public Vector <double>[] GetFeatures()
{
var allFeatures = new Vector <double> [NeuronsCount];
for (var i = 0; i < Weights.RowCount; i++)
{
var norm = Weights.Row(i).L2Norm();
var features = Weights.Row(i).Divide(norm);
allFeatures[i] = features;
}
return(allFeatures);
}
public Vector <double> ActivationFunctionPrime(Vector <double> output)
{
switch (CurrentActivationFunction)
{
case Network.ActivationFunction.Sigmoid:
return(HelperFunctions.SigmoidPrime(output));
case Network.ActivationFunction.Tanh:
return(HelperFunctions.TanhPrime(output));
default:
throw new ArgumentOutOfRangeException();
}
}
/// <summary>
///
/// </summary>
/// <param name="inputsCount">Number of inputs to each vector</param>
/// <param name="outputsCount">Same as neuron count in this layer</param>
/// <param name="activationFunction"></param>
/// <param name="initialWeightsRange"></param>
public HiddenLayer(
int inputsCount,
int outputsCount,
ActivationFunction activationFunction,
double initialWeightsRange
)
{
CurrentActivationFunction = activationFunction;
InputsCount = inputsCount;
NeuronsCount = outputsCount;
InitialWeightsRange = initialWeightsRange;
Weights = GetNewWeightsMatrix(false);
Biases = GetNewBiasesVector(false);
}
/// <summary>
/// Takes net as input (this layer output before activation function)
/// </summary>
/// <param name="output"></param>
/// <returns></returns>
public Vector <double> GetActivation(Vector <double> output)
{
return(ActivationFunction(output));
}
public Vector <double> GetNet(Vector <double> inputs)
{
//(O x I * I x 1) o 0 x 1 = O x 1
return(Weights * inputs + Biases);
}
/// <summary>
///
/// </summary>
/// <param name="inputs">Vector dimensions: I x 1</param>
/// <returns></returns>
public override Vector <double> Feedforward(Vector <double> inputs)
{
var output = GetNet(inputs);
return(GetActivation(output));
}
