Defining class_path before init_args does not work, but switching the order does work.

[basveeling]

🐛 Bug report

Using LightningCLI, when setting the model property in a yaml file, the order of class_path and init_args matters. class_path before init_args causes a ParsingError.

To reproduce

Setup a LightingCLI with a base class BaseModelClass(LightningModule)
train.py would be something along the lines of

LightningCLI(model_class=BaseModelClass, ...)

Config file a.yaml:

model:
   class_path: SubclassedModelClass
   init_args:
     a: b

Config file b.yaml

model:
   init_args:
     a: b
   class_path: SubclassedModelClass

Run it as follows
python train.py fit -c a.yaml

jsonargparse.util.ParserError: Configuration check failed :: Parser key "model": Type <class 'BaseModelClass'> expects: a class path (str); or a dict with a class_path entry; or a dict with init_args (if class path given previously). Got "Namespace(init_args=Namespace(a="b"))".

(Notice the missing class_path)

Then run with the alternative order
python train.py fit -c b.yaml
Which runs successfully.

Expected behavior

Both orders should work

Environment

• jsonargparse version 4.14.0
• pytorch_lightning 1.7.4
• python 3.8.14
• pip install jsonargparse[signatures]
• Ubuntu

[mauvilsa]

Thank you for reporting. I tried a few things, but unfortunately I am unable to reproduce it. From the description I wrote:

from pytorch_lightning import LightningModule
from pytorch_lightning.cli import LightningCLI

class BaseModelClass(LightningModule):
    def __init__(self, a: str = '-'):
        pass

class SubclassedModelClass(BaseModelClass):
    def __init__(self, **kwargs):
        super().__init__(**kwargs)

LightningCLI(
    model_class=BaseModelClass,
    subclass_mode_model=True,
)

If I run this with the two config files you posted, both ./cli.py fit -c a.yaml --print_config and ./cli.py fit -c b.yaml --print_config work correctly (--print_config added just to stop after parsing).

If by mistake subclass_mode_model=True is missing, it doesn't fail when it should. This is a bug which I will fix, but it is not the problem you reported. Please check again and complete the code in the description so that it can be reproduced.

[basveeling]

Thank you for looking into this and setting up a proper reproduction code snippet. Apologies for not doing this myself, I was short on time. I’ll try and run your example in our environment and see if we get the same issue.

…

On Sat, 24 Sep 2022 at 15:50, Mauricio Villegas ***@***.***> wrote: Thank you for reporting. I tried a few things, but unfortunately I am unable to reproduce it. From the description I wrote: from pytorch_lightning import LightningModulefrom pytorch_lightning.cli import LightningCLI class BaseModelClass(LightningModule): def __init__(self, a: str = '-'): pass class SubclassedModelClass(BaseModelClass): def __init__(self, **kwargs): super().__init__(**kwargs) LightningCLI( model_class=BaseModelClass, subclass_mode_model=True, ) If I run this with the two config files you posted, both ./cli.py fit -c a.yaml --print_config and ./cli.py fit -c b.yaml --print_config work correctly (--print_config added just to stop after parsing). If by mistake subclass_mode_model=True is missing, it doesn't fail when it should. This is a bug which I will fix, but it is not the problem you reported. Please check again and complete the code in the description so that it can be reproduced. — Reply to this email directly, view it on GitHub <#161 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAEDDD2J6W3A5UPLGECQCQ3V74BJHANCNFSM6AAAAAAQT7FFXE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[mauvilsa]

Any update on this?

[mauvilsa]

I assume you were not able to reproduce the issue. Closing.