Could a subclass type hint be a config file when using CLI?

[MendelXu]

I want to load a config file with the structure like:

input:
  class_path: src.input.KeyBoard
  init_args: input/keyboard.yaml

But it would not parse the file specified in init_args as a Dict. I also try to write it as

input:
  class_path: src.input.KeyBoard
  init_args:
    cfg: input/keyboard.yaml

and it still does not work. Is there any way to implement such function?

[mauvilsa]

Could you please add information about how you are creating the parser. Would be even better if it is a small code snippet that can be easily run.

Right now I am not sure if this can be easily implemented. Though in general I think this is not a good idea. The class_path and init_args jointly specify a subclass. In other words what init_args can be depends on the value of class_path. Having these two in separate files can make things confusing. It would advice to have instead:

# main.yaml
input: input/keyboard.yaml

# keyboard.yaml
class_path: src.input.KeyBoard
init_args:
  arg1: val1
  ...

Or maybe what you want is have a configurable class (not a subclass), so KeyBoard is fixed and in a separate file have its args. In this case having the args in a separate file would make sense.

[MendelXu]

Thanks for your replying. I create a small case like

#main.py
from jsonargparse import CLI


class B:
    def __init__(self, meta_info: dict):
        self.meta_info = meta_info


class A:
    def __init__(self, b: B = None, a_lot_of_params: dict=None):
        self.b = b
        self.a_lot_of_params = a_lot_of_params
        
    def hello(self):
        print('I am ok')

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

#cfg.yaml
b: b.yaml
a_lot_of_params: 
  a: 1
  b: 2

#b.yaml
class_path: __main__.B
init_args: 
  meta_info:
    something: true
    others: 1

By running python main.py --config cfg.yaml hello, I want to instantiate an object of class B for the class A.
However, it would return the error that

error: Parser key "b": Value "b.yaml" does not validate against any of the types in typing.Union[__main__.B, NoneType] :: Problem with given class_path "b.yaml" :: No module named 'b' :: Expected a <class 'NoneType'> but got "b.yaml"

[mauvilsa]

If you change line signatures.py#L295 to kwargs['action'] = ActionTypeHint(typehint=annotation, enable_path=True) it will work as you want. The unit tests of the package don't break with this change. But I should think a bit if this could result in some unexpected consequences. If not this could be added as a parameter of CLI to enable it or maybe it would make sense to be enabled by default.

[MendelXu]

Got it. Thank you very much!

[MendelXu]

OK. I find that in a more complicated situation, it still fails.

#main.py
from jsonargparse import CLI

class B:
    def __init__(self, meta_info: dict):
        self.meta_info = meta_info


class A:
    def __init__(self, b: B = None, a_lot_of_params: dict=None):
        self.b = b
        self.a_lot_of_params = a_lot_of_params
        
    def hello(self):
        print('I am ok')

class C(A):
    def __init__(self,b: B = None, a_lot_of_params: dict=None):
        pass

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

#cfg.yaml
class_path: C
init_args:
  b: b.yaml

And the error is

error: Conflicting namespace base: init_args.b, expected dict but is b.yaml.

[mauvilsa]

With the code snippet above I am unable to reproduce that error message. Furthermore, class components given to CLI do not behave like subclass type hints. The main config file can't start with a class_path and init_args pair. What is supported is that b could be a subclass of B.

[mauvilsa]

I am not sure what you are trying to achieve. If you want to be able to run A or C then maybe what you want might be CLI([A, C]) which would make each class a subcommand.

[MendelXu]

Sorry for the late reply. I am on a deadline and unable to deal with it now. I think I can close the issue and process it after my deadline. Thank you very much.