Reverse checkerboard pattern

1

I have a 2d-matrix and I want to split it using a checkerboard pattern. example:

1,  2,  3,  4
5,  6,  7,  8
9, 10, 11, 12

into

1,  3,
5,  7,
9, 11,

and

2,   4
6,   8
10, 12

this should be: a = x[:, :, 0::2, :] and b = x[:, :, 1::2, :]

but now if I have a and b, how can I revert the transformation?

2

Depending on your style preferences, there are

x = torch.empty(a.size(0), a.size(1), 2*a.size(2), a.size(3), dtype=a.dtype, device=a.device)
x[:, :, 0::2, :] = a
x[:, :, 1::2, :] = b

and

x = torch.stack([a, b], 3).view(a.size(0), a.size(1), 2*a.size(2), a.size(3))

you could try.
The latter worked better with JIT a while ago when I looked at tracing maskrcnn-benchmark.

Best regards

Thomas

P.S.: To me it looks more like a striped than a checkerboard pattern.

3

thanks for the quick answer!
ah, yeah. You’re right, that’s not a checkboard-pattern. But I think I can adapt it to a real checkerboard-pattern.