Update fail_untyped flag to work with dataclasses whose arguments are classese with untyped arguments.

[ioangatop]

🚀 Feature request

The fail_untyped works with classes and nested classes, but id does not work with dataclasses which have class arguments with untyped arguments in their __init__.

Motivation

Dataclasses come handy when we build complex CLI frameworks as it organise the input fields without the need of typing class_path and init_args. The powerful feature here is to have objects as arguments that could be configured directly from the CLI. However, even though when these objects have arguments that are typed it works great, but the majority of open source doesn't, and thus limits us quite a bit.

Here is a snip to reproduce and to make sense of the above, quite complex 😅, description:

# main.py
from dataclasses import dataclass
from jsonargparse import CLI

class Database:
    def __init__(self, version) -> None:  # <-- HERE version is untyped; for `version: str` works!
        self._version = version

@dataclass
class MySQLConfig:
    db: Database
    host: str = "localhost"
    port: int = 8080

def my_app(cfg: MySQLConfig) -> None:
    print(f"{cfg.host}, {cfg.port}, {cfg.db._version}")

if __name__ == "__main__":
    CLI(my_app, fail_untyped=False)

Command:

python3 main.py --cfg.host "localhost" --cfg.port 8080 --cfg.db Database --cfg.db.version "0.0.1"

mind that when changing version to version: str, it works as expected.

Thank you for you time and for maintaining this great library!

[mauvilsa]

Thank you for the proposal. Looks good!

[ioangatop]

@mauvilsa when you find the time, please check my PR #242 which fixes this!