[JIT] Unannotated NamedTuple Type should use TensorType::getInferred()

[eellison]

🐛 Bug / 🚀 Feature

import torch
from collections import namedtuple
_MyNamedTuple = namedtuple('_MyNamedTuple', ['value'])

@torch.jit.script
def foo():
    return _MyNamedTuple(1)
> Expected a value of type 'Tensor' for argument 'value' but instead found type 'int'.

It would be nice if we showed that the Tensor type was inferred because no type annotations were given for the named tuple. There was a user internally who was confused by this.

cc @gmagogsfm

[Lilyjjo]

Implementation pointers:

The error message is generated in torch/csrc/jit/frontend/schema_matching.cpp’s tryMatchArgument(...). There are multiple ways of solving this issue, so feel free to take a different approach. One way would be to get is_inferred_type() on the value’s type in this function to return true. The return value of is_inferred_type() is determined by /aten/src/ATen/core/type.cpp’s TensorType::create(...) that sets the is_inferred_ value. You would need to track down where in the code NamedTuple calls this TensorType::create(...) and set the is_inferred bool to be true. If this isn’t feasible, then generating the error message when the type refinement of the unannotated variable for the NamedTuple could also work.

The code that deals with the NamedTuple can be found by searching for 'NamedTuple' in:
torch/csrc/jit/python/python_sugared_value.cpp
torch/csrc/jit/frontend/sugared_value.cpp
torch/csrc/jit/frontend/ir_emitter.cpp

[Lilyjjo]

For reference, this is a way to give the needed type annotations to resolve the error:

import torch
from typing import NamedTuple
_MyNamedTuple = NamedTuple('_MyNamedTuple', [('value', int)])

@torch.jit.script
def foo():
    return _MyNamedTuple(1)