public float CalcCostFunction(float[] expected, CostFunctionType cType)
{
for (int i = 0; i < numberofOutputs; i++)
{
error[i] = outputs[i] - expected[i];
}
float sol = 0.0f;
switch (cType)
{
case CostFunctionType.MSE:
int cnt = 0; float sum = 0;
for (int i = 0; i < error.Length; i++)
{
sum += error[i];
cnt++;
}
sol = sum / cnt;
break;
case CostFunctionType.SSE:
sum = 0;
for (int i = 0; i < error.Length; i++)
{
sum += error[i];
}
sol = sum;
break;
default:
break;
}
return(sol);
}
public NetworkConfiguration(
CostFunctionType costFunctionType,
LearningRateDecayer decayer,
int inputDimenision,
int inputChannels)
{
Decayer = decayer;
CostFunctionType = costFunctionType;
InputDimenision = inputDimenision;
InputChannels = inputChannels;
}
public static ICostFunction Produce(CostFunctionType type)
{
switch (type)
{
case CostFunctionType.Quadratic: return(new QuadraticCostFunction());
case CostFunctionType.SoftMax: return(new SoftMaxCostFunction());
case CostFunctionType.CrossEntropy: return(new CrossEntropyCostFunction());
default: throw new Exception("Cost function is not supported");
}
}
public static (CostFunction, CostFunctionPrime) GetCostFunctions(CostFunctionType type)
{
switch (type)
{
case CostFunctionType.Quadratic: return(CostFunctions.QuadraticCost, CostFunctions.QuadraticCostPrime);
case CostFunctionType.CrossEntropy: return(CostFunctions.CrossEntropyCost, CostFunctions.CrossEntropyCostPrime);
case CostFunctionType.LogLikelyhood: return(CostFunctions.LogLikelyhoodCost, CostFunctions.CrossEntropyCostPrime);
default:
throw new InvalidOperationException("Unsupported cost function");
}
}
public NeuralNetwork(
int[] layerLengths,
float learnRate,
int batchSize,
float dropoutKeepRate,
float weightDecayRate,
ActivisionFunctionType activisionType,
ActivisionFunctionType activisionTypeOutput,
CostFunctionType costType,
InitializationType initializationType,
int initializationSeed,
float activisionCoeffitient,
int dropoutSeed
)
{
m_layerLengths = layerLengths;
m_layerCount = m_layerLengths.Length;
m_learnRate = learnRate;
m_activisionCoeffitient = activisionCoeffitient;
m_dropoutKeepRate = dropoutKeepRate;
m_weightDecayRate = weightDecayRate;
m_initDropoutSeed = dropoutSeed;
m_currentDropoutSeed = dropoutSeed;
if (batchSize <= 0)
{
Debug.Log("Info: batch size was <= 0 (" + batchSize + "). It was set to a default value of 1!");
m_batchSize = 1;
}
m_batchSize = batchSize;
m_activisionFunctionType = activisionType;
m_activisionFunctionTypeOutput = activisionTypeOutput;
m_costFunctionType = costType;
m_initializationType = initializationType;
SetActivisionFunction(activisionType);
InitializeBiases(null, initializationSeed);
for (int i = 1; i < m_layerCount; i++)
{
initializationSeed += m_layerLengths[i];
}
InitializeWeights(null, initializationSeed);
InitializeBatch();
InitializeBackPropagation();
//m_activisionFunctionType = ActivisionFunctionType.Sigmoid;
}
public static LayerFactory FullyConnected(
int neurons, ActivationFunctionType activation, CostFunctionType cost,
WeightsInitializationMode weightsMode = WeightsInitializationMode.GlorotUniform, BiasInitializationMode biasMode = BiasInitializationMode.Zero)
=> input => new OutputLayer(input, neurons, activation, cost, weightsMode, biasMode);
