private FlattenedImage[] GetCrossentropyError(double[] output, double[] expected, char c)
{
FlattenedImage[] result = new FlattenedImage[output.Length];
double[] expectedExp = MatrixUtils.GetExp(expected);
double expectedSum = expected.Sum();
//int cInt = (int)(c - 65);
//for (int i = 0; i < output.Length; i++)
//{
// if (i == cInt)
// {
// double log = -Math.Log(output[i]);
// result[i] = new FlattenedImage(1, new double[1] { log });
// }
// else
// {
// result[i] = new FlattenedImage(1, new double[1] { 0 });
// }
//}
for (int i = 0; i < output.Length; i++)
{
result[i] = new FlattenedImage(1, new double[1] {
expectedExp[i] / expectedSum - output[i]
});
}
return(result);
}
public override double Sum()
{
double result = 0;
for (int c = 0; c < NumberOfChannels; c++)
{
result += MatrixUtils.ElementSum(Channels[c].Values);
}
return(result);
}
public double[,] Convolve(FilteredImageChannel input, bool samePadding)
{
//double[,] flippedKernel = MatrixUtils.Rotate180(Values);
//int fullSize = Size * Size;
if (samePadding)
{
return(MatrixUtils.ConvolveSame(input.Values, Values));
}
return(MatrixUtils.Convolve(input.Values, Values));
//for (int i = 0; i < Size; i++)
//{
// for (int j = 0; j < Size; j++)
// {
// convResult[i, j] /= fullSize;
// }
//}
}
public FilteredImageChannel Backpropagate(FilteredImageChannel previous, FilteredImageChannel nextErrors, int totalKernels, double learningRate, bool samePadding)
{
//double[,] kernelDerivatives = new double[1,1];
//Task t = Task.Run(() =>
//{
// if (samePadding)
// {
// kernelDerivatives = MatrixUtils.ConvolveSame(previous.Values, nextErrors.Values);
// }
// else
// {
// kernelDerivatives = MatrixUtils.Convolve(previous.Values, nextErrors.Values);
// }
//});
//double[,] flippedKernel = MatrixUtils.Rotate180(Values);
////double[,] newErrors = MatrixUtils.Convolve(flippedKernel, nextErrors.Values);
//double[,] newErrors;
//if (samePadding)
//{
// newErrors = MatrixUtils.FullConvolutionSame(flippedKernel, nextErrors.Values);
//}
//else
//{
// newErrors = MatrixUtils.FullConvolution(flippedKernel, nextErrors.Values);
//}
//t.Wait();
//for (int i = 0; i < Size; i++)
//{
// for(int j = 0; j < Size; j++)
// {
// Values[i, j] -= learningRate * kernelDerivatives[i, j];
// }
//}
//ComputeElementSum();
//return new FilteredImageChannel(Size, newErrors);
//double[,] flippedErrors = MatrixUtils.Rotate180(nextErrors.Values);
double[,] deltaWeights = new double[1, 1];
Task t1 = Task.Run(() =>
{
if (samePadding)
{
deltaWeights = MatrixUtils.ConvolveSame(previous.Values, nextErrors.Values);
}
else
{
deltaWeights = MatrixUtils.Convolve(previous.Values, nextErrors.Values);
}
});
//double[,] flippedKernel = MatrixUtils.Rotate180(Values);
double[,] newErrors;
if (samePadding)
{
newErrors = MatrixUtils.Convolve(Values, nextErrors.Values);
}
else
{
newErrors = MatrixUtils.FullConvolution(Values, nextErrors.Values);
}
t1.Wait();
double[,] newWeights = new double[Size, Size];
for (int i = 0; i < Size; i++)
{
for (int j = 0; j < Size; j++)
{
newWeights[i, j] = Values[i, j] - learningRate * deltaWeights[i, j];
}
}
//Values = MatrixUtils.Rotate180(newWeights);
Values = newWeights;
return(new FilteredImageChannel(newErrors.GetLength(0), newErrors));
}
