private double feedForward(double[][,] input, Class cls)
{
Net.Calculate(input, m_Values);
var lidx = Net.LayerCount - 1;
var result = m_Values[lidx];
var errors = m_Errors[lidx];
var len = result.GetLength(0);
var output = new double[len];
for (int j = 0; j < len; j++)
{
output[j] = result[j][0, 0];
}
var expect = m_ExpectedOutputs[cls];
var llayer = Net[lidx];
for (int p = 0; p < llayer.OutputDepth; p++)
{
var ej = m_LossFunction.Derivative(p, output, expect);
var value = result[p][0, 0];
var deriv = (llayer.ActivationFunction != null) ? llayer.ActivationFunction.DerivativeFromValue(value) : 1;
errors[p][0, 0] = ej * deriv / m_BatchSize;
}
return(m_LossFunction.Value(output, expect) / m_BatchSize);
}
private double feedForward(NeuralNetwork net, double[] input, double[] expected)
{
var output = net.Calculate(input);
var llayer = net[net.LayerCount - 1];
var errors = m_Errors[net.LayerCount - 1];
for (int j = 0; j < m_OutputDim; j++)
{
var neuron = llayer[j];
var ej = m_LossFunction.Derivative(j, output, expected);
errors[j] = ej * neuron.Derivative;
}
return(m_LossFunction.Value(output, expected));
}
protected override void DoLossDerivative(Tensor correct, ILossFunction lossFunction, Tensor dy)
{
lossFunction.Derivative(this, correct, dy);
}