I have a 1D array:
[3, 4, 5, 6, 7, 20, 31, 32, 33, 34]
which I want to turn into a 2D interval array by merging every consecutive values into intervals:
[
[ 3, 7],
[20, 20],
[32, 34]
]
What would be a decent, possibly GPU-friendly, way to do it?
Not sure if this is ideal, but you can try cumsum() on the differences compared to 1. Then use that to slice the original data:
# mark the consecutive blocks
blocks = torch.cat([torch.tensor([0]),(t[1:] - t[:-1]) != 1]).cumsum(dim=0)
# where the blocks shift
mask = blocks[1:] != blocks[:-1]
out = torch.cat([t[:1], t[:-1][mask],t[1:][mask], t[-1:]]).reshape(-1,2)
Output:
tensor([[ 3, 7],
[20, 20],
[31, 34]])