same_padding

hybrid_learning.datasets.transforms.encoder.same_padding(kernel_size, hang_front=False)[source]

Calculate the left and right padding for mode 'same' for each dim and concat.

Mode 'same' here means Conv(kernel_size)(Pad(padding)(x)).size() == x.size().

Padding is distributed equally on both sides of a dimension. If unequal padding is needed in one dimension, by default (hang_front==False) the front gets padded by one pixel less than the back. To instead pad the front more, set hang_font==True.

Warning

Currently (Apr 2019), torch.nn.ZeroPad2d requires the padding in a special format: first width paddings, then height paddings. So the entries from kernel_size need to be swapped to obtain the correct padding as input for torch.nn.ZeroPad2d.

Parameters

  • kernel_size (Sequence[int]) – the list of kernel dimension sizes

  • hang_front (bool) – whether the front instead of the rear padding should be larger by 1 in case unequal padding is needed

Returns

padding tuple as (left dim0, right dim0, left dim1, right dim1, ...)

Return type

Tuple[int, …]