public MLPNN(TrainingDataMethodsAndParameters <T> trainingDataMethodsAndParameters,
NetWorkMethodsAndParameters <T, U> netWorkMethodsAndParameters, params int[] parameters) :
base(netWorkMethodsAndParameters)
{
SampleSize = trainingDataMethodsAndParameters.sampleSize; // ini of sample size for gradien decent
BatchSize = trainingDataMethodsAndParameters.batchSize; // ini the total batch size for gradient decent
ValidSize = trainingDataMethodsAndParameters.validSize; // ini the verification batch size for gradient decent
ComputeGradient = trainingDataMethodsAndParameters.computeGradient; // comptutes gradient and the cost function
GetInputOutput = trainingDataMethodsAndParameters.getInputOutput; // sets input output
ActivationFunctions = trainingDataMethodsAndParameters.activationFunctions; //sets Activation function for hidden layer
ActivationFunctionsOutputLayer = trainingDataMethodsAndParameters.activationFunctionsOutputLayer; //sets Activation function for hidden layer
GenerateEmptyMatrix = trainingDataMethodsAndParameters.generateEmptyMatrix; // generates the Matrix
epochNumber = trainingDataMethodsAndParameters.epochNumber; //pass epochnumber
learningRate = trainingDataMethodsAndParameters.learningRate; // setting learning rate
lTwoRegularization = trainingDataMethodsAndParameters.lTwoRegularization; // setting l2 regularization
GetVerfication = trainingDataMethodsAndParameters.getVerfication; // verification function
NNInputOutput = trainingDataMethodsAndParameters.nNInputOutput; // conveting float output into classification
ValidationFunction = trainingDataMethodsAndParameters.validationFunction; // how we compare the true output with the computed output
weightMatrix = new List <T>(NetworkSize.Length - 1); // initilaex the weigts
biasMatrix = new List <U>(NetworkSize.Length - 1); //// initilaex the bias
biasTempMatrix = new List <T>(NetworkSize.Length - 1); // initiate bias layers Matrix
errorValues = new List <T>(NetworkSize.Length - 1); //// inti errors
neuronOutputValues = new List <T>(NetworkSize.Length - 1); // ini outputes
neuronInputValues = new List <T>(NetworkSize.Length - 1); //ini inputs
this.parameters = parameters; // optiontl parameters// Now not sure what is going to be here
SetTheNetwrok(); // intilaizng
}
// call cosntructor
protected NNDataBase(NetWorkMethodsAndParameters <T, U> netWorkMethodsAndParameters) // make protected finally
{
GenerateWeights = netWorkMethodsAndParameters.generateWeights; // ini method for generating weights
GenerateBias = netWorkMethodsAndParameters.generateBias; // ini method for generating bias
MatMatMul = netWorkMethodsAndParameters.matMatMul; // ini method for Matrix x Matrix
MatMatMulTranspose = netWorkMethodsAndParameters.matMatMulTranspose; // ini method for Matrix Transpose x Matrix
MatMatMulSecondTranspose = netWorkMethodsAndParameters.matMatMulSecondTranspose; // ini method for Matrix x Matrix Transpose
MatMatMulPointWise = netWorkMethodsAndParameters.matMatMulPointWise; // ini method for Matrix x Matrix PointWise
MatVecMul = netWorkMethodsAndParameters.matVecMul; // ini method for Matrix x Vector
MatMatSum = netWorkMethodsAndParameters.matMatSum; // ini method for Matrix + Matrix
MatScalarMult = netWorkMethodsAndParameters.matScalarMult; // ini method mat x scalar
VectScalarMult = netWorkMethodsAndParameters.vectScalarMult; // ini method vec x scalar
PointWiseVecVecSum = netWorkMethodsAndParameters.pointWiseVecVecSum; // ini method for Shur product
VecToMat = netWorkMethodsAndParameters.vecToMat; // ini method for construction of Matrix with Vector
NetworkSize = netWorkMethodsAndParameters.networkSize; // number of layers, neurons
GenerateUnitVector = netWorkMethodsAndParameters.GenerateUnitVector; // generate unit vector
TransposeMatrix = netWorkMethodsAndParameters.TransposeMatrix; // matrix Transpose
}
public static void SetUpNNParameters(out NetWorkMethodsAndParameters <Matrix <float>, Vector <float> > parameters)
{
parameters = new NetWorkMethodsAndParameters <Matrix <float>, Vector <float> >();// creaty empty;
}
public static void SetUpNNParameters(out NetWorkMethodsAndParameters <float [][], float []> parameters)
{
parameters = null;
}
static void Main(string[] args)
{
#region Setting up binary data reader
// ImageFileReader imageFileReader = new ImageFileReader(3 * 1024, 10); // COLOR IMAGES
ImageFileReader imageFileReader = new ImageFileReader(28 * 28, 10, 16, 8); // MNIST DIGITS
#endregion
#region Initializng training Data
int sampleSize = 50; // this is for test 10; // the Size of MiniBatch ( so called Sample)
int inputRows = imageFileReader.Size; //10; // matrix sizes
int numberOfHiddenLayers = 1; // 1; // number of hidden layers
int[] hiddenLayers = new int[numberOfHiddenLayers];
for (int i = 0; i < numberOfHiddenLayers; i++)
{
hiddenLayers[i] = inputRows / (int)Math.Pow(2, i + 1) / 3; // defining number of hidden layers
}
int outputRows = imageFileReader.SizeOfClassification; // 1; // outputSizes // imageFileReader.SizeOfClassification
int parameterToPass = 1; // if positive, random Normal generated weights and matrices, or unifrom otherwise
int parameterSampleGenerator = 0; // if positive, sample is fixed// gradient decent
int epochNumber = 20; // 300; // number of attempts
double learningRate = 2; // learning rate
int[] NetWorkSize = new int[2 + numberOfHiddenLayers]; // { rows, hiddenLayers outputRows }; // inti networkSize
NetWorkSize[0] = inputRows; // ini input layer
NetWorkSize[1 + numberOfHiddenLayers] = outputRows; // ini output layer
for (int i = 1; i < 1 + numberOfHiddenLayers; i++)
{
NetWorkSize[i] = hiddenLayers[i - 1]; // ini of NN architecture
}
NetWorkSize[1] = 30; //manual, comment after all
#endregion
#region SetUpTheNetWorkMethods
NetWorkMethodsAndParameters <Matrix <float>, Vector <float> > parametersForNN = //iniParamtersMethod to Pass to a NN
new NetWorkMethodsAndParameters <Matrix <float>, Vector <float> >(); // creaty empty
TrainingDataMethodsAndParameters <Matrix <float> > methodsForNN = //iniParamtersMethods to Pass to a Training
new TrainingDataMethodsAndParameters <Matrix <float> >(); // create empty
// ini for NN
parametersForNN.generateWeights = RandomGenerators.GenerateRandomMatrix; // pass weights generators.
parametersForNN.generateBias = RandomGenerators.GenerateRandomVector; // pass bias generators.
parametersForNN.vecToMat = MathLibrary.VectorToMatrix; // vectorToMatrix for biases
parametersForNN.matMatSum = MathLibrary.MatMatSum; // sum matrix and matrix
parametersForNN.matMatMul = MathLibrary.MatMatMul; // product matrix and matrix
parametersForNN.matMatMulTranspose = MathLibrary.MatMatMulTranspose; // product matrixT and matrix
parametersForNN.matMatMulSecondTranspose = MathLibrary.MatMatMulSecondTranspose; // product matrix and matrixT
parametersForNN.matMatMulPointWise = MathLibrary.MatMatMulPointWise; // product matrix and matrix, PointWise
parametersForNN.matScalarMult = MathLibrary.MatScalarMul; // matrix x scalar product delegate
parametersForNN.matVecMul = MathLibrary.MatVecMul; // matrix x vector;
parametersForNN.vectScalarMult = MathLibrary.VecScalarMul; // matrix x scalar product delegate
parametersForNN.pointWiseVecVecSum = MathLibrary.VectorVectorSum; // vect + vect;
parametersForNN.GenerateUnitVector = MathLibrary.GenerateUnitVector; // generate unit vector
parametersForNN.networkSize = NetWorkSize; // passint networkSize
parametersForNN.TransposeMatrix = MathLibrary.TranposeMatrix; // Matrix Transpose
// ini methods for propagation and feeding
methodsForNN.sampleSize = sampleSize; // setting sampleSize
methodsForNN.validSize = imageFileReader.ValidSize; // setting sampleSize
methodsForNN.batchSize = imageFileReader.BatchSize; // BatchSize for now disabled
methodsForNN.epochNumber = epochNumber; // number of grad steps
methodsForNN.learningRate = learningRate; // learning rate
methodsForNN.getVerfication = imageFileReader.VerInOutSample; //computing verification
methodsForNN.validationFunction = MathLibrary.ValidatePrediction; // validator
methodsForNN.nNInputOutput = MathLibrary.GenerateMatrix; // generate input rule
methodsForNN.generateEmptyMatrix = MathLibrary.GenerateMatrix; // null matrix generator
methodsForNN.getInputOutput = imageFileReader.InputSample; // TrainingDataSet<Matrix<float>>.SetTrainingDataBenchmark; //setting InputOutput// imageFileReader.InputSample(sampleSize, null);
methodsForNN.computeGradient = CostFunction.GradientQuadratic; // CostFunction.GradientEntropy;// passing a gradient function // fix this issue tommorow
ActivationFunctions.ReturnActivationFunction(out methodsForNN.activationFunctions, 0, "Sigmoid", "Sigmoid"); // fixed // passing activation functions // hidden layers
ActivationFunctions.ReturnActivationFunction(out methodsForNN.activationFunctionsOutputLayer, 1, "Sigmoid", "Sigmoid"); // fixed // passing activation functions // output layers
MLPNN <Matrix <float>, Vector <float> > myFullNN = new MLPNN <Matrix <float>, Vector <float> >(methodsForNN, parametersForNN, parameterToPass, parameterSampleGenerator); // iniFullNetwork to Test
#endregion
#region Optimization and Output
myFullNN.Optimization(); // running optimization procedure // printes cost function
Console.WriteLine($"Accuracy is {myFullNN.Accuracy(10000,imageFileReader.SizeOfVerification)} %"); // ok it is working or it seems so.
#endregion
//// print Input
//PrintMatrix(myFullNN.ShowInputLayer(), inputRows, sampleSize, "\nPrint InputLayer");
//// print ComputedOutput
//PrintMatrix(myFullNN.ShowComputedOutputLayer(), outputRows, sampleSize, "\nPrint ComputedOutputLayer", 0);
//// print TrueOutput
//PrintMatrix(myFullNN.ShowOutputLayer(), outputRows, sampleSize, "\nPrint TrueOutputLayer");
//// print RefinedComputedOutput
//PrintMatrix(MathLibrary.GenerateMatrix(myFullNN.ShowComputedOutputLayer()), outputRows, sampleSize, "\nPrint ComputedOutputLayerModifiedByfunction");
//Console.WriteLine($"The Number of CorrectAnswers is" +
// $" {(int)MathLibrary.ValidatePrediction(myFullNN.ShowOutputLayer(), MathLibrary.GenerateMatrix(myFullNN.ShowComputedOutputLayer()))}");
//Console.WriteLine("\nPress Enter to Finish");
//Console.ReadKey();
}
