private static TrainingState ClassifyData(NeuralNetwork neuralNetwork, UserInput userInput, long epoch)
{
// calculate the output for each training example
foreach (DataInstance trainingExample in userInput.DataInstances)
{
neuralNetwork.InputNodes.ForEach(input => input.CalculateValue(trainingExample));
neuralNetwork.HiddenNodes.ForEach(hidden => hidden.CalculateValue(neuralNetwork.Weights));
neuralNetwork.OutputNodes.ForEach(output => output.CalculateValue(neuralNetwork.Weights));
// associated all the calculated outputs (1 for our current case) with the training example
trainingExample.CalculatedClasses.Clear();
neuralNetwork.OutputNodes.ForEach(output => trainingExample.CalculatedClasses.Add(output.Value));
}
// calculate RMSE for all training examples
List<double> RMSEs = userInput.DataInstances.ConvertAll(trainingExample => trainingExample.RMSE);
// return the results
return new TrainingState()
{
Epoch = epoch,
MaxRMSE = RMSEs.Max(),
AvgRMSE = RMSEs.Average(),
PercentCorrect = (float)RMSEs.Where(rmse => rmse < userInput.ErrorMargin).Count() / (float)RMSEs.Count
};
}
public void Fixed_200Epochs()
{
UserInput userInput = new UserInput()
{
DataFilePath = ".\\TestData\\iris.csv",
ErrorMargin = 0.05,
HiddenNodeCount = 10,
LearningRate = 0.1,
MaxEpochs = 200,
MaxInitialWeight = 0.02,
MinInitialWeight = 0.02
};
NeuralNetwork net = new NeuralNetwork(userInput);
Backpropagation.Calculate(net, userInput);
Assert.AreEqual(net.TrainingState.Epoch, 200);
Assert.LessOrEqual(Math.Abs(net.TrainingState.MaxRMSE - 0.3539866109402714), 0.0000001);
Assert.LessOrEqual(Math.Abs(net.TrainingState.AvgRMSE - 0.08665010792931953), 0.0000001);
Assert.LessOrEqual(Math.Abs(net.TrainingState.PercentCorrect - .43), 0.05);
}
public void Fixed_10Epochs()
{
UserInput userInput = new UserInput()
{
DataFilePath = ".\\TestData\\iris.csv",
ErrorMargin = 0.05,
HiddenNodeCount = 10,
LearningRate = 0.1,
MaxEpochs = 10,
MaxInitialWeight = 0.02,
MinInitialWeight = 0.02
};
NeuralNetwork net = new NeuralNetwork(userInput);
Backpropagation.Calculate(net, userInput);
Assert.AreEqual(net.TrainingState.Epoch, 10);
Assert.LessOrEqual(Math.Abs(net.TrainingState.MaxRMSE - 0.7647903291625834), 0.0000001);
Assert.LessOrEqual(Math.Abs(net.TrainingState.AvgRMSE - 0.4198381630410543), 0.0000001);
Assert.LessOrEqual(Math.Abs(net.TrainingState.PercentCorrect - 0), 0.05);
}
public static void Calculate(NeuralNetwork neuralNetwork, UserInput userInput)
{
LogHelper.WriteDebug("Backpropagation.Calculate - Enter");
long epoch = 0;
// do until the termination condition is met
while (++epoch <= userInput.MaxEpochs && neuralNetwork.TrainingState.MaxRMSE >= userInput.ErrorMargin)
{
LogHelper.WriteDebug("Starting epoch {0}", epoch);
// iterate through the training examples
foreach (DataInstance trainingExample in userInput.DataInstances)
{
LogHelper.WriteDebug("Processing training example: {0}", trainingExample);
// 1.
LogHelper.WriteDebug("Propagating input forward through network.");
neuralNetwork.InputNodes.ForEach(input => input.CalculateValue(trainingExample));
neuralNetwork.HiddenNodes.ForEach(hidden => hidden.CalculateValue(neuralNetwork.Weights));
neuralNetwork.OutputNodes.ForEach(output => output.CalculateValue(neuralNetwork.Weights));
// 2. propagate error backward through the network
LogHelper.WriteDebug("Calculating output unit error values.");
neuralNetwork.OutputNodes.ForEach(output => output.CalculateError(trainingExample));
LogHelper.WriteDebug("Calculating hidden unit error values.");
neuralNetwork.HiddenNodes.ForEach(hidden => hidden.CalculateError(neuralNetwork.Weights));
LogHelper.WriteDebug("Updating network weights.");
neuralNetwork.Weights.ForEach(weight => weight.CalculateValue(userInput.LearningRate));
}
// classify the data based on the updated weights
LogHelper.WriteDebug("Running classification with updated weights.");
neuralNetwork.TrainingState = ClassifyData(neuralNetwork, userInput, epoch);
}
LogHelper.WriteDebug("Backpropagation.Calculate - Exit");
}
public void Fixed()
{
UserInput userInput = new UserInput()
{
DataFilePath = ".\\TestData\\xor.csv",
ErrorMargin = 0.1,
HiddenNodeCount = 2,
LearningRate = 0.1,
MaxEpochs = 10,
MaxInitialWeight = 0.6,
MinInitialWeight = 0.6
};
NeuralNetwork net = new NeuralNetwork(userInput);
Backpropagation.Calculate(net, userInput);
Assert.AreEqual(net.TrainingState.Epoch, 10);
Assert.LessOrEqual(Math.Abs(net.TrainingState.MaxRMSE - 0.5432215742673802), 0.0000001);
Assert.LessOrEqual(Math.Abs(net.TrainingState.AvgRMSE - 0.4999911891911738), 0.0000001);
Assert.LessOrEqual(Math.Abs(net.TrainingState.PercentCorrect - 0), 0.05);
}
private void trainButton_Click(object sender, EventArgs e)
{
UpdateUserInputFromControls();
InitializeResultsFile();
output.Clear();
NeuralNetwork net = new NeuralNetwork(_userInput);
net.TrainingStateChanged += new NeuralNetwork.TrainingStateChangedEventHandler(net_TrainingStateChanged);
Backpropagation.Calculate(net, _userInput);
}
