Support multi-inheritance classes

[florianblume]

🚀 Feature request

As far as I can tell, currently there's not support for multi-inheritance classes (but of course let me know if I'm wrong or implementing something wrong). The following example throws an error:

from jsonargparse import ArgumentParser
import yaml

class ModelBaseClass():
    def __init__(self, batch_size: int=0):
        self.batch_size = batch_size

class ImageModel(ModelBaseClass):
    def __init__(self, image_size: int=0, **kwargs):
        super().__init__(**kwargs)
        self.image_size = image_size

class ClassModel(ModelBaseClass):
    def __init__(self, num_classes: int=0):
        self.num_classes = num_classes

class ImageClassModel(ImageModel, ClassModel):
    def __init__(self, **kwargs):
        super().__init__(**kwargs)

parser = ArgumentParser(
    prog='app',
    description='Description for my app.'
)

parser.add_subclass_arguments(ModelBaseClass, 'model')

model = yaml.safe_dump({
    'class_path': '__main__.ImageClassModel',
    'init_args': {
        'image_size': 5,
        'num_classes': 10
    }
})

print(parser.parse_args([f'--model={model}']))

It used to work in the fork I made off of version 4.8 but somehow stopped to work now.

Motivation

I guess the example makes it pretty obvious why this is a neat feature, it just makes sense to separate the classes like this (maybe you have a classifier that is not classifying images) but also compose them again.

Pitch

It would be cool if it would work again like in version 4.8 :) I'd also happy to help out if it's a simpler fix but it might be faster/nicer coded if you do it yourself ;)

Alternatives

Not really any that I can think of.

[mauvilsa]

This worked before, making it a regression. So changed the issue from enhancement to bug.

[mauvilsa]

This is fixed in commit 57fe548. Note that there is a mistake in the code snippet you included above. Creating an instance of ImageClassModel will not call ModelBaseClass.__init__, which means that batch_size will not be set. This can be observed by running list(ImageClassModel().__dict__.keys()) which gives ['num_classes', 'image_size']. The problem is that ClassModel is missing super().__init__(**kwargs). Without fixing that code, jsonargparse will work the same way, not accepting the parameter batch_size.

[florianblume]

Nice, works perfectly. Thanks about mentioning the bug in the example script, it's not actual code but I wrote it from scratch, that's probably why this error is in there.

I'd love to directly implement a fix next time but every time I look at your changes I get that I don't understand the project well enough for that, yet... So thanks a lot for your work!