protected override void BackPropInternal(Tensor outputGradient)
{
Tensor.Conv2DInputsGradient(outputGradient, Kernels, Stride, InputsGradient[0]);
Tensor.Conv2DKernelsGradient(Inputs[0], outputGradient, Stride, Tensor.PaddingType.Valid, KernelsGradient);
if (UseBias)
{
BiasGradient.Add(outputGradient.SumBatches());
}
}
protected override void BackPropInternal(Tensor outputGradient)
{
// for explanation watch https://www.youtube.com/watch?v=8H2ODPNxEgA&t=898s
// each input is responsible for the output error proportionally to weights it is multiplied by
Weights.Transposed().Mul(outputGradient, InputsGradient[0]);
WeightsGradient.Add(outputGradient.Mul(Inputs[0].Transposed()).SumBatches(), WeightsGradient);
if (UseBias)
{
BiasGradient.Add(outputGradient.SumBatches(), BiasGradient);
}
}
//计算Weights和Bias的梯度
private void ComputeGradient(TensorOld error)
{
var inputData = ForwardInput.GetRawValues();
var errorData = error.GetRawValues();
var features = WeightsGradient.shape[0];
Parallel.For(0, features, i =>
{
Parallel.For(0, UnitCount, j =>
{
var weightSum = 0d;
var biasSum = 0d;
for (int k = 0; k < sampleStartIndex.Length; k++)
{
weightSum += inputData[sampleStartIndex[k] + i] * errorData[errorStartIndex[k] + j];
biasSum += errorData[errorStartIndex[k] + j];
}
WeightsGradient[i, j] = weightSum;
BiasGradient.GetRawValues()[j] = biasSum;
});
});
}
