Typed NamedTuple with member function generating another instance of class fails at runtime with NameError

[lwolfsonkin]

I simplified the code just for proof of concept:

class Hypothesis(NamedTuple):
    logprob: float
    def attach(self) -> Hypothesis:
        return Hypothesis(logprob=4)

While mypy doesn't complain at this, the python interpreter crashes with a NameError saying:

Traceback (most recent call last):
  File "./decode.py", line 31, in <module>
    class Hypothesis(NamedTuple):
  File "./decode.py", line 34, in Hypothesis
    def attach(self) -> Hypothesis:
NameError: name 'Hypothesis' is not defined

Unfortunately, I dropped the return type annotation as a result (which mypy didn't complain about) in order to make it runnable.

[emmatyping]

This is because the name Hyposthesis is not defined inside its own class definition body. Mypy accepts this due to supporting this in stubs being needed. You can change the return annotation to be -> 'Hypothesis' and it will both work at runtime and be understood by mypy.

[lwolfsonkin]

Thanks for the tip! Interesting. So, the name Hypothesis is not defined in the class body, but is defined in methods within that class body?

[emmatyping]

Methods are part of the class body. So this also will not work

class Example:
    def test(a: Example):
        pass

nor this

class Example:
    def test(a) -> Example:
        pass

However, the same use of quotes can be used.

[lwolfsonkin]

huh. Cool. Definitely good to know. Thanks a lot @ethanhs!

[gvanrossum]

I do think there's a mypy bug here -- outside stubs, this should be rejected.

[JukkaL]

I do think there's a mypy bug here -- outside stubs, this should be rejected.

Yes, and we already have a mypy issue about this (python/mypy#948).

[ilevkivskyi]

@gvanrossum

I do think there's a mypy bug here -- outside stubs, this should be rejected.

...or PEP 563 can be accepted :-)