Support for discriminative union

[julien-blanchon]

Hey, in my workflow I'm using a lot the following pattern

from dataclasses import dataclass
from typing import Annotated, Literal
from pydantic import Field
from datargs import parse


@dataclass
class DatasetImage:
    type: Literal["image"]
    path: str
    resolution: tuple[int, int]


@dataclass
class DatasetVideo:
    type: Literal["video"]
    path: str
    resolution: tuple[int, int]
    fps: int


@dataclass
class Config:
    dataset: Annotated[DatasetImage | DatasetVideo, Field(discriminator="type")]


if __name__ == "__main__":
    args = parse(Config)
    print(args)

It's working well with cyclopts but not with datargs.
cyclopts require a discriminator Field from pydantic to work.

Any opinion to add this to datargs ?

[roee30]

I imagine you're expecting a parser that expects different arguments depending on the value of --type. It seems argparse only supports this by parsing twice, first with .parse_known_args(). This does not seem to me standard or common usage, so I am not inclined to support it.

However, you can use this:

from dataclasses import dataclass
from typing import Literal
from datargs import parse, make_parser


@dataclass
class DatasetImage:
    path: str
    resolution: tuple[int, int]


@dataclass
class DatasetVideo:
    path: str
    resolution: tuple[int, int]
    fps: int


@dataclass
class Config:
    type: Literal["image", "video"]


if __name__ == "__main__":
    args, rest = make_parser(Config).parse_known_args()
    args = parse(DatasetVideo if args.type == "video" else DatasetImage, rest)
    print(args)

Alternatively, with slight modifications, you can get similar behavior with subcommands:

class Image: ...
class Video: ...
class Args:
    command: Image | Video

# usage:
python script.py image --resolution ...
python script.py video --fps ...