public void IfTargetLengthDoesNotMatchNumOutputs_Throw()
{
var nn = SampleInputs.GetSample1HiddenLayerPerceptron();
Action action = () => nn.CalculateGradients(SampleInputs.GetSampleInputs(), new[] { 0.0 });
action.ShouldThrow <NeuralNetworkException>()
.WithMessage($"*Argument 'targets' should have width {nn.NumOutputs}; was 1.*");
}
public void ShouldIncludeInputsWithBias()
{
var inputs = SampleInputs.GetSampleInputs();
var nn = SampleInputs.GetSample1HiddenLayerPerceptron();
var result = nn.FeedForward(inputs);
var expected = new double[inputs.Length + 1];
Array.Copy(inputs, expected, inputs.Length);
expected[inputs.Length] = 1;
result.InputWithBias.Should().Equal(expected);
}
public void ShouldCalculateCorrectly_OneHiddenLayer()
{
var nn = SampleInputs.GetSample1HiddenLayerPerceptron();
var result = nn.FeedForward(SampleInputs.GetSampleInputs());
result.HiddenLayers.Should().HaveCount(1);
var expectedHidden = new[] { 0.336375544336332, 0.347853774202261, 0.359228665810271, 1.0 };
result.HiddenLayers[0].ShouldApproximatelyEqual(expectedHidden, 1e-12);
var expectedOutput = new[] { 0.449090580148409, 0.550909419851591 };
result.Output.ShouldApproximatelyEqual(expectedOutput, 1e-12);
}
public void ShouldCalculateGrads_OneHiddenLayerCase()
{
// See MultilayerPerceptronTests.xlsx in this folder for calculations of
// expected outputs.
var nn = SampleInputs.GetSample1HiddenLayerPerceptron();
var result = nn.CalculateGradients(SampleInputs.GetSampleInputs(), SampleInputs.GetSampleTargets());
result.Should().NotBeNullOrEmpty();
result.Should().HaveCount(2);
var expectedHiddenGrads = new[] { 0.00488574841598145, 0.00977149683196289, 0.0488574841598144, 0.00484248141480251, 0.00968496282960501, 0.0484248141480251, 0.00479817184666825, 0.0095963436933365, 0.0479817184666825 };
var expectedOutputGrads = new[] { -0.185312455982592, -0.191635920938954, -0.197902455875597, -0.550909419851591, 0.185312455982592, 0.191635920938954, 0.197902455875597, 0.550909419851591 };
result[1].ShouldApproximatelyEqual(expectedOutputGrads, 1e-12);
result[0].ShouldApproximatelyEqual(expectedHiddenGrads, 1e-12);
}
public void ShouldCalculateGrads_TwoHiddenLayersCase()
{
// See MultilayerPerceptronTests.xlsx in this folder for calculations of
// expected outputs.
var nn = SampleInputs.GetSample2HiddenLayerPerceptron();
var result = nn.CalculateGradients(SampleInputs.GetSampleInputs(), SampleInputs.GetSampleTargets());
result.Should().NotBeNullOrEmpty();
result.Should().HaveCount(3);
var expectedHidden1Grads = new[] { 0.00197922538436067, 0.00395845076872135, 0.0197922538436067, 0.00200603359153654, 0.00401206718307308, 0.0200603359153654, 0.00203160809301684, 0.00406321618603367, 0.0203160809301684 };
var expectedHidden2Grads = new[] { 0.00976178390929107, 0.0100948877912437, 0.0104249927460735, 0.0290204923445045, 0.00720465406099808, 0.00745050033849661, 0.00769413326721845, 0.0214184835440782 };
var expectedOutputGrads = new[] { -0.390542035775992, -0.44184361704256, -0.561729160466126, 0.390542035775992, 0.44184361704256, 0.561729160466125 };
result[2].ShouldApproximatelyEqual(expectedOutputGrads, 1e-12);
result[1].ShouldApproximatelyEqual(expectedHidden2Grads, 1e-12);
result[0].ShouldApproximatelyEqual(expectedHidden1Grads, 1e-12);
}
public void ShouldCalculateCorrectly_TwoHiddenLayers()
{
var nn = SampleInputs.GetSample2HiddenLayerPerceptron();
var result = nn.FeedForward(SampleInputs.GetSampleInputs());
result.HiddenLayers.Should().HaveCount(2);
var expectedHidden1 = new[] { 0.336375544336332, 0.347853774202261, 0.359228665810271, 1.0 };
result.HiddenLayers[0].ShouldApproximatelyEqual(expectedHidden1, 1e-12);
var expectedHidden2 = new[] { 0.695249709756778, 0.78657767504168, 1 };
result.HiddenLayers[1].ShouldApproximatelyEqual(expectedHidden2, 1e-12);
var expectedOutput = new[] { 0.438270839533874, 0.561729160466125 };
result.Output.ShouldApproximatelyEqual(expectedOutput, 1e-12);
}