[TorchScript] fail to convert list to tuple

jit
1

Hi, in my torch.jit.script I would like to convert a list to a tuple, but it seems not to be the easy way.

The size of the list is fixed - although not in a way that jit prefers: which seems to be the problem
On Python 3.8 with PyTorch 1.9.1

result = []
val = torch.full((1,1),1.)
result.append(val) # size() = 1
return tuple(result)

… fails with Runtimeerror: cannot statically infer the expected size of a list in this context:

same as

val = torch.full((1,1),1.)
result:List[Tensor] = val * 1 
return tuple(result)

…fails with Runtimeerror: cannot statically infer the expected size of a list in this context:

Can someone please tell me how to do this in torch.jit.script (not trace!)
Thx

2

as error suggests, tuple sizes are static(hardcoded) when “compiled”, so you have to do something like
return (result[0],result[1])
and there may be a need to annotate this as Tuple[Tensor,Tensor] in some contexts

3

Ok, that’s meant by “statically inferred”. Coming from the C++ side, I would have expected some counterpart of C++

result[4] = {val}
retval = tuple(result)

like
result:List[Tensor,4] = val * 4

Thank you for your help

4

if you’re comparing to C++, List is like a vector<Tensor> of dynamic size, hence you cannot initialize specialized N-tuples from it, without runtime checks and metaprogramming anyway.
Though I agree that TorchScript is lacking in some aspects, in this case the workaround is to return a list.

5

this still fails for me:

from torch import Tensor
import torch

def some_rotation_op(rotation:Tensor):
  [Rxz, Ryz, Rzz] = rotation[:, -1]  # <------ whether this is a list makes no difference
  ...
  return Rxz

torch.jit.script(some_rotation_op, (torch.rand( 3, 3),))