////////////////////////////////////////////////////////////////////////////////////////////////////
/// <summary> Backward CPU. </summary>
///
/// <param name="y"> A NdArray to process. </param>
/// <param name="x"> A NdArray to process. </param>
////////////////////////////////////////////////////////////////////////////////////////////////////
private void BackwardCpu([CanBeNull] NdArray y, [CanBeNull] NdArray x)
{
Beta.ClearGrad();
Gamma.ClearGrad();
for (int i = 0; i < ChannelSize; i++)
{
for (int j = 0; j < y.BatchCount; j++)
{
Beta.Grad[i] += y.Grad[i + j * y.Length];
Gamma.Grad[i] += y.Grad[i + j * y.Length] * Xhat[j * ChannelSize + i];
}
}
if (Verbose)
{
RILogManager.Default?.ViewerSendWatch("Learning", (IsTrain ? "Yes" : "No"));
}
if (IsTrain)
{
// with learning
int m = y.BatchCount;
for (int i = 0; i < ChannelSize; i++)
{
Real gs = Gamma.Data[i] / Std[i];
for (int j = 0; j < y.BatchCount; j++)
{
Real val = (Xhat[j * ChannelSize + i] * Gamma.Grad[i] + Beta.Grad[i]) / m;
x.Grad[i + j * ChannelSize] += gs * (y.Grad[i + j * y.Length] - val);
}
}
}
else
{
// No learning
for (int i = 0; i < ChannelSize; i++)
{
Real gs = Gamma.Data[i] / Std[i];
AvgMean.Grad[i] = -gs * Beta.Grad[i];
AvgVar.Grad[i] = -0.5 * Gamma.Data[i] / AvgVar.Data[i] * Gamma.Grad[i];
for (int j = 0; j < y.BatchCount; j++)
{
x.Grad[i + j * ChannelSize] += gs * y.Grad[i + j * y.Length];
}
}
}
}
