//Kernel initialization
public override void Init(int nGPU)
{
layers = new List <MyAbstractLayer>();
if (Owner.SortedChildren == null)
{
Owner.InitGroup.Execute(); // TODO - horrible hack
}
foreach (var i in Owner.SortedChildren)
{
if (i is MyRBMInputLayer || i is MyRBMLayer)
{
layers.Add((MyAbstractLayer)i);
if (i is MyRBMInputLayer)
{
((MyRBMInputLayer)i).Init(nGPU);
}
else
{
((MyRBMLayer)i).Init(nGPU);
}
}
}
InputLayer = (MyRBMInputLayer)layers[0];
step = 0;
}
//Kernel initialization
public override void Init(int nGPU)
{
layers = new List<MyAbstractLayer>();
if (Owner.SortedChildren == null)
Owner.InitGroup.Execute(); // TODO - horrible hack
foreach (var i in Owner.SortedChildren)
{
if (i is MyRBMInputLayer || i is MyRBMLayer)
{
layers.Add((MyAbstractLayer)i);
if (i is MyRBMInputLayer)
((MyRBMInputLayer)i).Init(nGPU);
else
((MyRBMLayer)i).Init(nGPU);
}
}
InputLayer = (MyRBMInputLayer)layers[0];
step = 0;
}
//Task execution
public override void Execute()
{
// sampling between input and hidden layer
if (CurrentLayerIndex <= 0)
{
InputLayer.RBMInputForwardAndStore();
MyRBMInputLayer visible = InputLayer;
hidden = (MyRBMLayer)layers[CurrentLayerIndex + 1];
if (hidden.Dropout > 0)
{
hidden.CreateDropoutMask();
}
// forward to visible and hidden layers and store for sampling biases
hidden.RBMForwardAndStore(SigmoidSteepness);
// sample positive weight data
hidden.RBMSamplePositive();
if (RandomHidden)
{
hidden.RBMRandomActivation();
}
// do k-step Contrastive Divergence (go back and forth between visible and hidden)
for (int i = 0; i < CD_k - 1; i++)
{
// back
hidden.RBMBackward(visible.Bias, SigmoidSteepness);
if (RandomVisible)
{
visible.RBMRandomActivation();
}
// and forth
hidden.RBMForward(SigmoidSteepness, true);
// randomly activate the just updated hidden neurons if needed
if (RandomHidden)
{
hidden.RBMRandomActivation();
}
}
// in the last (= k-th) step of CD-k, we use probabilistic modeling -> we sample both with probability, not random activation
hidden.RBMBackward(visible.Bias, SigmoidSteepness);
hidden.RBMForward(SigmoidSteepness, true);
if (hidden.StoreEnergy)
{
hidden.Energy.Fill(0);
}
// update biases of both layers based on the sampled data stored by RBMForwardAndStore()
visible.RBMUpdateBiases(LearningRate, Momentum, WeightDecay);
hidden.RBMUpdateBiases(LearningRate, Momentum, WeightDecay);
// sample negative weight data AND adapt weights at the same time
hidden.RBMUpdateWeights(LearningRate, Momentum, WeightDecay);
}
// sampling between hidden layers
else if (layers.Count > 2) // if number of hidden layers is greater than 1)
{
InputLayer.RBMInputForward();
// hidden layers, up to the one that will be visible (excluding)
for (int i = 1; i < CurrentLayerIndex; i++)
{
((MyRBMLayer)layers[i]).RBMForward(SigmoidSteepness, false);
if (RandomPrevious)
{
((MyRBMLayer)layers[i]).RBMRandomActivation();
}
}
MyRBMLayer visible = (MyRBMLayer)layers[CurrentLayerIndex];
hidden = (MyRBMLayer)layers[CurrentLayerIndex + 1];
if (hidden.Dropout > 0)
{
hidden.CreateDropoutMask();
}
// forward to visible and hidden layers and store for sampling biases
visible.RBMForwardAndStore(SigmoidSteepness);
hidden.RBMForwardAndStore(SigmoidSteepness);
// sample positive weight data
hidden.RBMSamplePositive();
if (RandomHidden)
{
hidden.RBMRandomActivation();
}
// do k-step Contrastive Divergence (go back and forth between visible and hidden)
for (int i = 0; i < CD_k - 1; i++)
{
// back
hidden.RBMBackward(visible.Bias, SigmoidSteepness);
if (RandomVisible)
{
visible.RBMRandomActivation();
}
// and forth
hidden.RBMForward(SigmoidSteepness, true);
// randomly activate the just updated hidden neurons if needed
if (RandomHidden)
{
hidden.RBMRandomActivation();
}
}
// in last (= k-th) step of CD-k, we use probabilistic modeling -> we sample both with probability, not random activation
hidden.RBMBackward(visible.Bias, SigmoidSteepness);
hidden.RBMForward(SigmoidSteepness, true);
if (hidden.StoreEnergy)
{
hidden.Energy.Fill(0);
}
// update biases of both layers based on the sampled data stored by RBMForwardAndStore()
visible.RBMUpdateBiases(LearningRate, Momentum, WeightDecay);
hidden.RBMUpdateBiases(LearningRate, Momentum, WeightDecay);
// sample negative weight data AND adapt weights at the same time
hidden.RBMUpdateWeights(LearningRate, Momentum, WeightDecay);
}
else
{
MyLog.ERROR.WriteLine("Wrong CurrentLayerIndex parameter. There are " + layers.Count + " total layers, can't sample from " + CurrentLayerIndex);
}
MyLog.DEBUG.WriteLine("RBM initialization between layers [" + CurrentLayerIndex + ";" + (CurrentLayerIndex + 1) + "], step " + step + ".");
++step;
}
