AvgPoolNdBackward

Autograd node for N-dimensional average pooling.

Each output element is the arithmetic mean of a $K_0 \times \cdots \times K_{N-1}$ window of the (optionally zero-padded) input. No activations need to be saved — the backward pass is a uniform scatter that distributes each output gradient over the $\prod_i K_i$ input elements of its window.