public void CanCalculateBackProp__Given__LayersOfVariousSizes(double[] inputs, double[] inputToHiddenWeights, double[] hiddenToOutputWeights, double target)
{
var net = NNSigmoid.FromFlatWeightArrays(inputs.Length, inputToHiddenWeights, hiddenToOutputWeights);
var input = new MatrixD(inputs.Length, 1, inputs);
var outSize = hiddenToOutputWeights.Length * inputs.Length / inputToHiddenWeights.Length;
var targets = new MatrixD(outSize, 1, target);
var calculatedDeltas = net.backprop(input, targets);
}
public void InitalisesWithRandomWeightsAndBiases()
{
var net = new NNSigmoid(new[] { 10, 20, 13 });
net.Biases.ShouldAll(b =>
{
b.ByRows().Where(r => r.Length > 1).ShouldAll(
row =>
{
var avg = row.Average().ShouldBeBetween(0.1, 0.9);
row.Average(x => Math.Abs(x - avg)).ShouldBeBetween(0.1, 0.7);
});
});
}
public void Given__223_SigmoidNetwork(double[] inputs, double[] inputToHiddenWeights, double[] hiddenToOutputWeights, double target)
{
var net = NNSigmoid.FromFlatWeightArrays(inputs.Length, inputToHiddenWeights, hiddenToOutputWeights);
var input = new MatrixD(inputs.Length, 1, inputs);
var targets = new MatrixD(3, 1, target);
var calculatedDeltas = net.backprop(input, targets);
var exOutputDelta1 = -0.040681125112339026d;
var exOutputDelta2 = hiddenToOutputWeights[1].Equals(0) ? 0 : exOutputDelta1;
var exOutputDelta3 = hiddenToOutputWeights[2].Equals(0) ? 0 : exOutputDelta1;
var exHiddenDelta00 = -0.0023685025015371172d;
var exHiddenDelta01 = -0.0075927347073177845d;
//calculatedDeltas.Item1.First().ShouldEqualByValue(expectedDeltas.HiddenWeights);
//calculatedDeltas.Item1.Last().ShouldEqualByValue(expectedDeltas.OutputWeights);
calculatedDeltas.Item2.First().ShouldEqualByValue(new MatrixD(2, 1, 0));
calculatedDeltas.Item2.Last().ShouldEqualByValue(new MatrixD(3, 1, 0));
}
public void Given__221_SigmoidNetwork(double[] inputs, double[] inputToHiddenWeights, double[] hiddenToOutputWeights, double target)
{
var net = NNSigmoid.FromFlatWeightArrays(inputs.Length, inputToHiddenWeights, hiddenToOutputWeights);
var input = new MatrixD(inputs.Length, 1, inputs);
var targets = new MatrixD(1, 1, target);
var calculatedDeltas = net.backprop(input, targets);
/*
* 0.680267196698649d, 0.663738697404353d, 0.690283492907644d
*/
var output = net.feedforward(input);
var exOutputDelta = -0.040681125112339026d;
var exHiddenDelta0 = -0.0023685025015371172d;
var exHiddenDelta1 = -0.0075927347073177845d;
var expectedDeltaValues = new
{
OutputBiases = new MatrixD(1, 1, exOutputDelta),
HiddenBiases = new MatrixD(2, 1, new [] { -0.00260919207722715d, -0.00782757623168145d }),
OutputWeights = new MatrixD(
new[, ]
{
{ -0.027674034938717864 }, // -0.040681125112339026d * 0.680267196698649d +/- last two d.p.s
{ -0.027001636991007407 } // -0.040681125112339026d * 0.663738697404353d +/- last two d.p.s
}),
HiddenWeights = new MatrixD(
new[, ] {
{ -0.000913217227029503d, -0.0026574571475612243d },
{ -0.0021316522513834054d, -0.0068334612365860059d }
}),
};
//
calculatedDeltas.Item2.Last().ShouldEqualByValue(expectedDeltaValues.OutputBiases, "OutputBiases");
calculatedDeltas.Item1.Last().ShouldEqualByValue(expectedDeltaValues.OutputWeights, "OutputWeights");
calculatedDeltas.Item2.First().ShouldEqualByValue(expectedDeltaValues.HiddenBiases, "HiddenBiases");
calculatedDeltas.Item1.First().ShouldEqualByValue(expectedDeltaValues.HiddenWeights, "HiddenWeights");
}
