_convolution_double_backward() - What does it do?

Can someone explain to me what that function _convolution_double_backward() does? And when it is called?
It is located in Convolution.cpp

1 Like


As you can see here, it defines the backward of the backward function for convolution.
It is used if you need higher order derivatives through a conv layer.
The autograd could build it automatically from the backward implementation of conv but this custom function is much faster.