What doesn't pytorch support negative slicing?

Fei_Liu · May 10, 2023, 10:20pm

e.g. x[::-1] is not allowed.

ValueError: step must be greater than zero

Are there any special design reasons from the beginning that supporting this brings in some technical difficulties? There are also some issue trackers that go back and forth for years on this question, e.g.:

github.com/pytorch/pytorch

Support negative step sizes for slicing

opened 11:44AM - 10 Jun 21 UTC

pmeier

triaged module: python array api

The [array API specification stipulates](https://data-apis.org/array-api/latest/…API_specification/indexing.html#slice-syntax) that the slicing syntax needs to be equivalent to Pythons [`slice()`](https://docs.python.org/3/library/functions.html#slice) syntax. That includes negative step sizes that PyTorch does not support: ```python import torch t = torch.rand((3,)) t[::-1] ``` ``` ValueError: step must be greater than zero ``` ## Additional context This was already proposed in #229, but finally dismissed in https://github.com/pytorch/pytorch/issues/229#issuecomment-548092342 in favor of `torch.flip()`. Note that this is only a substitution for a negative step size of 1 (`::-1`) rather than negative step sizes in general. cc @mruberry @rgommers @pmeier

torch.flip is for sure one work around, but it has an extra copy that are not necessary most of the times. The main question is, what stops us from the negative slicing? It seems so common (and natural) in other numerical packages (numpy and many other things) and I cannot really come up with any valid reason on skipping the support in pytorch.