public void BackPropagate(Matrix y, float learningRate)
{
Console.WriteLine("\n * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * \n");
Console.WriteLine($" B A C K P R O P A P A G A T I O N");
Console.WriteLine("\n * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * \n");
Console.WriteLine();
// debug
var c = NeurNetMath.C(A[net.L - 1], y, SquaredMeanError.CostFunction)
.DumpToConsole($"\n{CostType.SquaredMeanError} C =");
var cTotal = NeurNetMath.CTotal(A[net.L - 1], y, SquaredMeanError.CostFunction);
Console.WriteLine($"\nCTotal = {cTotal}\n");
// Iterate backwards over each layer (skip input layer).
for (int l = net.L - 1; l > 0; l--)
{
dadz_OfLayer[l] = NeurNetMath.dadz(A[l], Z[l], net.Derivations[l])
.DumpToConsole($"\ndadz{l} =");
Matrix error;
if (l == net.L - 1)
{
// .. and C0 instead of a[i] and t as parameters here?
error = NeurNetMath.delta_Output(cTotal, Z[l], SquaredMeanError.DerivationOfCostFunction);
}
else
{
error = NeurNetMath.delta_Hidden(Delta[l + 1], net.W[l + 1], A[l], Z[l], net.Derivations[l]);
}
Delta[l] = error.DumpToConsole($"\ndelta{l} =");
}
// Adjust weights and biases (skip input layer).
for (int l = 1; l < net.L; l++)
{
// E or delta ?
net.W[l] = GetCorrectedWeights(net.W[l], A[l - 1], Delta[l], learningRate)
.DumpToConsole($"\nadjusted w{l} =");
net.B[l] = GetCorrectedWeights(net.B[l], A[l - 1], Delta[l], learningRate)
.DumpToConsole($"\nadjusted b{l} =");
}
}
} // redundant?
#endregion
#region methods
public Matrix FeedForwardAndGetOutput(Matrix input)
{
Console.WriteLine("\n * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * \n");
Console.WriteLine($" F E E D F O R W A R D");
Console.WriteLine("\n * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * \n");
Console.WriteLine();
// wa: Separate inp layer from 'layers' ?!
A[0] = input.DumpToConsole($"\nA[0] = "); //new Matrix(input.ToArray());
Console.WriteLine("\n - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ");
// iterate over layers (skip input layer)
for (int i = 1; i < net.L; i++)
{
Z[i] = NeurNetMath.z(net.W[i].DumpToConsole($"\nW{i} = "), A[i - 1].DumpToConsole($"\nA{i-1} = "), net.B[i].DumpToConsole($"\nB{i} = ")).DumpToConsole($"\nZ{i} = ");
A[i] = NeurNetMath.a(Z[i], net.Activations[i]).DumpToConsole($"\nA{i} = ");
Console.WriteLine("\n - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ");
}
return(A.Last());
}
