I have a tensor x, that looks like this:
x = tensor([ 1, 2, 3, 4, 5],
[ 6, 7, 8, 9, 10]
[ 11, 12, 13, 14, 15])
I'm trying to switch the first two and last two numbers of each tensor, like this:
x = tensor([ 4, 5, 3, 1, 2],
[ 9, 10, 8, 6, 7],
[ 14, 15, 13, 11, 12])
How could I do this with torch.roll()? How would I switch 3 instead of 1?
Not sure if that can be done with torch.roll alone... However, you can expect the desired result by using a temporary tensor and a pair assignment:
>>> x = torch.arange(1, 16).reshape(3,-1)
tensor([[ 1, 2, 3, 4, 5],
[ 6, 7, 8, 9, 10],
[11, 12, 13, 14, 15]])
>>> tmp = x.clone()
# swap the two sets of columns
>>> x[:,:2], x[:,-2:] = tmp[:,-2:], tmp[:,:2]
Such that tensor x has been mutated as:
>>> x
tensor([[ 4, 5, 3, 1, 2],
[ 9, 10, 8, 6, 7],
[14, 15, 13, 11, 12]])
You can pull off this operation with torch.roll and some indexing:
>>> x = torch.arange(1, 21).reshape(4,-1)
tensor([[ 1, 2, 3, 4, 5],
[ 6, 7, 8, 9, 10],
[11, 12, 13, 14, 15],
[16, 17, 18, 19, 20]])
>>> rolled = x.roll(-2,0)
tensor([[11, 12, 13, 14, 15],
[16, 17, 18, 19, 20],
[ 1, 2, 3, 4, 5],
[ 6, 7, 8, 9, 10]])
# overwrite columns [1,-1[ from rolled with those from x
>>> rolled[:, 1:-1] = x[:, 1:-1]
Such that at this end you get:
>>> rolled
tensor([[11, 2, 3, 4, 15],
[16, 7, 8, 9, 20],
[ 1, 12, 13, 14, 5],
[ 6, 17, 18, 19, 10]])
torch.roll this time around. Please look at my edit above for a full walkthrough.