private void btnLoadMNIST_Click(object sender, EventArgs e)
{
try {
/*
* Use the DataPoint[]
* Get double[][] from DataPoint[]
* array of array, should be easily converted to double[][]
*/
//String trainDir = "..\\..\\..\\..\\..\\..\\handouts\\data\\trainingAll60000";
String trainDir = "..\\..\\..\\..\\..\\..\\handouts\\data\\train";
//String testDir = "..\\..\\..\\..\\..\\..\\handouts\\data\\testAll10000";
Stopwatch sw = new Stopwatch();
sw.Start();
DataPoint[] data = ImageReader.ReadAllDataScaled(trainDir);
double[][] trainData = ImageReader.GetData(data);
sw.Stop();
MessageBox.Show("Time taken to read the trainer data " + sw.ElapsedMilliseconds.ToString());
int[] layers = { 100, trainData[0].Count() }; // neurons in hidden layer, ouput layer
sparse_encoder = new Network(trainData[0].Count(), layers); // # of inputs
sparse_encoder.randomizeAll();
sparse_encoder.LearningAlg.ErrorTreshold = 0.0001f;
sparse_encoder.LearningAlg.MaxIteration = 10000;
//sparse_encoder = Network.load("sparse_encoder");
sw.Restart();
sparse_encoder.LearningAlg.Learn(trainData, trainData);
sw.Stop();
MessageBox.Show("Done training...Time taken " + sw.ElapsedMilliseconds.ToString());
/* Save the auto encoder learn */
sparse_encoder.save("sparse_encoder");
double[][] expectedOutputs = ImageReader.ExpectedOutput(data);
int[] nnLayers = { 100, expectedOutputs[0].Length }; // neurons in hidden layer, ouput layer
nn = new Network(trainData[0].Count(), nnLayers); // # of inputs
/* No need to randmize, get weight and baise from sparse_encoder */
for (int i = 0; i < sparse_encoder.layers[0].NumNeurons; i++)
{
nn.layers[0].Neurons[i].weights = sparse_encoder.layers[0].Neurons[i].weights;
nn.layers[0].Neurons[i].Bias = sparse_encoder.layers[0].Neurons[i].Bias;
}
//nn.randomizeAll();
nn.LearningAlg.ErrorTreshold = 0.0001f;
nn.LearningAlg.MaxIteration = 10000;
sw.Restart();
nn.LearningAlg.Learn(trainData, expectedOutputs);
sw.Stop();
MessageBox.Show("Done training...Time taken " + sw.ElapsedMilliseconds.ToString());
nn.save("nn_ae");
}
catch (Exception ex)
{
MessageBox.Show(ex.Message);
}
}
private void btnLoadPCA_Click(object sender, EventArgs e)
{
try {
/*
* STEPS
* 1- Convert image to grayscale
* 2- Convert to 2-D image i.e. conversion to vector
*/
//String trainDir = "..\\..\\..\\..\\..\\..\\handouts\\data\\trainingAll60000";
String trainDir = "..\\..\\..\\..\\..\\..\\handouts\\data\\train";
//String testDir = "..\\..\\..\\..\\..\\..\\handouts\\data\\testAll10000";
//string trainDir = "..\\..\\..\\..\\..\\..\\handouts\\AttDataSet\\ATTDataSet\\Training";
Stopwatch sw = new Stopwatch();
sw.Start();
DataPoint[] data = ImageReader.ReadAllDataUnscaled(trainDir);
//double[][] trainData = ImageReader.ReadAllData(trainDir);
double[][] trainDataOrig = ImageReader.GetData(data);
double[][] trainData = PCA.Transpose(trainDataOrig, trainDataOrig[0].Length);
sw.Stop();
MessageBox.Show("Time taken to read the trainer data " + sw.ElapsedMilliseconds.ToString());
/*
* STEPS:
* 3- Compute the mean vector of all test images
* 4- Subtract mean vector from each image
* 5- Compute covariant matrix of all test images
* pass the vector through the PCA
* then pass that data through NN
*/
sw.Restart();
iMean = PCA.FindMean(trainData);
PCA.SubMean(trainData, iMean);
double[][] covariance = PCA.Covariance(trainData);
/* Compute the eigan values (values are sorted) */
PCALib.Matrix mapackMatrix = new PCALib.Matrix(covariance);
PCALib.IEigenvalueDecomposition EigenVal = mapackMatrix.GetEigenvalueDecomposition();
/* select the top 50 Eigen values */
int top = 50;
#if DEBUG
/*
* we don't need the eigen values
* because the eigen vactors are already
* calculated by mapack library
*/
double[] topEigen = new double[top];
PCA.GetTopN(EigenVal.RealEigenvalues, topEigen, top);
#endif // DEBUG
/* get Eigen vector */
double[][] EigenVector = PCA.GetEigenVector(EigenVal.EigenvectorMatrix, top);
/* multiply eigen vector with vector that has mean substracted */
EigenFaceImage = PCA.Multiply(trainData, EigenVector);
/* Project each image on to reduced top dimensional space */
double[][] transposeInput = PCA.Transpose(EigenFaceImage, EigenFaceImage[0].Length);
double[][] transposeTrainData = PCA.Transpose(trainData, trainData[0].Length);
projectionInput = PCA.Multiply(transposeTrainData, EigenFaceImage);
sw.Stop();
MessageBox.Show("Done PCA...Time taken " + sw.ElapsedMilliseconds.ToString()); // 256094 normal vs 211514 parallel
double[][] image = PCA.ConvertToPixels(transposeInput);
int iNo = 0;
foreach (Control obj in groupbox1.Controls)
{
if (obj is PictureBox)
{
obj.BackgroundImage = PCA.Draw(image, iNo++, trainDir);
}
}
double[][] expectedOutputs = ImageReader.ExpectedOutput(data);
int[] layers = { 50, 10 }; // neurons in hidden layer, ouput layer
nn = new Network(projectionInput[0].Length, layers); // # of inputs
nn.randomizeAll();
nn.LearningAlg.ErrorTreshold = 0.0001f;
nn.LearningAlg.MaxIteration = 10000;
sw.Restart();
nn.LearningAlg.Learn(projectionInput, expectedOutputs);
sw.Stop();
MessageBox.Show("Done training...Time taken " + sw.ElapsedMilliseconds.ToString());
nn.save("nn_pca");
}
catch (Exception ex)
{
MessageBox.Show(ex.Message);
}
}
