/// <summary>
/// Create new NeuralNetworkTrainer.
/// </summary>
/// <param name="nn">NeuralNetwork to train.</param>
/// <param name="inputDat">Input data.</param>
/// <param name="targetDat">Target data.</param>
/// <param name="maxUnrollLen">Memory state unroll times, for recurring layers.</param>
/// <param name="losType">Loss calculation type, NeuralNetworkTrainer.LOSS_TYPE_AVERAGE/MAX/CROSSENTROPY.</param>
public NeuralNetworkTrainer(NeuralNetwork nn, float[][] inputDat, float[][] targetDat, int maxUnrollLen, int losType)
{
neuralNetwork = nn;
inputData = inputDat;
targetData = targetDat;
maxUnrollLength = maxUnrollLen;
if (maxUnrollLength < 1)
{
maxUnrollLength = 1;
}
lossType = losType;
//check for recurring layer, if need to stack and unroll
if (nn.outputLayer.recurring)
{
hasRecurring = true;
}
else
{
for (int i = 0; i < nn.hiddenLayers.Length; i++)
{
if (nn.hiddenLayers[i].recurring)
{
hasRecurring = true;
break;
}
}
}
derivatives.Setup(nn);
adagradMemory.Setup(nn);
adagradMemory.Reset();
int tunrollLen = maxUnrollLength;
if (!hasRecurring)
{
tunrollLen = 1;
}
stackedRuntimeContext = new NeuralNetworkContext[tunrollLen];
stackedFullContext = new NeuralNetworkFullContext[tunrollLen];
stackedDerivativeMemory = new NeuralNetworkPropagationState[tunrollLen];
for (int i = 0; i < stackedRuntimeContext.Length; i++)
{
stackedRuntimeContext[i] = new NeuralNetworkContext();
stackedRuntimeContext[i].Setup(nn);
stackedFullContext[i] = new NeuralNetworkFullContext();
stackedFullContext[i].Setup(nn);
stackedDerivativeMemory[i] = new NeuralNetworkPropagationState();
stackedDerivativeMemory[i].Setup(nn, stackedRuntimeContext[i], stackedFullContext[i], derivatives);
}
}
/// <summary>
/// Create new NeuralNetworkQLearning system.
/// </summary>
/// <param name="nn">NeuralNetwork to train.</param>
/// <param name="inputDat">Input data.</param>
/// <param name="targetDat">Target data.</param>
public NeuralNetworkQLearning(NeuralNetwork nn, int maxUnrollLen, string sessionsFileName)
{
neuralNetwork = nn;
learningSessionsFile = sessionsFileName;
maxUnrollLength = maxUnrollLen;
if (maxUnrollLength < 1)
{
maxUnrollLength = 1;
}
//check for recurring layer, if need to stack and unroll
for (int i = 0; i < nn.hiddenLayers.Length; i++)
{
if (nn.hiddenLayers[i].recurring)
{
hasRecurring = true;
break;
}
}
derivatives.Setup(nn);
if (!hasRecurring)
{
maxUnrollLength = 1;
}
stackedRuntimeContext = new NeuralNetworkContext[maxUnrollLength];
stackedFullContext = new NeuralNetworkFullContext[maxUnrollLength];
stackedDerivativeMemory = new NeuralNetworkPropagationState[maxUnrollLength];
for (int i = 0; i < stackedRuntimeContext.Length; i++)
{
stackedRuntimeContext[i] = new NeuralNetworkContext();
stackedRuntimeContext[i].Setup(nn);
stackedFullContext[i] = new NeuralNetworkFullContext();
stackedFullContext[i].Setup(nn);
stackedDerivativeMemory[i] = new NeuralNetworkPropagationState();
stackedDerivativeMemory[i].Setup(nn, stackedRuntimeContext[i], stackedFullContext[i], derivatives);
}
/*
* if (hasRecurring)
* {
* recurringMemoryState = new float[nn.hiddenLayers.Length][];
* for (int i = 0; i < nn.hiddenLayers.Length; i++)
* {
* if (nn.hiddenLayers[i].recurring)
* {
* recurringMemoryState[i] = new float[nn.hiddenLayers[i].numberOfNeurons];
* }
* }
* }*/
}
public void Setup(NeuralNetwork nn, NeuralNetworkContext context, NeuralNetworkFullContext fullCtx, NeuralNetworkDerivativeMemory derivMem)
{
//initialize memory buffers
state = new float[nn.hiddenLayers.Length][];
weights = new float[nn.hiddenLayers.Length + 1][];
biases = new float[nn.hiddenLayers.Length + 1][];
buf = new float[nn.hiddenLayers.Length + 1][];
recurrBuf = new float[nn.hiddenLayers.Length][];
biasMems = derivMem.biasMems;
weightMems = derivMem.weightMems;
recurrWeightMems = derivMem.recurrWeightMems;
recurrWeights = new float[nn.hiddenLayers.Length][];
derivativeMemory = derivMem;
for (int i = 0; i < nn.hiddenLayers.Length; i++)
{
state[i] = new float[nn.hiddenLayers[i].numberOfNeurons];
weights[i] = nn.hiddenConnections[i].weights;
biases[i] = nn.hiddenLayers[i].biases;
if (i == 0)
{
buf[i] = context.inputData;
}
else
{
buf[i] = fullCtx.hiddenBuffer[i - 1];
}
if (nn.hiddenLayers[i].recurring)
{
recurrWeights[i] = nn.hiddenRecurringConnections[i].weights;
recurrBuf[i] = fullCtx.hiddenRecurringBuffer[i];
}
}
int lid = nn.hiddenLayers.Length;
weights[lid] = nn.outputConnection.weights;
biases[lid] = nn.outputLayer.biases;
if (lid > 0)
{
buf[lid] = fullCtx.hiddenBuffer[lid - 1];
}
else
{
buf[lid] = context.inputData;
}
}
/// <summary>
/// Create new NeuralNetworkGenerator.
/// </summary>
/// <param name="nn">NeuralNetwork to train.</param>
public NeuralNetworkGenerator(NeuralNetwork nn, int maxUnrollLen)
{
neuralNetwork = nn;
maxUnrollLength = maxUnrollLen;
//check for recurring layer, if need to stack and unroll
for (int i = 0; i < nn.hiddenLayers.Length; i++)
{
if (nn.hiddenLayers[i].recurring)
{
hasRecurring = true;
break;
}
}
derivatives.Setup(nn);
if (hasRecurring)
{
recurringBPBuffer = new float[nn.hiddenLayers.Length][];
for (int i = 0; i < recurringBPBuffer.Length - 1; i++)
{
if (nn.hiddenLayers[i].recurring)
{
recurringBPBuffer[i] = new float[nn.hiddenLayers[i].numberOfNeurons];
}
}
}
else
{
maxUnrollLen = 1;
}
stackedRuntimeContext = new NeuralNetworkContext[maxUnrollLen];
stackedFullContext = new NeuralNetworkFullContext[maxUnrollLen];
stackedDerivativeMemory = new NeuralNetworkPropagationState[maxUnrollLen];
for (int i = 0; i < maxUnrollLen; i++)
{
stackedRuntimeContext[i] = new NeuralNetworkContext();
stackedRuntimeContext[i].Setup(nn);
stackedFullContext[i] = new NeuralNetworkFullContext();
stackedFullContext[i].Setup(nn);
stackedDerivativeMemory[i] = new NeuralNetworkPropagationState();
stackedDerivativeMemory[i].Setup(nn, stackedRuntimeContext[i], stackedFullContext[i], derivatives);
stackedDerivativeMemory[i].inputMem = new float[nn.inputLayer.numberOfNeurons];
}
}
/// <summary>
/// Run neural network backwards calculating derivatives to use for adagrad or generation.
/// </summary>
/// <param name="target"></param>
/// <param name="context"></param>
/// <param name="fullContext"></param>
/// <param name="derivMem"></param>
public void ExecuteBackwards(float[] target, NeuralNetworkContext context, NeuralNetworkFullContext fullContext, NeuralNetworkPropagationState propState, int lossType, int crossEntropyTarget)
{
//prepare for back propagation
for (int i = 0; i < propState.state.Length; i++)
{
Utils.Fill(propState.state[i], 0.0f);
}
//back propagation + calculate max loss
int lid = hiddenLayers.Length;
float lossAvg = 0.0f;
for (int i = 0; i < target.Length; i++)
{
float deriv = context.outputData[i] - target[i];
if (lossType == NeuralNetworkTrainer.LOSS_TYPE_MAX)
{
float aderiv = Math.Abs(deriv);
if (aderiv > lossAvg)
{
lossAvg = aderiv;
}
}
else if (lossType == NeuralNetworkTrainer.LOSS_TYPE_AVERAGE)
{
lossAvg += Math.Abs(deriv);
}
backpropagate(lid, i, deriv, propState);
}
if (lossType == NeuralNetworkTrainer.LOSS_TYPE_AVERAGE)
{
lossAvg /= (float)target.Length;
}
else
{
if (lossType == NeuralNetworkTrainer.LOSS_TYPE_CROSSENTROPY && crossEntropyTarget != -1)
{
lossAvg = (float)-Math.Log(context.outputData[crossEntropyTarget]);
if (float.IsInfinity(lossAvg))
{
lossAvg = 1e8f;
}
}
}
propState.loss = lossAvg;
propState.derivativeMemory.SwapBPBuffers();
int k = lid;
while (k-- > 0)
{
int l = hiddenLayers[k].numberOfNeurons;
while (l-- > 0)
{
backpropagate(k, l, propState.state[k][l], propState);
}
}
}
//execute neural network and save all calculation results in fullContext for adagrad
/// <summary>
/// Execute neural network and save all calculation results in fullContext for adagrad
/// </summary>
/// <param name="input"></param>
/// <param name="context"></param>
/// <param name="fullContext"></param>
public void Execute_FullContext(NeuralNetworkContext context, NeuralNetworkFullContext fullContext)
{
float[] input = context.inputData,
output = context.outputData,
hidden = context.hiddenData;
float[][] hiddenRecurring = context.hiddenRecurringData;
int i, weightIndex, recurringWeightIndex;
NeuronActivationFunction activeFunc;
if (hiddenLayers.Length > 0)
{
int lastNumNeurons = 0;
float[] weights, biases, recurringWeights;
for (i = 0; i < hiddenLayers.Length; i++)
{
weights = hiddenConnections[i].weights;
biases = hiddenLayers[i].biases;
activeFunc = hiddenLayers[i].activationFunction;
float[] ina;
int alen;
if (i == 0)
{
ina = input;
alen = input.Length;
}
else
{
ina = hidden;
alen = lastNumNeurons;
}
if (hiddenLayers[i].recurring)
{
//recurring hidden layer
float[] hrec = hiddenRecurring[i];
recurringWeights = hiddenRecurringConnections[i].weights;
//copy over data needed for training
Array.Copy(hrec, fullContext.hiddenRecurringBuffer[i], hrec.Length);
weightIndex = 0;
recurringWeightIndex = 0;
int k = biases.Length;
while (k-- > 0)
{
float ov = biases[k];
int j = alen;
while (j-- > 0)
{
ov += ina[j] * weights[weightIndex++];
}
j = hrec.Length;
while (j-- > 0)
{
ov += hrec[j] * recurringWeights[recurringWeightIndex++];
}
hidden[k] = activeFunc(ov);
}
Array.Copy(hidden, hrec, biases.Length);
}
else
{
//non recurring hidden layer
weightIndex = 0;
int k = biases.Length;
while (k-- > 0)
{
float ov = biases[k];
int j = alen;
while (j-- > 0)
{
ov += ina[j] * weights[weightIndex++];
}
hidden[k] = activeFunc(ov);
}
}
Array.Copy(hidden, fullContext.hiddenBuffer[i], biases.Length);
lastNumNeurons = biases.Length;
}
activeFunc = outputLayer.activationFunction;
//last output layer
//run input to output layer connection
weights = outputConnection.weights;
biases = outputLayer.biases;
weightIndex = 0;
recurringWeightIndex = 0;
i = output.Length;
while (i-- > 0)
{
float ov = biases[i];
//input connections
int k = lastNumNeurons;
while (k-- > 0)
{
ov += hidden[k] * weights[weightIndex++];
}
output[i] = activeFunc(ov);
}
}
else
{
activeFunc = outputLayer.activationFunction;
//run input to output layer connection with recurring output
float[] weights = outputConnection.weights,
biases = outputLayer.biases;
weightIndex = 0;
recurringWeightIndex = 0;
i = output.Length;
while (i-- > 0)
{
float ov = biases[i];
//input connections
int k = input.Length;
while (k-- > 0)
{
ov += input[k] * weights[weightIndex++];
}
output[i] = activeFunc(ov);
}
}
}