public NeuralNetwork(uint numInputs, uint numHiddenNodes, uint numOutputs, uint batches = 1)
{
NumInputs = numInputs;
NumHiddenNodes = numHiddenNodes;
NumOutputs = numOutputs;
m_ActivationFunction = new ActivationFunctions();
Layers = new Layer[3];
if (batches < 1)
{
batches = 1;
}
TotalBatches = batches;
ResetBatch();
InputLayer = Layers[0] = new Layer(numInputs, 0, m_ActivationFunction, batches);
HiddenLayer = Layers[1] = new Layer(numHiddenNodes, numInputs, m_ActivationFunction, batches);
OutputLayer = Layers[Layers.Length - 1] = new Layer(numOutputs, numHiddenNodes, m_ActivationFunction, batches, true);
ConfusionMatrix = new ConfusionMatrix(numOutputs);
Reset();
}
public void Start()
{
//Read the data from the files
var file1DataRaw = AlgorithmHelpers.ReadMatrixFromFile(@"class_1.dat");
var file2DataRaw = AlgorithmHelpers.ReadMatrixFromFile(@"class_2.dat");
ClassA = AlgorithmHelpers.ScaleDown(AlgorithmHelpers.ChooseFeatures(file1DataRaw));
ClassB = AlgorithmHelpers.ScaleDown(AlgorithmHelpers.ChooseFeatures(file2DataRaw));
m_maxX = ClassA.Max(0)[0] > ClassB.Max(0)[0] ? ClassA.Max(0)[0] : ClassB.Max(0)[0];
m_minX = ClassA.Min(0)[0] < ClassB.Min(0)[0] ? ClassA.Min(0)[0] : ClassB.Min(0)[0];
//Fill the charts with the data
m_view.ShowTrainingData(ClassA, ClassB);
//Clear the list view
m_view.ClearListView();
//Fill it with the confusion matrix for each algorithm per iteration
var statistics = new ConfusionMatrix[4, 5];
//Merge the two data sets
//and run kmeans
var kmeans = new KMeansClustering(AlgorithmHelpers.MergeArrays(file1DataRaw, file2DataRaw),
m_view.Iterations,
m_view.ThetaStep);
var idx = kmeans.Classify();
m_view.ClustersTextUpdate(idx.Distinct().Length.ToString());
m_view.ZeroProgressBar();
m_view.StepProgressBar();
//Partition m_iterations times and run the algorithms
for (int i = 0; i < m_iterations; ++i)
{
m_view.PerformStep();
//Partition its class to training and testing set
var partitions = new DataPartition[] { AlgorithmHelpers.Partition(ClassA), AlgorithmHelpers.Partition(ClassB) };
//Create the training data
var trainingPair = AlgorithmHelpers.CreateDataPair(partitions[0].Item1, partitions[1].Item1);
var trainingSet = trainingPair.Item1;
var trainingOutput = trainingPair.Item2;
//Create the testing data
var testingPair = AlgorithmHelpers.CreateDataPair(partitions[0].Item2, partitions[1].Item2);
var testingSet = testingPair.Item1;
var testingOutput = testingPair.Item2;
//Some functions need the training output to be a vector of doubles
var doubleTO = trainingOutput
.Select(x => new[] { Convert.ToDouble(x) })
.ToArray();
for (int k = 1; k < 3; ++k)
{
var nn = new KNearestNeighboursRuntime(k, trainingSet, trainingOutput, testingSet, testingOutput);
if (BestKNN == null)
{
BestKNN = nn.Execute();
}
else
{
var iter = nn.Execute();
if (iter.Accuracy > BestKNN.Accuracy)
{
BestKNN = iter;
}
}
}
var perceptron = new PerceptronRuntime(trainingSet, doubleTO, testingSet, testingOutput);
perceptron.Finished += perceptron_Finished;
var leastSquare = new LeastSquaresRuntime(AlgorithmHelpers.JaggedToMD(trainingSet), AlgorithmHelpers.JaggedToMD(doubleTO), AlgorithmHelpers.JaggedToMD(testingSet), testingOutput);
var neuralNetwork = new ParallelNeuralNetworkRuntime(trainingSet, doubleTO, testingSet, testingOutput);
neuralNetwork.Finished += neuralNetwork_Finished;
//Compute the confusion matrices for the four classifiers
statistics[0, i] = perceptron.Execute();
statistics[1, i] = leastSquare.Execute();
//Use the most accurate K of KNN
statistics[2, i] = BestKNN;
statistics[3, i] = neuralNetwork.Execute();
}
//Update the classifier lines in the charts
//with the most accurate of the 5 iterations
m_view.ChartUpdate("", "classifier", MostAccuratePerceptron.Item1);
m_view.ChartUpdate("", "classifier1", MostAccurateNN.Item1);
//Process the array with the Confusion Matrices
//and update the list view
var processed = AlgorithmHelpers.ProcessStatistics(statistics);
m_view.UpdateStatisticsListView(processed);
}
