protected LayerBackwardCompute(LayerForwardCompute forwardCompute, ConnectedLayer connectedLayer)
{
Contract.Requires(forwardCompute != null);
Contract.Requires(connectedLayer != null);
Contract.Requires((connectedLayer.StructuralElementFlags & NNStructuralElement.BackwardImplementation) != 0 &&
(connectedLayer.StructuralElementFlags & NNStructuralElement.GradientInformation) != 0);
Init(forwardCompute, connectedLayer);
}
private void Init(LayerForwardCompute forwardCompute, ConnectedLayer connectedLayer)
{
ForwardCompute = forwardCompute;
InitLowerErrorValueAccessItems(connectedLayer);
ErrorBuffer = connectedLayer.ErrorBuffer;
GradientBuffers = connectedLayer.GradientBuffers;
GradientSumBuffers = connectedLayer.GradientSumBuffers;
BiasGradientValueIndex = connectedLayer.BiasGradientValueIndex;
BiasGradientSumValueIndex = connectedLayer.BiasGradientSumValueIndex;
OutputErrorBuffer = connectedLayer.OutputErrorBuffer;
}
private unsafe void SetGradientValue(float* valueBuffer, LayerForwardCompute comp, int iValueIndex, int jLayerIndex, int jValueIndex, float gradient)
{
if (jLayerIndex == -1)
{
// Bias
Debug.Assert(jValueIndex == -1);
Debug.Assert(comp.BiasGradientValueIndex != null);
valueBuffer[comp.BiasGradientValueIndex.Value] = gradient;
valueBuffer[comp.BiasGradientSumValueIndex.Value] += gradient;
}
else
{
int wvIndex = WeightAccessor.GetWeightValueIndex(jValueIndex, iValueIndex, comp.OutputBuffer.Size);
valueBuffer[comp.GradientBuffers[jLayerIndex].MinValue + wvIndex] = gradient;
valueBuffer[comp.GradientSumBuffers[jLayerIndex].MinValue + wvIndex] += gradient;
}
}
private unsafe float GetInputValue(float* valueBuffer, LayerForwardCompute comp, int jLayerIndex, int jValueIndex)
{
if (jLayerIndex == -1)
{
// Bias
Debug.Assert(jValueIndex == -1);
Debug.Assert(comp.BiasValueIndex != null);
return valueBuffer[comp.BiasValueIndex.Value];
}
var accessItem = comp.InputValueAccessItems[jLayerIndex];
var valueIndex = accessItem.InputBufferBeginIndex + jValueIndex;
Debug.Assert(valueIndex <= accessItem.InputBufferBeginIndex + accessItem.InputSize);
return valueBuffer[valueIndex];
}
unsafe private void ComputePValues(float* valueBuffer, LayerForwardCompute comp, IntRange[] valueRelatedPBuffs, IntRange[] valueRelatedPrevPBuffs, int iLayerIndex, int iValueIndex, int jLayerIndex, int jValueIndex, float* eVector)
{
int outputLayerIndex = valueRelatedPBuffs.Length - 1;
Debug.Assert(outputLayerIndex == ForwardComputes.Length - 1);
Debug.Assert(outputLayerIndex == ForwardComputes[ForwardComputes.Length - 1].ConnectedLayerIndex);
for (int kLayerIndex = 0; kLayerIndex < valueRelatedPBuffs.Length; kLayerIndex++)
{
float gradient = 0.0f;
bool computeGradient = kLayerIndex == outputLayerIndex && eVector != null;
var currentPBuffValueRange = valueRelatedPBuffs[kLayerIndex];
var currentPrevPBuffValueRange = valueRelatedPrevPBuffs[kLayerIndex];
int currentPBuffValueRangeSize = currentPBuffValueRange.Size;
for (int kValueIndex = 0; kValueIndex < currentPBuffValueRangeSize; kValueIndex++)
{
// i: iLayerIndex, iValueIndex
// j: jLayerIndex, jValueIndex
// k: kLayerIndex, kValueIndex
float netDeriv_k = GetNetDerivValue(valueBuffer, kLayerIndex, kValueIndex);
float sum = 0.0f;
var comp_k = ForwardComputes[kLayerIndex];
foreach (var upperNonInputLayerInfo in comp_k.UpperNonInputLayerInfos)
{
int lLayerIndex = upperNonInputLayerInfo.LayerIndex;
var accessItem = comp_k.InputValueAccessItems[upperNonInputLayerInfo.WeightedErrorBufferIndex];
for (int lValueIndex = 0; lValueIndex < accessItem.InputSize; lValueIndex++)
{
// i: iLayerIndex, iValueIndex
// j: jLayerIndex, jValueIndex
// k: kLayerIndex, kValueIndex
// l: lLayerIndex, lValueIndex
float weight_k_l = valueBuffer[accessItem.WeightBufferBeginIndex + WeightAccessor.GetWeightValueIndex(lValueIndex, kValueIndex, currentPBuffValueRangeSize)];
float p_i_j_l = valueBuffer[valueRelatedPrevPBuffs[lLayerIndex].MinValue + lValueIndex];
sum += weight_k_l * p_i_j_l;
}
}
if (iLayerIndex == kLayerIndex && iValueIndex == kValueIndex)
{
sum += GetInputValue(valueBuffer, comp, jLayerIndex, jValueIndex);
}
if (computeGradient)
{
gradient += eVector[kValueIndex] * (valueBuffer[valueRelatedPBuffs[kLayerIndex].MinValue + kValueIndex] = netDeriv_k * sum);
}
else
{
valueBuffer[valueRelatedPBuffs[kLayerIndex].MinValue + kValueIndex] = netDeriv_k * sum;
}
}
if (computeGradient)
{
//Console.WriteLine(gradient);
SetGradientValue(valueBuffer, comp, iValueIndex, jLayerIndex, jValueIndex, gradient);
}
}
}