internal virtual void Initialize()
{
Random random;
unchecked
{
random = new Random((int)DateTime.UtcNow.Ticks);
}
// TODO: add support for bias values
for (int li = 0; li < this.layers.Count - 1; li++)
{
NeuralNetworkLayer currentLayer = this.layers[li];
NeuralNetworkLayer nextLayer = this.layers[li + 1];
for (int i = 0; i < currentLayer.Inputs.Length; i++)
{
if (li < this.layers.Count - 1)
{
for (int j = 0; j < nextLayer.Inputs.Length; j++)
{
//currentLayer.Weights[i, j] = random.NextDouble() * 4 - 2;
currentLayer.Weights[i, j] = random.NextDouble();
currentLayer.LastWeightDelta[i, j] = 0.0;
}
}
currentLayer.Errors[i] = double.MinValue;
}
}
}
public NeuralNetworkBuilder AddHiddenLayer(int numberOfNodes, ActivationFunction activationFunction, bool addBias)
{
if (this.currentNetwork.InputLayer == null)
{
throw new NeuralNetworkConfigurationException("You can't add a hidden layer begore adding an input one!");
}
if (this.currentNetwork.OutputLayer != null)
{
throw new NeuralNetworkConfigurationException("You can't add a hidden layer after the output one was defined!");
}
var previousLayer = this.currentNetwork.Layers.Last();
previousLayer.Weights = new double[previousLayer.Inputs.Length, numberOfNodes + (addBias ? 1 : 0)];
previousLayer.LastWeightDelta = new double[previousLayer.Inputs.Length, numberOfNodes + (addBias ? 1 : 0)];
NeuralNetworkLayer hiddenLayer = new NeuralNetworkLayer();
hiddenLayer.Inputs = new double[numberOfNodes + (addBias ? 1 : 0)];
hiddenLayer.Errors = new double[numberOfNodes + (addBias ? 1 : 0)];
hiddenLayer.ActivationFunction = activationFunction;
this.currentNetwork.Layers.Add(hiddenLayer);
return(this);
}
public NeuralNetworkBuilder AddOutputLayer(int numberOfOutputs, ActivationFunction activationFunction)
{
if (this.currentNetwork.InputLayer == null)
{
throw new NeuralNetworkConfigurationException("You can't add an output layer before adding and input one");
}
if (this.currentNetwork.OutputLayer != null)
{
throw new NeuralNetworkConfigurationException("You can't add more than one output layers");
}
var previousLayer = this.currentNetwork.Layers.Last();
previousLayer.Weights = new double[previousLayer.Inputs.Length, numberOfOutputs];
previousLayer.LastWeightDelta = new double[previousLayer.Inputs.Length, numberOfOutputs];
NeuralNetworkLayer outputLayer = new NeuralNetworkLayer();
outputLayer.Inputs = new double[numberOfOutputs];
outputLayer.Errors = new double[numberOfOutputs];
outputLayer.ActivationFunction = activationFunction;
outputLayer.Weights = null;
this.currentNetwork.Layers.Add(outputLayer);
this.currentNetwork.OutputLayer = outputLayer;
return(this);
}
public NeuralNetworkBuilder AddInputLayer(int numberOfInputs, ActivationFunction activationFunction, bool addBias)
{
if (this.currentNetwork.InputLayer != null)
{
throw new NeuralNetworkConfigurationException("You can't have more than one input layers!");
}
NeuralNetworkLayer inputLayer = new NeuralNetworkLayer();
inputLayer.Inputs = new double[numberOfInputs + (addBias ? 1 : 0)];
inputLayer.Errors = new double[numberOfInputs + (addBias ? 1 : 0)];
inputLayer.ActivationFunction = activationFunction;
// postpone initialization of weights
this.currentNetwork.Layers.Add(inputLayer);
this.currentNetwork.InputLayer = inputLayer;
return(this);
}
public double[] Run(double[] inputValues)
{
for (int i = 0; i < inputValues.Length; i++)
{
this.inputLayer.Inputs[i] = inputValues[i];
}
for (int li = 0; li < this.layers.Count - 1; li++)
{
NeuralNetworkLayer currentLayer = this.layers[li];
NeuralNetworkLayer nextLayer = this.layers[li + 1];
for (int i = 0; i < nextLayer.Inputs.Length; i++)
{
double sum = 0.0;
for (int j = 0; j < currentLayer.Inputs.Length; j++)
{
sum += currentLayer.Inputs[j] * currentLayer.Weights[j, i];
}
nextLayer.Inputs[i] = currentLayer.ActivationFunction.Forward(sum);
}
}
return(outputLayer.Inputs);
}