/// <summary>
/// Gets the activation value of the neuron
/// </summary>
/// <param name="activationTable">An activation table with all activations up to L-1 filled in</param>
/// <returns>the float value of the neuron activation</returns>
public float GetActivation(float[,] activationTable)
{
var z = GetInputValue(activationTable);
var activation = _activationFunction.GetValue(z); //activation = p.activationMethod(z)
return(activation);
}
/// <summary>
/// Returns the computed value of each unit in the layer
///
/// For each unit, take input from upstream layer, multiply by weights
/// and then apply transfer function
/// </summary>
/// <returns></returns>
public virtual double GetValue(long idxUnit, long row)
{
double net = 0;
for (int idxUp = 0; idxUp < Weights.Length - 1; idxUp++)
{
net += Weights[idxUp][idxUnit]
* GetValueUpstreamLayer(idxUp, row);
}
//For Bias term
net += Weights[Weights.Length - 1][idxUnit] * 1.0;
return(_activationFunction.GetValue(net));
}
/// <summary>
/// Computes the FilterMap by convolution over each Input array
/// from Upstream layer
///
/// This function contains the sliding convolution window logic
/// </summary>
/// <param name="upStreamLayer"></param>
public override void ComputeValueMap()
{
int filterSize = _weights.Length; //Filter is same as window
//Slide the filter Window to compute value for each Cell in ValueMap (VM)
for (int idxRowVM = 0; idxRowVM < GetValueMapNoOfRows(); idxRowVM++)
{ //idxRowVM
for (int idxColVM = 0; idxColVM < GetValueMapNoOfColumns(); idxColVM++)
{ //idxColVM
double sum = 0;
int idxUpValueMapCol = 0, idxUpValueMapRow = 0;
//Set the filter window co-ordinates in the upstream value map
//relative to the current Value Map
int upMapLeftCol = idxColVM * _stride;
int upMapLeftRow = idxRowVM * _stride;
//TODO: Assign value
for (int idxFilter = 0; idxFilter < _upStreamLayer.GetNumberOfFilterUnits();
idxFilter++) //For each Filter Unit Upstream
{
SingleFilterUnit upFilterUnit =
(SingleFilterUnit)_upStreamLayer.GetFilterUnit(idxFilter);
for (int idxFilterRow = 0; idxFilterRow < filterSize; idxFilterRow++)
{
for (int idxFilterCol = 0; idxFilterCol < filterSize; idxFilterCol++)
{
//Update value
idxUpValueMapCol = upMapLeftCol + idxFilterCol;
idxUpValueMapRow = upMapLeftRow + idxFilterRow;
//Multiply upStream FilterMap by current Filter weights
sum += _weights[idxFilterCol][idxFilterRow] *
upFilterUnit.GetValueMapAtIndex(idxUpValueMapCol,
idxUpValueMapRow);
} //idxFilterCol
//Do for bias term
sum += _bias * 1.0;
} //idxFilterRow
} //idxFilter
//Normalize Sum
sum = sum / (filterSize * 2); //divide sum by filter Grid size
//Apply activation function - this saves time
sum = _activationFunction.GetValue(sum);
ValueMap[idxColVM][idxRowVM] = sum; //Assign value to FilterMap
} //idxColFM
} //idxRowFM
}
public IList <double> GetOutputs(IList <double> inputs, IActivationFunction activationFunction, double bias)
{
if (inputs.Count != InputSize)
{
throw new ArgumentOutOfRangeException(nameof(inputs),
"Inputs size should be the same as layer input size.");
}
var output = new double[OutputSize];
for (var row = 0; row < OutputSize; row++)
{
var value = bias;
for (var col = 0; col < InputSize; col++)
{
value += WeightMatrix[row, col] * inputs[col];
}
output[row] = activationFunction.GetValue(value);
}
return(output);
}
