Is it possible to drop init_args?

[florianblume]

I've tried to find out in the code whether there are technical reasons for why we need init_args in configs to pass the init arguments for subclasses. I assume there is but I can't see it and was also wondering whether there's a way around it.

My issue is somewhat related to #89.

I'm using Pytorch Lightning and would like to pass the init arguments to my LightningModule subclasses in the same way I pass them to the trainer:

trainer:
    gpus: 2
    deterministic: true
    ...

since this way I can omit init_args in the command line (which makes commands pretty long) and I can pass overwrite parameters in additional configs without having to specify the class_path again (it complains when I only pass init_args). This drove me to putting more and more stuff into the config file I edit before running experiments, instead of having most of it in the default config file.

I assume that you don't need the init_args part when you don't pass subclasses (as in the case of the trainer class) but is there a way to achieve this for subclasses, too, that I'm missing? I guess I could fake it by inserting init_args into the config before class instantiation but that feels a bit wrong. Could it be implemented properly? I'd also be happy to help.

[mauvilsa]

Yes, it is possible and certainly related to #89. The initial concept was simple, a subclass is a single argument (one action class instance for argparse) which requires to have both class_path and init_args. Extending this as suggested here and in #89 is considerably more complex. The limitation is only that this needs to be properly thought through, designed, consider how this interacts with all other features, implemented, tested, etc. Essentially it is not a small task.

Help is certainly welcome. I will try to create a list of things that I think need to be done to get this going, hopefully with you and more people contributing.

[florianblume]

Ok cool, keep me posted when you've got a list of what to do and we can see where I can jump in!

[mauvilsa]

In #89 (comment) it was suggested to keep init_args but not have class_path. Some relevant explanation about constraints is this other #89 (comment).

[florianblume]

Yes, thanks for the extensive comment! As soon as there are tasks to do I'd be happy to help implement them.

[mauvilsa]

With commit 44eb5df it is now possible to give init_args in a config without specifying the class_path. It is possible to give nested in an init_args or directly without nesting in init_args. See the unit test for an example:

jsonargparse/jsonargparse_tests/test_typehints.py

Lines 625 to 643 in 44eb5df

	def test_init_args_without_class_path(self):
	parser = ArgumentParser()
	parser.add_argument('--config', action=ActionConfigFile)
	parser.add_argument('--cal', type=Calendar)
	config = """cal:
	class_path: TextCalendar
	init_args:
	firstweekday: 2
	"""
	cal = """init_args:
	firstweekday: 3
	"""
	cfg = parser.parse_args([f'--config={config}', f'--cal={cal}'])
	self.assertEqual(cfg.cal.init_args, Namespace(firstweekday=3))
	cfg = parser.parse_args([f'--config={config}', f'--cal={cfg.cal.init_args.as_dict()}'])
	self.assertEqual(cfg.cal.init_args, Namespace(firstweekday=3))

[florianblume]

Love it!