public static double TotalError(NeuralNetwork net, TrainingSet trainingSamples)
{
var netOutputs = trainingSamples.Select(x => net.ForwardPass(x.Input));
var expectedOutpus = trainingSamples.Select(x => x.Ideal);
var trainingSampleErrors = expectedOutpus.Zip(netOutputs, TrainingSampleError);
return(trainingSampleErrors.Sum() / trainingSamples.Count);
}
private static void PerformTrainingEpoch(NeuralNetwork net, TrainingConfiguration config, TrainingSet trainingSet)
{
trainingSet.ForEach(sample =>
{
var netOutput = net.ForwardPass(sample.Input);
var outputDeltas = netOutput.Zip(sample.Ideal, (actual, ideal) => OutputDelta(ideal, actual)).ToList();
var outputWeightsAdjustments = outputDeltas
.Select(d => net.HiddenLayer.Select(h => h.LastOutput).Select(h => OutputWeightDelta(d, h)));
var outputDeltaPairs = net.OutputLayer.Zip(outputDeltas, (output, oDelta) => new { output, oDelta });
var outputWeightedDeltaSums = net.HiddenLayer
.Select((_, idx) => outputDeltaPairs
.Select(outputDeltaPair => outputDeltaPair.output.Weights[idx] * outputDeltaPair.oDelta)
.Sum());
var hiddenDeltaPairs = net.HiddenLayer.Select(h => h.LastOutput)
.Zip(outputWeightedDeltaSums, (h, delta) => new { HiddenOutput = h, Delta = delta });
var hiddenWeightsAdjustments =
hiddenDeltaPairs.Select(hd => sample.Input.Select(i => HiddenWeightDelta(hd.Delta, hd.HiddenOutput, i)));
UpdateWeights(outputWeightsAdjustments, hiddenWeightsAdjustments, net, config);
});
}