public ConvolutionalLayer(int inputRows, int inputCols, int kernelRows, int kernelCols)
{
var hiddenSize = (
(inputRows - kernelRows) *
(inputCols - kernelCols)
);
int kernelSize = kernelRows * kernelCols;
Kernel = Enumerable.Repeat(0, kernelSize).Select(i => NNFunctions.GetRandom()).ToArray();
}
ForwardPropogate(Matrix inputs)
{
InputValues = inputs;
int iStart = KernelSide / 2;
int iEnd = inputs.NumCols - iStart;
int jEnd = inputs.NumRows - iStart;
Matrix outputs = new Matrix(jEnd - iStart, iEnd - iStart);
int kIStart = 0;
int kJStart = 0;
int kIEnd = KernelSide;
int kJEnd = KernelSide;
int outputsI = 0;
int outputsJ = 0;
for (int inputsI = iStart; inputsI < iEnd; ++inputsI)
{
for (int inputsJ = iStart; inputsJ < jEnd; ++inputsJ)
{
// iterating through the section of the image,
// applying the kernel to it
int kernelIndex = 0;
double total = 0;
for (int kI = kIStart; kI < kIEnd; ++kI)
{
for (int kJ = kJStart; kJ < kJEnd; ++kJ)
{
// multiplying the pixel by the kernel weight
if (NNFunctions.GetRandom(false) > Parameters.DropoutRate)
{
total += Kernel[kernelIndex] * inputs[kI, kJ];
}
++kernelIndex;
}
}
outputs[outputsI, outputsJ] = Math.Tanh(total * Parameters.DropoutMultiplier);
++outputsJ;
++kJStart;
++kJEnd;
}
++outputsI;
++kIStart;
++kIEnd;
outputsJ = 0;
kJStart = 0;
kJEnd = KernelSide;
}
return(null);
}
