public IErrorBackPropagationTraining SetUpXOrTraining(double learningRate = 0.5d, double bias = 0d, double momentum = 0d, double slopeMultiplier = 1d)
{
var weightSetterMock = new Mock <IWeightSetter>();
var stepsMock = new Mock <ErrorBackPropagationStepsDependencyFactory>().As <IErrorBackPropagationStepsDependencyFactory>();
stepsMock.CallBase = true;
stepsMock.Setup(s => s.CreateWeightSetter(It.IsAny <double>(), It.IsAny <double>())).Returns(weightSetterMock.Object);
var dependencyFactory = new ErrorBackPropagationDependencyFactory(stepsMock.Object);
var errorBackPropagationTraining = new ErrorBackPropagationBuilder()
.With.ANewLayerOfInputUnits(2)
.ConnectedTo.ANewLayerOfHiddenUnitsOptimizedForTrainingPatterns(_trainingPatterns)
.With.UnitActivation <IdentityUnitActivationTraining>()
.ConnectedTo.ANewLayerOfOutputUnits(1)
.With.OutputUnitActivation <IdentityUnitActivationTraining>()
.And.Bias(bias)
.And.LearningRate(learningRate)
.And.Momentum(momentum)
.And.SlopeMultiplier(slopeMultiplier)
.And.SlopeMultiplier(1d)
.And.SetupNetwork()
.And.NameEverything()
.And.ReadyForTraining(dependencyFactory);
var randomWeights = new Queue <double>(new[] { -1d, -0.5d, 0.5d, 1d, 0.3, 0.2, 0.1, 0.5, 0.2, 0.3, 0.4, -1, -0.9 });
weightSetterMock.Setup(wi => wi.SetWeight(It.IsAny <IConnectionUnderTraining>())).Callback <IConnectionUnderTraining>(c =>
{
var weight = randomWeights.Dequeue();
c.Weight = weight;
randomWeights.Enqueue(weight);
});
return(errorBackPropagationTraining);
}
public async Task ErrorBackPropagationCorrectPerceptronUsed(Type expectedPerceptronType)
{
var oneHot = typeof(OneHotPerceptron) == expectedPerceptronType;
var chain = new ErrorBackPropagationBuilder()
.With.ANewLayerOfInputUnits(3)
.ConnectedTo.ANewLayerOfHiddenUnits(2)
.With.UnitActivation <IdentityUnitActivationTraining>()
.ConnectedTo.ANewLayerOfOutputUnits(1)
.With.OutputUnitActivation <IdentityUnitActivationTraining>()
.And.LearningRate(LearningRate)
.And.Bias(Bias)
.And.SlopeMultiplier(SlopeMultiplier)
.And.Momentum(Momentum);
if (oneHot)
{
chain = chain.And.UseOneHotEncoding();
}
var perceptron = await chain
.And.SetupNetwork(n => _network = n)
.And.ReadyForTraining().TrainAsync(new List <TrainingPattern>(), 999d, 1);
perceptron.GetType().Should().Be(expectedPerceptronType);
}
public static IErrorBackPropagationTraining SetUpXOrTrainingSingleFold <TUnitActivation>(double learningRate = 0.5d, double bias = 0d, double momentum = 0d, double slopeMultiplier = 1d, bool batch = false, Action <TNetwork> networkCallback = null)
where TUnitActivation : IUnitActivationTrainingSingleFold, new()
{
var inventoryAndChaining = new ErrorBackPropagationBuilder()
.With.ANewLayerOfInputUnits(2)
.ConnectedTo.ANewLayerOfHiddenUnits(2).With.UnitActivation <TUnitActivation>()
.ConnectedTo.ANewLayerOfOutputUnits(1).With.OutputUnitActivation <TUnitActivation>();
return(SetUpXOrTraining(inventoryAndChaining, learningRate, bias, momentum, slopeMultiplier, batch, networkCallback));
}
private static IErrorBackPropagationTraining SetUpXOrTraining(double learningRate = 0.5d, double bias = 0d, double momentum = 0d, double slopeMultiplier = 1d, bool oneHot = false, Action <TNetwork> networkCallback = null)
{
var weightSetterMock = new Mock <IWeightSetter>();
var stepsMock = new Mock <ErrorBackPropagationStepsDependencyFactory>().As <IErrorBackPropagationStepsDependencyFactory>();
stepsMock.CallBase = true;
stepsMock.Setup(s => s.CreateWeightSetter(It.IsAny <double>(), It.IsAny <double>())).Returns(weightSetterMock.Object);
var dependencyFactory = new ErrorBackPropagationDependencyFactory(stepsMock.Object);
TNetwork network = null;
var chain = new ErrorBackPropagationBuilder()
.With.ANewLayerOfInputUnits(2)
.ConnectedTo.ANewLayerOfHiddenUnits(3).With.UnitActivationMultiFold <SoftmaxUnitActivationTraining>()
.ConnectedTo.ANewLayerOfOutputUnits(2).With.OutputUnitActivationMultiFold <SoftmaxUnitActivationTraining>()
.And.NetworkErrorFunction <CrossEntropyErrorFunction>()
.And.Bias(bias)
.And.LearningRate(learningRate)
.And.Momentum(momentum)
.And.SlopeMultiplier(slopeMultiplier);
if (oneHot)
{
chain = chain.And.UseOneHotEncoding();
}
var errorBackPropagationTraining = chain
.And.NetworkErrorFunction <DifferenceErrorFunction>()
.And.SetupNetwork(n =>
{
network = n;
if (networkCallback != null)
{
networkCallback(n);
}
})
.And.NameEverything()
.And.ReadyForTraining(dependencyFactory);
// This whole ceremony ensures weight setting is deterministic (same weight to same connection every time even if weights are set asynchronously)
var connections = network.Skip(1).Select(units => units.SelectMany(u => u.IncomingConnections.Select(ic => ic.Properties)));
var randomWeights = new List <double> {
-1d, -0.5d, 0.5d, 1d, 0.3, 0.2, 0.1, 0.5, 0.2, 0.3, 0.4, -1, -0.9, 0.1, 0.5, 1.2, 1.3, 1.1, 0.35, 1.1, 0.1, 0.2, 0.3, -1d, -0.5d, 0.5d, 1d, 0.3, 0.2, 0.1, 0.5, 0.2, 0.3, 0.4, -1, -0.9, 0.1, 0.5, 1.2, 1.3, 1.1, 0.35, 1.1, 0.1, 0.2, 0.3
};
weightSetterMock.Setup(wi => wi.SetWeight(It.IsAny <IConnectionUnderTraining>())).Callback <IConnectionUnderTraining>(c =>
{
var randomWeightIndex = connections.SelectMany(x => x).ToList().IndexOf(c);
c.Weight = randomWeights[randomWeightIndex];
});
return(errorBackPropagationTraining);
}
public void SetUp()
{
_trainingPatterns = new List <TrainingPattern>
{
new TrainingPattern(new [] { 0d, 0d }, new [] { 0d }),
new TrainingPattern(new [] { 1d, 0d }, new [] { 1d }),
new TrainingPattern(new [] { 0d, 1d }, new [] { 1d }),
new TrainingPattern(new [] { 1d, 1d }, new [] { 0d }),
};
var inventoryAndChaining = new ErrorBackPropagationBuilder()
.With.ANewLayerOfInputUnits(2)
.ConnectedTo.ANewLayerOfHiddenUnits(3).With.UnitActivation <SigmoidUnitActivationTraining>()
.ConnectedTo.ANewLayerOfOutputUnits(1).With.OutputUnitActivation <ReluUnitActivationTraining>();
_errorBackPropagationTraining = XOrSetUp.SetUpXOrTraining(inventoryAndChaining, learningRate: 0.8d);
}
