How to add custom parsing

[Conchylicultor]

I have a code like:

@dataclasses.dataclass
class Args:
  model_cls: type[Model] = Transformer

I would like to support this through CLI. I have a mapping str -> type[Model], such as the user can pass a string --model_cls=Transformer that I can then normalize to the model class.

Something like:

@dataclasses.dataclass
class Args:
  model_cls: type[Model] = field(default=Transformer, parsing_type=str)

  def __post_init__(self):
    if isinstance(self.model_cls, str):
      self.model_cls = _NAME_TO_CLS[self.model]

However I do not know how to make simple_parsing parse model_cls as a str.

[lebrice]

Hi @Conchylicultor , thanks for posting!

I suggest you take a look at the subgroups feature

Otherwise you can also have a field with simple_parsing.field(choices=dict_with_str_to_models.keys(), default="transformer")

[Conchylicultor]

Thank you for the answer, unfortunately in my case, neither subgroup nor choices can be used as models can be dynamically registered (user can pass arbitrary modules --cfg.model="module.to.import.MyModel"), so I do not know the exact mapping in advance. The example from my original post was simplified.

I ended up using a hack to trick simple parsing to interpret the field as a str:

@dataclasses.dataclass
class Args:
  if typing.TYPE_CHECKING:  # For static type checking
    model_cls: type[Model] = ...
  else:  # For simple parsing
    model_cls: str = Transformer

So I'm fine at the moment. Keeping this bug open in case a cleaner solution is planned but I'll also understand if this bug was closed as there's a workaround.