PyTorch - Functions

• torch.nn.functional.pad — PyTorch 2.6 documentation
  • confusing because
    1. the tuples passed to the pad argument are flat even though you’re specifying basically pairs for each dimension
    2. if you pad more than one dimension, your padding integers from pad are applied to dimensions in a back-to-front order, where the first pair are the left and right pads for the final dimension which are nominally the “columns” (and then the subsequent pairs pad the “rows”, then “batch”, then idk what you’d call it)