private void UpdateLayer(int PWhichLayer, bool PForward, AutoencoderWeights PWeights)
{
int beginlayer = PWhichLayer - 1;
if (PForward)
{
Utility.WithinBounds("Cannot update this layer!!!", PWhichLayer, 1, numlayers);
}
else
{
Utility.WithinBounds("Cannot update this layer!!!", PWhichLayer, 0, numlayers - 1);
beginlayer = PWhichLayer + 1;
}
RBMLayer thislayer = layers[PWhichLayer];
RBMLayer previouslayer = layers[beginlayer];
double input = 0;
double[] states = previouslayer.GetStates();
for (int i = 0; i < thislayer.Count; i++)
{
for (int j = 0; j < previouslayer.Count; j++)
{
if (!PForward)
{
input += PWeights.GetWeightSet(beginlayer - 1).GetWeight(i, j) * states[j];
}
else
{
input += PWeights.GetWeightSet(beginlayer).GetWeight(j, i) * states[j];
}
}
thislayer.SetState(i, input);
input = 0;
}
}
private static void TrainBias(RBMLayer PLayer, int PWhich, double PPosPhase, double PNegPhase)
{
double biaschange = (learnrate.mombiases * PLayer.GetBiasChange(PWhich)) + (learnrate.lrbiases * (PPosPhase - PNegPhase));
PLayer.SetBiasChange(PWhich, biaschange);
PLayer.SetBias(PWhich, PLayer.GetBias(PWhich) + biaschange);
}
public RBM(RBMLayer PVisibles, RBMLayer PHiddens, RBMLearningRate PLearnRate, IWeightInitializer PWeightInit)
{
numvisibles = PVisibles.Count;
numhiddens = PHiddens.Count;
InitLayers(PVisibles, PHiddens);
InitWeights(PWeightInit);
InitTrainingData();
learnrate = PLearnRate;
}
public RBM(RBM PA)
{
visibles = (RBMLayer)PA.visibles.Clone();
hiddens = (RBMLayer)PA.hiddens.Clone();
weights = (RBMWeightSet)PA.weights.Clone();
learnrate = PA.learnrate;
trainingdata = PA.trainingdata;
numvisibles = PA.numvisibles;
numhiddens = PA.numhiddens;
}
private void AddLayer(RBMLayer PLayer)
{
learnrate.prelrbiases.Add(0.001);
learnrate.premombiases.Add(0.5);
learnrate.finelrbiases.Add(0.001);
if (layers.Count >= 1)
{
learnrate.prelrweights.Add(0.001);
learnrate.premomweights.Add(0.5);
learnrate.finelrweights.Add(0.001);
}
layers.Add(PLayer);
}
private void UpdateLayerBackwardPreTrain(int PWhich)
{
Utility.WithinBounds("Cannot update this layer!!!", PWhich, numlayers - 1);
RBMLayer thislayer = layers[PWhich];
RBMLayer nextlayer = layers[PWhich + 1];
double input = 0;
double[] states = nextlayer.GetStates();
for (int i = 0; i < thislayer.Count; i++)
{
for (int j = 0; j < nextlayer.Count; j++)
{
input += recognitionweights.GetWeightSet(PWhich).GetWeight(i, j) * states[j];
}
thislayer.SetState(i, input);
input = 0;
}
}
public static void Train(RBMLayer PLayerVis, RBMLayer PLayerHid, TrainingData PData
, RBMLearningRate PLearnRate, RBMWeightSet PWeightSet)
{
weightset = PWeightSet;
learnrate = PLearnRate;
for (int i = 0; i < PLayerVis.Count; i++)
{
for (int j = 0; j < PLayerHid.Count; j++)
{
TrainWeight(i, j, CalculateTrainAmount(PData.posvis[i], PData.poshid[j]
, PData.negvis[i], PData.neghid[j]));
}
TrainBias(PLayerVis, i, PData.posvis[i], PData.negvis[i]);
}
for (int j = 0; j < PLayerHid.Count; j++)
{
TrainBias(PLayerHid, j, PData.poshid[j], PData.neghid[j]);
}
}
private void InitLayers(RBMLayer PVisibles, RBMLayer PHiddens)
{
if (PVisibles == null)
{
throw new Exception("You need a visible layer...");
}
if (PHiddens == null)
{
throw new Exception("You need a hidden layer...");
}
if (PVisibles.Count <= 0)
{
throw new Exception("You need at least one visible neuron...");
}
if (PHiddens.Count <= 0)
{
throw new Exception("You need at least one hidden neuron...");
}
hiddens = PHiddens;
visibles = PVisibles;
}
