Indexing with LongTensor is inconsistent with that of BoolTensor

1

As of today, we can use long, byte or bool tensor as indices. but indexing with byte tensor is deprecated.
But indexing with longtensor doesn’t give same answer as indexing with booltensor.

x = torch.rand(3, 3)
>>> x
tensor([[0.0857, 0.9867, 0.0330],
        [0.9295, 0.9787, 0.4835],
        [0.9376, 0.3851, 0.8934]])
y = torch.tensor([[1, 1, 1],
        [0, 0, 0],
        [0, 0, 0]])
>>> y
tensor([[1, 1, 1],
        [0, 0, 0],
        [0, 0, 0]])

indexing with booltensor gives

>>> x[y.bool()]
tensor([0.0857, 0.9867, 0.0330])

indexing with longtensor gives

>>> x[y]
tensor([[[0.9295, 0.9787, 0.4835],
         [0.9295, 0.9787, 0.4835],
         [0.9295, 0.9787, 0.4835]],

        [[0.0857, 0.9867, 0.0330],
         [0.0857, 0.9867, 0.0330],
         [0.0857, 0.9867, 0.0330]],

        [[0.0857, 0.9867, 0.0330],
         [0.0857, 0.9867, 0.0330],
         [0.0857, 0.9867, 0.0330]]])

Is this an expected behavior? Is there something am i doing wrong?
I want indexing with longtensor to give same result as indexing with booltensor.

Thank you in advance!

2

Yes, this is expected behavior and e.g. explained in numpy - integer array indexing and numpy - boolean array indexing.

1 Like