Support for using a negative dimension in torch.stack currently works as follows:

if dim < 0:
    dim += sequence[0].dim()

This is not consistent with the behaviour of torch.unsqueeze, which uses dim += input.dim() + 1. The unsqueeze behaviour is better since it is possible to use negative indexing to insert a new dimension as the last dimension.

Example

torch.stack with dim=-1 adds a new dimension in the second-to-last position. This is confusing. It is not possible to use negative indexing to add a dimension in the last position (ie can't achieve a 3x2 result).

>>> a = torch.Tensor([1, 2, 3])
>>> b = torch.Tensor([4, 5, 6])
>>> torch.stack([a, b], -1)

 1  2  3
 4  5  6
[torch.FloatTensor of size 2x3]

Contrast this to using torch.unsqueeze and torch.cat with dim=-1, which adds a new dimension in the last position:

>>> a = torch.Tensor([1, 2, 3])
>>> b = torch.Tensor([4, 5, 6])
>>> torch.cat([a.unsqueeze(-1), b.unsqueeze(-1)], -1)

 1  4
 2  5
 3  6
[torch.FloatTensor of size 3x2]