public NeuralNetworkDenseLayer(NeuralNetworkNeuron[] neurons, NeuralNetworkActivationFunctionType activationFunctionType)
{
Type = NeuralNetworkLayerType.Dense;
Neurons = neurons;
ActivationFunctionType = activationFunctionType;
Weights = new float[Neurons.Length][];
for (int i = 0; i < Neurons.Length; i++)
{
Weights[i] = Neurons[i].Weights;
}
}
public float ForwardPropagation(float[] inputs, NeuralNetworkActivationFunctionType activationFunctionType)
{
float sum = 0;
#if NETFRAMEWORK
NeuralNetworkAPI.Neuron(inputs, Weights, Weights.Length, ref sum);
#else
for (int i = 0; i < Weights.Length; i++)
{
sum += inputs[i] * Weights[i];
}
;
#endif
return(NeuralNetworkAPI.ActivationFunction(activationFunctionType, sum + Bias));
}
public NeuralNetworkConv2DLayer(
NeuralNetworkKernel[] kernels,
int width, int height,
int strideX, int strideY,
NeuralNetworkActivationFunctionType activationFunctionType,
NeuralNetworkPaddingType paddingType
)
{
Type = NeuralNetworkLayerType.Conv2D;
Kernels = kernels;
Width = width;
Height = height;
StrideX = strideX;
StrideY = strideY;
ActivationFunctionType = activationFunctionType;
PaddingType = paddingType;
}
public static float ActivationFunction(NeuralNetworkActivationFunctionType type, float value)
{
switch (type)
{
case NeuralNetworkActivationFunctionType.ReLU:
return(value < 0 ? 0 : value);
case NeuralNetworkActivationFunctionType.Softmax:
return(value);
case NeuralNetworkActivationFunctionType.Sigmoid:
return((float)(1 / (1 + Math.Exp(-value))));
case NeuralNetworkActivationFunctionType.Linear:
return(value);
default:
throw new NotImplementedException(type.ToString());
}
}
