Registered functions with pydantic class arguments can't be parsed

[timothyjlaurent]

Consider this:

from pydantic import BaseModel

@registry.dim('v1')
class Dim(BaseModel):
   length: int

@registry.architecture('v1')
def load_architecture(foo: str, dim: Dim):
    ...

config str = """
[model]
@architecture = 'v1'
length = 6
[model.dim]
@dim = 'v1'
length = 6
"""

So the problem with this is that thinc.Config will generate an ArgumentModel from the registered architecture function

https://github.com/explosion/thinc/blob/adeb99ab7f005228c5a45a8c755929a6bb081d78/thinc/config.py#L910

in config._fill

        validation.update(result.dict(exclude=exclude_validation))
        filled, final = cls._update_from_parsed(validation, filled, final) # <- this breaks

in the first line both validation and results both have the dim attribute at the desired pydantic class.

    dim: <Dim length=6>

however on the call to results.dict() the pydantic model is transformed to a dictionary as is the behavior of Pydantic .dict method. Then in _update_from_parsed()
https://github.com/explosion/thinc/blob/adeb99ab7f005228c5a45a8c755929a6bb081d78/thinc/config.py#L932

            if isinstance(value, dict):
                filled[key], final[key] = cls._update_from_parsed(
                    value, filled[key], final[key]
                )

It iterates over validation items, thinks that dim is a dict and tries to set the key on the pydantic model in final[key].

I think there needs to be some way for Config to tell apart configs that come with a pydantic Schema and those that come from functions that happen to contain pydantic models. (Luckily I also have attrs in this project so could avoid erroneously casting it to a dict after figuring out the problem)

[honnibal]

Fixed in v1.2