public void Train_WithoutInvalidOutputNeuronIds_ThrowsException()
{
// Arrange
var trainingIteration = new NetworkTrainingIteration();
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 0
});
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 1
});
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 2
});
// 2 output neurons
trainingIteration.Outputs.Add(new NetworkTrainingOutput()
{
NeuronId = 8
});
trainingIteration.Outputs.Add(new NetworkTrainingOutput()
{
NeuronId = -1
});
_trainingIterations.Add(trainingIteration);
_trainingIterations.Add(trainingIteration);
// Act
_network.Train(_trainingIterations);
}
public void Train_WithoutOutputsForEachOutputNeuron_ThrowsException()
{
// Arrange
var trainingIteration = new NetworkTrainingIteration();
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 0
});
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 1
});
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 2
});
trainingIteration.Outputs.Add(new NetworkTrainingOutput());
_trainingIterations.Add(trainingIteration);
// Act
_network.Train(_trainingIterations);
}
public void Train_WithNullDatasetOutputs_ThrowsException()
{
// Arrange
var trainingIteration = new NetworkTrainingIteration
{
Outputs = null
};
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 0
});
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 1
});
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 2
});
_trainingIterations.Add(trainingIteration);
// Act
_network.Train(_trainingIterations);
}
public void Train_WithEmptyDatasetInputs_ThrowsException()
{
// Arrange
var trainingIteration = new NetworkTrainingIteration();
_trainingIterations.Add(trainingIteration);
// Act
_network.Train(_trainingIterations);
}
public void Train_WithValidDataset_ReturnsImprovingCosts()
{
//for (int i = 0; i < 1000; i++)
//{
// Arrange
var trainingIteration = new NetworkTrainingIteration();
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 0, ActivationLevel = .75
});
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 1, ActivationLevel = .75
});
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 2, ActivationLevel = .75
});
// 2 output neurons
trainingIteration.Outputs.Add(new NetworkTrainingOutput()
{
NeuronId = 8, ExpectedActivationLevel = .25
});
trainingIteration.Outputs.Add(new NetworkTrainingOutput()
{
NeuronId = 9, ExpectedActivationLevel = .25
});
_trainingIterations.Add(trainingIteration);
_trainingIterations.Add(trainingIteration);
_trainingIterations.Add(trainingIteration);
_trainingIterations.Add(trainingIteration);
_trainingIterations.Add(trainingIteration);
// Act
_network.RandomizeNetwork();
var trainingIterations = _network.Train(_trainingIterations).ToList();
// Assert
var lastTrainingSet = trainingIterations.First();
foreach (var trainingSet in trainingIterations)
{
Assert.IsTrue(trainingSet.TrainingCost <= lastTrainingSet.TrainingCost);
lastTrainingSet = trainingSet;
}
//}
}
public void Train_WithValidDataset_UpdatesWeightsAndBiases()
{
// Arrange
var trainingIteration = new NetworkTrainingIteration();
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 0, ActivationLevel = .75
});
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 1, ActivationLevel = .75
});
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 2, ActivationLevel = .75
});
// 2 output neurons
trainingIteration.Outputs.Add(new NetworkTrainingOutput()
{
NeuronId = 8, ExpectedActivationLevel = .25
});
trainingIteration.Outputs.Add(new NetworkTrainingOutput()
{
NeuronId = 9, ExpectedActivationLevel = .25
});
_trainingIterations.Add(trainingIteration);
_trainingIterations.Add(trainingIteration);
var inputNeuron = _network.Layers.First().Neurons.First();
var inputNeuronOutgoingConnection = inputNeuron.Connections.OfType <IOutgoingConnection>().First();
var neuronBiasBefore = inputNeuron.Bias;
var connectionWeightBefore = inputNeuronOutgoingConnection.Weight;
// Act
_network.RandomizeNetwork();
_network.Train(_trainingIterations);
// Assert
Assert.IsTrue(neuronBiasBefore != inputNeuron.Bias);
Assert.IsTrue(connectionWeightBefore != inputNeuronOutgoingConnection.Weight);
}
public void Train_WithValidDataset_ReturnsIterationCosts()
{
// Arrange
var trainingIteration = new NetworkTrainingIteration();
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 0, ActivationLevel = .75
});
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 1, ActivationLevel = .75
});
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
NeuronId = 2, ActivationLevel = .75
});
// 2 output neurons
trainingIteration.Outputs.Add(new NetworkTrainingOutput()
{
NeuronId = 8, ExpectedActivationLevel = .25
});
trainingIteration.Outputs.Add(new NetworkTrainingOutput()
{
NeuronId = 9, ExpectedActivationLevel = .25
});
_trainingIterations.Add(trainingIteration);
_trainingIterations.Add(trainingIteration);
_trainingIterations.Add(trainingIteration);
_trainingIterations.Add(trainingIteration);
_trainingIterations.Add(trainingIteration);
// Act
_network.RandomizeNetwork();
var costs = _network.Train(_trainingIterations);
// Assert
Assert.IsNotNull(costs);
Assert.IsTrue(costs.Count() == 5);
}
/// <summary>
/// Converts each of the <paramref name="source"/> items to a trainingIteration.
/// </summary>
/// <param name="source">The source dataset that is to be converted to training iterations.</param>
/// <returns>Returns the converted <paramref name="source"/> items as trainingIterations.</returns>
private IEnumerable <INetworkTrainingIteration> ConvertToTrainingIterations(IEnumerable <GetTrainingDataset_Result> source)
{
var trainingIterations = new List <INetworkTrainingIteration>();
if (source != null && source.Count() >= 1)
{
var properties = source.First().GetType().GetProperties();
var inputProperties = properties.Where(p => p.Name.StartsWith("I_"));
var outputProperties = properties.Where(p => p.Name.StartsWith("O_"));
// Convert each dataset entry into a network training iteration
foreach (var entry in source)
{
var trainingIteration = new NetworkTrainingIteration();
// Dynamically add all inputs
foreach (var inputProperty in inputProperties)
{
trainingIteration.Inputs.Add(new NetworkTrainingInput()
{
ActivationLevel = (bool)inputProperty.GetValue(entry) ? 1.0 : 0.0, Description = inputProperty.Name
});
}
// Dynamically add all outputs
foreach (var outputProperty in outputProperties)
{
trainingIteration.Outputs.Add(new NetworkTrainingOutput()
{
ExpectedActivationLevel = (bool)outputProperty.GetValue(entry) ? 1.0 : 0.0, Description = outputProperty.Name
});
}
trainingIterations.Add(trainingIteration);
}
}
return(trainingIterations);
}
