Override init_args for class in config file via CLI

[cgoliver]

I have a config like:

trainer:
  logger:
      - class_path: pytorch_lightning.logger.CSVLogger
         init_args: 
           name: "bob"

I would like to use this config and override the value of one of the init_args like so:

python main.py fit --config config.yaml --trainer.logger.name="joe"

However this causes an error that looks like this:

python3 main.py fit --config configs/synthetics.yaml --trainer.logger.name="joe"
usage: main.py [-h] [-c CONFIG] [--print_config [={comments,skip_null,skip_default}+]] {fit,validate,test,predict,tune} ...
main.py: error: Parser key "trainer.logger": Value "[Namespace(class_path='pytorch_lightning.loggers.CSVLogger', init_args=Namespace(save_dir=None, name='joe', version=None, prefix='', flush_logs_every_n_steps=100))]" does not validate against any of the types in typing.Union[pytorch_lightning.loggers.logger.Logger, typing.Iterable[pytorch_lightning.loggers.logger.Logger], bool]:
  - Type <class 'pytorch_lightning.loggers.logger.Logger'> 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(class_path='pytorch_lightning.loggers.CSVLogger', init_args=Namespace(save_dir=None, name='joe', version=None, prefix='', flush_logs_every_n_steps=100))]".
  - Problem with given class_path "pytorch_lightning.loggers.CSVLogger":
    - Configuration check failed :: Key "save_dir" is required but not included in config object or its value is None.
  - Expected a <class 'bool'> but got "[Namespace(class_path='pytorch_lightning.loggers.CSVLogger', init_args=Namespace(save_dir=None, name='joe', version=None, prefix='', flush_logs_every_n_steps=100))]"

Version:

pytorch-lightning==1.7.1

[tshu-w]

Please try to update your jsonargparse through pip install -U jsonargparse, it supports overriding init_args after this commit

[cgoliver]

Thanks! I have updated my jsonargparse jsonargparse==4.14.0

But not sure what the new syntax would be, and running things as in the original post is still causing the same error.

Would you be able to kindly provide an example of the config file and a CLI which would override one of the config file init_args?

[tshu-w]

trainer:
  logger:
    class_path: CSVLogger
    init_args:
      save_dir: "."
      name: "bob"

from pytorch_lightning.cli import LightningCLI

from pytorch_lightning.demos.boring_classes import DemoModel, BoringDataModule

def run():
    LightningCLI(
        DemoModel,
        BoringDataModule,
        trainer_defaults={"max_epochs": 2},
        save_config_callback=False,
    )


if __name__ == "__main__":
    run()

python cli.py fit --config config.yaml --trainer.logger.name="alice" --print_config=skip_null

[cgoliver]

That's perfect. Thanks!