Python2 TorchScript type annotation tuple vs Tuple

jit
1

I have a torch script that accepts a tuple[Tensor],

@torch.jit.script
def smt_pred(confs, child, child_sizes, threshold, obj,
             height, width, a):
    # type: (Tuple[Tensor],Tuple[Tensor],Tuple[Tensor],Tensor,Tensor,Tensor,Tensor,Tensor) -> Tuple[Tensor,Tensor]

When I call the code in Python2 I get the error

RuntimeError: smt_pred() Expected a value of type ‘Tuple[Tensor]’ for argument ‘confs’ but instead found type ‘tuple’.

tuple is same as Tuple and I could not find how to annotate confs to be tuple of tensors. Is it possible in Python2?

Changing Tuple to List works in Python2, so this is only about Tuple

2

Added an issue here because List works (and I just have to pass list(confs) as a workaround)

3

Copied from the issue:

I’m guessing that you’re passing in a tuple that doesn’t look like (torch.ones(2, 2),). For a Tuple the type needs to be fixed and completely specified and the elements can be different types. In TorchScript, a, b, and c below are all different types

# type is `Tuple[int, int, int]
a = (1, 2, 3)

# type is `Tuple[str, int, int]
b = ('hi', 2, 3)

# type is `Tuple[int]`
c = (2,)

For lists, the element types must all be the same, which is why it can be any length but you only need to specify List[Tensor] or List[int]. You can read more about it here.