Optional class instances

[hadipash]

Hello! I would like to ask what is a good way to handle optional class instances? For example, instantiating a validation dataset. In the original Python code, it would look like this:

if args.val_dataset is not None:
    # instantiate the validation set class

However, add_class_arguments() doesn't support optional instances. To work around this, I'm currently using add_subclass_arguments(required=False), but I find this approach suboptimal. Is there a better way to handle situations like this in jsonargparse? Thank you.

[mauvilsa]

The only way is to make a subclass. Why exactly do you find it suboptimal?

[hadipash]

Because the class I would like to instantiate is known in advance and is not a subclass. Moreover, running python train.py --help will output this

--valid.dataset.help CLASS_PATH_OR_NAME
                        Show the help for the given subclass of ImageVideoDataset and exit.
--valid.dataset CONFIG | CLASS_PATH_OR_NAME | .INIT_ARG_NAME VALUE
                        One or more arguments specifying "class_path" and "init_args" for any subclass of ImageVideoDataset. (type: Optional[ImageVideoDataset], default: null, known
                        subclasses: moviegen.dataset.ImageVideoDataset)

instead of the list of arguments. I can use python train.py --valid.dataset.help ImageVideoDataset to see them, but it's less intuitive in my opinion.

It will be nice if add_class_arguments() and add_function_arguments() supported required parameter as well. Thank you!

[mauvilsa]

It will be nice if add_class_arguments() and add_function_arguments() supported required parameter as well. Thank you!

The way these work is that they add a group of arguments corresponding to the parameters of the function/class. But all added arguments are independent so they will always be part of the parsed namespace. It is not possible to have an optional group in argparse.

Because the class I would like to instantiate is known in advance and is not a subclass.

Since v4.29.0 it is possible to provide to a subclass without providing class_path. In such a case the base class is used. So at least in a config file the content and individual command line arguments can be the same as for add_class_arguments. Though, --print_config/dump/save would have class_path and init_args.

Moreover, running python train.py --help will output this

Somewhat related I have been working on #287 (comment). My idea is that in the future there wouldn't be a difference between subclasses and dataclasses. A part of this that I was planning is to make the class path for help optional, and if not given, then the help for the base class is printed, i.e. it would look like

--valid.dataset.help [CLASS_PATH_OR_NAME]

and running cli.py --valid.dataset.help would give the help.

I do agree that it is not ideal even with these possible improvements. But note that inevitably these automatic creation of parsers from signatures will have some limitations and compromises.

[hadipash]

Thank you for your detailed reply. jsonargparse significantly simplifies parsers creation and reduces boilerplate code. So, its minor limitations are not a big trade-off. 👍