStrategy = new VSetLearningStrategy(learningSettings);
BackpropagationManager backpropMan = new BackpropagationManager(network, trainSet, testSet);
var learningResult = backpropMan.Run(learningStrategy, learningSettings, this);
NormalizeDataBack(network, trainSet, testSet);
resultsBySettings[learningSettings].Add(new SingleRunReport(
network, DateTime.Now, learningResult, trainSet, testSet));
}
// use train error as criterion to simulate real-life ("we cannot use test set results yet - test set is the future")
resultsBySettings[learningSettings].RemoveHighestValues(
r => r.LearningResult.FinalTrainError, discardCount);
}
EngineResult result = new EngineResult();
result.ResultsBySettings = resultsBySettings;
result.Eid = eid;
result.WorstDiscardedCount = discardCount;
return(result);
}
