private void CalculateGradientForSingleTrainingExample(Network network, IErrorFunction errorFunction, ref List <List <NeuronData> > intermediateResults, float[] trainingInput, float[] trainingDesiredOutput)
{
List <float[]> activations = new List <float[]>();
List <float[]> zValues = new List <float[]>();
network.Compute(this, trainingInput, ref activations, ref zValues, false); //dont flush working cache
var lastLayerGradient = intermediateResults.Last();
List <float> delta_k_holder = new List <float>();
CalculateOutputLayerGradient(network, errorFunction, ref lastLayerGradient, ref delta_k_holder, activations, trainingInput, zValues, trainingDesiredOutput);
for (int i = network.layers.Count - 2; i >= 0; --i)
{
var layerGradient = intermediateResults[i];
CalculateHiddenLayerGradient(network, i, ref layerGradient, ref delta_k_holder, i == 0 ? trainingInput : activations[i - 1], zValues);
}
}
private void CalculateOutputLayerGradient(Network network, IErrorFunction errorFunction, ref List <NeuronData> gradientData, ref List <float> delta_k_vector, List <float[]> activations, float[] trainingInput, List <float[]> zValues, float[] desiredOutput)
{
var prevActivations = activations.Count <= 1 ? trainingInput : activations[activations.Count - 2];
int lastLayerWeightCount = network.layers.Last().GetWeightsPerNeuron();
int lastLayerNeuronCount = network.layers.Last().GetNeuronCount();
for (int i = 0; i < lastLayerNeuronCount; i++)
{
float outputValue = activations.Last()[i];
float delta_k = errorFunction.CalculateDelta(zValues.Last()[i], outputValue, desiredOutput[i], network.activationFunction);
var gradientDataItem = gradientData[i];
//Assert(gradientData[i].weights.Length == prevActivations.Length);
for (int j = 0; j < lastLayerWeightCount; j++)
{
gradientDataItem.weights[j] += delta_k * prevActivations[j];
}
gradientDataItem.bias += delta_k;
delta_k_vector.Add(delta_k);
}
}
