//Backward Propagation.
public Matrix Backwards(Matrix yin, Boolean end = false, Boolean learn = true)
{
Matrix y = yin;
#if ADD_BIAS
if (!end)
{
y = Matrix.RemoveBias(yin);
}
#endif
if (y.Rows != batchsize)
{
throw new Exception("Batch-size is inconsistant.");
}
if (y.Columns != nodes.Columns)
{
Debug.Log("Y: " + y.Columns + " nodes: " + nodes.Columns);
throw new Exception("Inconsistant number of inputs for backward.");
}
Matrix err;
Matrix delta;
//This only works on the last layer. >:
if (end)
{
err = y - nodes; //Only supports "last"
}
else
{
err = y; //Takes in error from prior result.
}
delta = Matrix.HadamardProd(err, Matrix.Transform(func.deriv, nodes)); //Adjust for the slope
if (learn)
{
weights = weights + (Matrix.Transpose(x) * delta); //Apply the values.
}
return(delta * Matrix.Transpose(weights));
}
private void BackwardPropagation()
{
int last = hidLayer.results.Count - 1;
errorOutput = targetOutput - outputResult;
//do initial fix for out put to last hidden layer
Matrix deltaLayer = Matrix.HadamardProd(Matrix.Transform(SigmoidPrime, outputSum), errorOutput);
Matrix layerChanges = Matrix.Transpose(hidLayer.results[last]) * deltaLayer * learningRate;
Matrix previousWieght = hiddenOutputWeights;
hiddenOutputWeights += layerChanges;
for (int i = 0; i < last; i++)
{
deltaLayer = Matrix.HadamardProd(deltaLayer * Matrix.Transpose(previousWieght), Matrix.Transform(SigmoidPrime, hidLayer.sums[last - i]));
layerChanges = Matrix.Transpose(hidLayer.results[last - i]) * deltaLayer * learningRate;
previousWieght = hidLayer.wieghts[last - 1 - i];
hidLayer.wieghts[last - 1 - i] += layerChanges;
}
deltaLayer = Matrix.HadamardProd(deltaLayer * Matrix.Transpose(previousWieght), Matrix.Transform(SigmoidPrime, hidLayer.sums[0]));
layerChanges = Matrix.Transpose(inputLayer) * deltaLayer * learningRate;
inputHiddenWeights += layerChanges;
}
