Cannot manipulate dict

[rob-hen]

🐛 Bug report

To reproduce

python file mwe.py:

from jsonargparse import CLI
from typing import Any, Dict
from dataclasses import dataclass


@dataclass
class ModelConfig:
    data: Dict[str, Any] = None


def method(model: ModelConfig = None):

    print(model.data)


if __name__ == "__main__":
    CLI(method)

yaml file mwe.yaml:

model:
  class_path: ModelConfig
  init_args:
    data:
      A: 1
      B: 2

python mwe.py --config mwe.yaml outputs:
{'A': 1, 'B': 2}

However,

python mwe.py --config mwe.yaml --model.init_args.data.A=4 produces the error message:

usage: mwe.py [-h] [--config CONFIG] [--print_config[=flags]] [--model.help CLASS_PATH_OR_NAME] [--model MODEL]
mwe.py: error: Parser key "model":
  Does not validate against any of the Union subtypes
  Subtypes: (<class '__main__.ModelConfig'>, <class 'NoneType'>)
  Errors:
    - Problem with given class_path '__main__.ModelConfig':
      - Unrecognized arguments: --data.A=4
    - Expected a <class 'NoneType'>
  Given value type: <class 'jsonargparse.util.NestedArg'>
  Given value: NestedArg(key='data.A', val='4')

Expected behavior

The dictionary should have been modified to {'A': 4, 'B': 2}

Environment

• jsonargparse version: 4.20.1:
• Python version: 3.9:
• How was jsonargparse installed: pip install jsonargparse.
• OS (e.g., Linux): Linux

[mauvilsa]

Thank you very much for reporting! Indeed there is a bug, though there are inconsistencies in the description. Thus, some clarifications are needed. Dataclasses don't behave like subclasses (see #287), so the config mwe.yaml should not have class_path and init_args. Similarly the command line should be --model.data.A=4 without init_args. And by using None as default, the type hint becomes optional, which is where the actual bug is. Support for dataclasses nested in a type was recently added (#267).

A slightly simpler reproduction which avoids using features unrelated to the bug is:

from jsonargparse import ArgumentParser
from typing import Any, Dict, Optional
from dataclasses import dataclass

@dataclass
class ModelConfig:
    data: Optional[Dict[str, Any]] = None

parser = ArgumentParser()
parser.add_argument("--model", type=Optional[ModelConfig], default=ModelConfig(data={"A": 1, "B": 2}))
parser.parse_args(["--model.data.A=4"])
print(cfg.model.data)

I have created pull request #290 with a fix. @rob-hen please have a look.