Hum.... I'm not sure we want to disable this behavior altogether, because in the case where the indexing slice is in the first dimension, it kinda works as is now (returns an empty tensor).

A better fix I think would be to check if the slice size is 0, and returns an empty tensor straight away?

For example, I think we should have the same behavior for the moment as in the following:

a = torch.rand(3, 3)
print(a[[]])
print(a[:, []])

Both prints an empty tensor, and I think this is the expected behavior for now.

I haven't thought through the implications of @fmassa's comment, but it seems consistent with how we have handled 0-sized things in the past: all sizes that would include a 0 if we supported it are collapsed to (0,).

One issue with returning a 0 index tensor is when it's not the last index you specify. For example:

i = slice(0)
x = torch.Tensor(range(8)).view(2, 2, 2)
x[i, 0]
# RuntimeError: dimension out of range (expected to be in range of [-1, 0], but got 1

Can't we short-circuit indexing if any of the indexing elements is a slice of size zero?

Its quite annoying to have checks everywhere in the code to skip some operations if the result of the operation is an empty tensor, and with the current setup, many things work already.

@pytorchbot retest this please

@li-roy Don't forget about this PR!

@pytorchbot retest this please

@pytorchbot retest this please

@pytorchbot retest this please