private static void DoBackpropagation(Layer outputLayer, double[] targetOutputs, ErrorFunctionType errorFunctionType, double learningRate, double momentumMagnitude)
{
var backwardsPassDeltas = UpdateOutputLayer(outputLayer, targetOutputs, errorFunctionType, learningRate, momentumMagnitude);
foreach (var t in outputLayer.PreviousLayers)
{
RecurseBackpropagation(t, backwardsPassDeltas, momentumMagnitude);
}
}
private static Dictionary <Node, double> UpdateOutputLayer(Layer outputLayer, double[] targetOutputs, ErrorFunctionType errorFunctionType, double learningRate, double momentumMagnitude)
{
var errorFunctionDifferential = ErrorFunctionResolver.ResolveErrorFunctionDifferential(errorFunctionType);
var deltas = new Dictionary <Node, double>();
for (var i = 0; i < outputLayer.Nodes.Count; i++)
{
var node = outputLayer.Nodes[i];
var delta = Math.MinMagnitude(errorFunctionDifferential.Invoke(targetOutputs[i], node.Output), 100)
* outputLayer.ActivationFunctionDifferential(node.Output)
* learningRate;
deltas.Add(node, delta);
foreach (var(prevNode, weightForPrevNode) in node.Weights)
{
UpdateNodeWeight(prevNode, weightForPrevNode, delta, momentumMagnitude);
}
foreach (var(_, weightForPrevLayer) in node.BiasWeights)
{
UpdateBiasNodeWeight(weightForPrevLayer, delta, momentumMagnitude);
}
}
return(deltas);
}
public static void Backpropagate(this Layer outputLayer, Dictionary <Layer, double[]> inputs, double[] targetOutputs, ErrorFunctionType errorFunctionType, double learningRate, double momentumMagnitude = 0d)
{
outputLayer.CalculateOutputs(inputs);
DoBackpropagation(outputLayer, targetOutputs, errorFunctionType, learningRate, momentumMagnitude);
}
/// <summary>
/// Returns the differential of the error function supplied
/// Function returned is of the following signature: (target, actual) => differential of error
/// </summary>
/// <param name="errorFunctionType"></param>
/// <returns></returns>
public static Func <double, double, double> ResolveErrorFunctionDifferential(ErrorFunctionType errorFunctionType) => errorFunctionType switch
{
public static void NegativeSample(this Layer outputLayer, int inputIndex, int outputIndex, bool isPositiveTarget, ErrorFunctionType errorFunctionType, double learningRate, double momentumMagnitude = 0d)
{
outputLayer.CalculateIndexedOutput(inputIndex, outputIndex, 1);
var targetOutput = isPositiveTarget ? 1 : 0;
var deltas = NegativeSampleOutput(outputLayer, targetOutput, outputIndex, errorFunctionType, learningRate, momentumMagnitude);
foreach (var previousLayer in outputLayer.PreviousLayers)
{
foreach (var previousPreviousLayer in previousLayer.PreviousLayers)
{
RecurseNegativeSample(previousLayer, previousPreviousLayer, inputIndex, deltas, momentumMagnitude);
}
}
}
private static Dictionary <Node, double> NegativeSampleOutput(Layer outputLayer, double targetOutput, int outputIndex, ErrorFunctionType errorFunctionType, double learningRate, double momentumMagnitude)
{
var errorFunctionDifferential = ErrorFunctionResolver.ResolveErrorFunctionDifferential(errorFunctionType);
var outputNode = outputLayer.Nodes[outputIndex];
var delta = errorFunctionDifferential.Invoke(targetOutput, outputNode.Output)
* outputLayer.ActivationFunctionDifferential(outputNode.Output)
* learningRate;
foreach (var(prevNode, weightForPrevNode) in outputNode.Weights)
{
Backpropagation.UpdateNodeWeight(prevNode, weightForPrevNode, delta, momentumMagnitude);
}
foreach (var(_, weightForPrevLayer) in outputNode.BiasWeights)
{
Backpropagation.UpdateBiasNodeWeight(weightForPrevLayer, delta, momentumMagnitude);
}
return(new Dictionary <Node, double> {
{ outputNode, delta }
});
}
