public void InterConnectionWeightRenewal(double learningRate, double decayRate)
{
InterConnectionMatrix += LayerActivationMatrix.Transpose() * LayerErrorMatrix * learningRate;
Parallel.For(0, InterConnectionMatrix.RowCount, index =>
{
InterConnectionMatrix[index, index] = 0;
});
if (decayRate > 0)
{
Parallel.For(0, InterConnectionMatrix.RowCount, rowIndex =>
{
for (int columnIndex = 0; columnIndex < InterConnectionMatrix.ColumnCount; columnIndex++)
{
if (InterConnectionMatrix[rowIndex, columnIndex] > 0)
{
InterConnectionMatrix[rowIndex, columnIndex] -= decayRate;
}
else if (InterConnectionMatrix[rowIndex, columnIndex] < 0)
{
InterConnectionMatrix[rowIndex, columnIndex] += decayRate;
}
}
});
InterConnectionMatrix.CoerceZero(decayRate);
}
}
public virtual void Duplicate(Layer cloneLayer)
{
cloneLayer.LayerStroageMatrix = LayerStroageMatrix.Clone();
cloneLayer.LayerActivationMatrix = LayerActivationMatrix.Clone();
cloneLayer.LayerErrorMatrix = LayerErrorMatrix.Clone();
cloneLayer.BiasMatrix = BiasMatrix.Clone();
cloneLayer.InterConnectionMatrix = InterConnectionMatrix.Clone();
}
public void CleanUpBackwordProcess(double momentum, double learningRate, double decayRate)
{
CleanupLayerErrorMatrix = (LayerErrorMatrix * CleanupToLayerWeightMatrix.Transpose())
.PointwiseMultiply(CleanupLayerActivationMatrix.PointwiseMultiply(1 - CleanupLayerActivationMatrix));
CleanupBiasMatrix += CleanupLayerErrorMatrix.ColumnSums().ToRowMatrix() * learningRate;
CleanupToLayerWeightMatrix += CleanupLayerActivationMatrix.Transpose() * LayerErrorMatrix * learningRate;
LayerToCleanupWeightMatrix += LayerActivationMatrix.Transpose() * CleanupLayerErrorMatrix * learningRate;
if (decayRate > 0)
{
Parallel.For(0, CleanupBiasMatrix.ColumnCount, columnIndex =>
{
if (CleanupBiasMatrix[0, columnIndex] > 0)
{
CleanupBiasMatrix[0, columnIndex] -= decayRate;
}
else if (CleanupBiasMatrix[0, columnIndex] < 0)
{
CleanupBiasMatrix[0, columnIndex] += decayRate;
}
});
CleanupBiasMatrix.CoerceZero(decayRate);
Parallel.For(0, CleanupToLayerWeightMatrix.RowCount, rowIndex =>
{
for (int columnIndex = 0; columnIndex < CleanupToLayerWeightMatrix.ColumnCount; columnIndex++)
{
if (CleanupToLayerWeightMatrix[rowIndex, columnIndex] > 0)
{
CleanupToLayerWeightMatrix[rowIndex, columnIndex] -= decayRate;
}
else if (CleanupToLayerWeightMatrix[rowIndex, columnIndex] < 0)
{
CleanupToLayerWeightMatrix[rowIndex, columnIndex] += decayRate;
}
}
});
CleanupToLayerWeightMatrix.CoerceZero(decayRate);
Parallel.For(0, LayerToCleanupWeightMatrix.RowCount, rowIndex =>
{
for (int columnIndex = 0; columnIndex < LayerToCleanupWeightMatrix.ColumnCount; columnIndex++)
{
if (LayerToCleanupWeightMatrix[rowIndex, columnIndex] > 0)
{
LayerToCleanupWeightMatrix[rowIndex, columnIndex] -= decayRate;
}
else if (LayerToCleanupWeightMatrix[rowIndex, columnIndex] < 0)
{
LayerToCleanupWeightMatrix[rowIndex, columnIndex] += decayRate;
}
}
});
LayerToCleanupWeightMatrix.CoerceZero(decayRate);
}
}
public void Test(Matrix <double> targetPattern, double activationCriterion, double inactivationCriterion, out List <double> meanSEList, out List <double> meanSSList, out List <double> meanCEList, out List <bool> correctnessList, out List <double> meanActivationList, out List <double> meanAUActivationList, out List <double> meanIUActivationList)
{
meanSEList = ((targetPattern - LayerActivationMatrix).PointwisePower(2).RowSums() / 2.0 / UnitCount).ToList();
meanSSList = ((((LayerActivationMatrix.PointwiseMultiply(LayerActivationMatrix.PointwiseLog()) + ((1 - LayerActivationMatrix).PointwiseMultiply((1 - LayerActivationMatrix).PointwiseLog()))) / Math.Log(2)) + 1).RowSums() / UnitCount).ToList();
meanCEList = (((targetPattern.PointwiseMultiply(LayerActivationMatrix.PointwiseLog()) + (1 - targetPattern).PointwiseMultiply((1 - LayerActivationMatrix).PointwiseLog())) / Math.Log(Math.E)).RowSums() * -1 / UnitCount).ToList();
meanActivationList = (LayerActivationMatrix.RowSums() / UnitCount).ToList();
correctnessList = new List <bool>(LayerActivationMatrix.RowCount);
meanAUActivationList = new List <double>(LayerActivationMatrix.RowCount);
meanIUActivationList = new List <double>(LayerActivationMatrix.RowCount);
for (int index = 0; index < LayerActivationMatrix.RowCount; index++)
{
correctnessList.Add(true);
meanAUActivationList.Add(0);
meanIUActivationList.Add(0);
}
Matrix <double> activeUnitMatrix = DenseMatrix.Create(LayerActivationMatrix.RowCount, LayerActivationMatrix.ColumnCount, 0);
Matrix <double> inactiveUnitMatrix = DenseMatrix.Create(LayerActivationMatrix.RowCount, LayerActivationMatrix.ColumnCount, 0);
Matrix <double> countActive = DenseMatrix.Create(LayerActivationMatrix.RowCount, 1, 0);
Matrix <double> countInactive = DenseMatrix.Create(LayerActivationMatrix.RowCount, 1, 0);
for (int rowIndex = 0; rowIndex < LayerActivationMatrix.RowCount; rowIndex++)
{
for (int columnIndex = 0; columnIndex < LayerActivationMatrix.ColumnCount; columnIndex++)
{
if (targetPattern[rowIndex, columnIndex] > activationCriterion)
{
if (LayerActivationMatrix[rowIndex, columnIndex] < activationCriterion)
{
correctnessList[rowIndex] = false;
}
activeUnitMatrix[rowIndex, columnIndex] = 1;
countActive[rowIndex, 0]++;
}
if (targetPattern[rowIndex, columnIndex] < inactivationCriterion)
{
if (LayerActivationMatrix[rowIndex, columnIndex] > activationCriterion)
{
correctnessList[rowIndex] = false;
}
inactiveUnitMatrix[rowIndex, columnIndex] = 1;
countInactive[rowIndex, 0]++;
}
}
}
meanAUActivationList = LayerActivationMatrix.PointwiseMultiply(activeUnitMatrix).RowSums().ToList();
meanIUActivationList = LayerActivationMatrix.PointwiseMultiply(inactiveUnitMatrix).RowSums().ToList();
for (int rowIndex = 0; rowIndex < LayerActivationMatrix.RowCount; rowIndex++)
{
if (countActive[rowIndex, 0] > 0)
{
meanAUActivationList[rowIndex] = meanAUActivationList[rowIndex] / countActive[rowIndex, 0];
}
else
{
meanAUActivationList[rowIndex] = double.NaN;
}
if (countInactive[rowIndex, 0] > 0)
{
meanIUActivationList[rowIndex] = meanIUActivationList[rowIndex] / countInactive[rowIndex, 0];
}
else
{
meanIUActivationList[rowIndex] = double.NaN;
}
}
}