What I want to do is something like this:
import torch
a = torch.arange(120).reshape(2, 3, 4, 5)
b = torch.cat(list(a), dim=2)
I want to know:
You want to:
Reduce the number of copies: in this specific scenario, copies need to be made since we are rearranging the layout of our underlying data.
Reduce or remove any torch.Tensor -> non-torch.Tensor conversions: this will be a pain point when working with a GPU since you're transferring data in and out of the device.
You can perform the same operation by permuting the axes such that axis=0 goes to axis=-2 (the before the last axis), then flattening the last two axes:
>>> a.permute(1,2,0,3).flatten(-2)
tensor([[[ 0, 1, 2, 3, 4, 60, 61, 62, 63, 64],
[ 5, 6, 7, 8, 9, 65, 66, 67, 68, 69],
[ 10, 11, 12, 13, 14, 70, 71, 72, 73, 74],
[ 15, 16, 17, 18, 19, 75, 76, 77, 78, 79]],
[[ 20, 21, 22, 23, 24, 80, 81, 82, 83, 84],
[ 25, 26, 27, 28, 29, 85, 86, 87, 88, 89],
[ 30, 31, 32, 33, 34, 90, 91, 92, 93, 94],
[ 35, 36, 37, 38, 39, 95, 96, 97, 98, 99]],
[[ 40, 41, 42, 43, 44, 100, 101, 102, 103, 104],
[ 45, 46, 47, 48, 49, 105, 106, 107, 108, 109],
[ 50, 51, 52, 53, 54, 110, 111, 112, 113, 114],
[ 55, 56, 57, 58, 59, 115, 116, 117, 118, 119]]])