LightningCLI: how to configure logger using cmd-line args?

[slinnarsson]

I would like to change the names of the logging directories from the default "version_{n}" to something of my own choosing. How can I do this using command-line arguments to LightningCLI?

I know I can set the logger using trainer.logger but setting logger args e.g. trainer.logger.version does not work (unrecognized argument). So how can I pass args to the logger?

[rohitgr7]

cc @carmocca

[carmocca]

See my reply here: #10574 (comment)

We'll be adding support for shorthand notation shortly too: #11533

[carmocca]

After the 1.6 release the following will work too:

--trainer.logger TensorBoardLogger \
--trainer.logger.save_dir 'lightning_logs' \
--trainer.logger.version 'my_subdir' \

[slinnarsson]

Btw the above code makes the logs go to 'lightning_logs/default/my_subdir'. I.e. there's an extra directory level, but Tensorboard picks up the logs without problem. I think 'default' comes from the 'name' parameter to TensorBoardLogger, so if you want to change it, you have to provide that as well.

[YoniSchirris]

Which version did this work @slinnarsson ? in 1.5.1 I'm getting

Traceback (most recent call last):
  File "/dlup_lightning_mil/tools/train.py", line 25, in <module>
    main()
  File "/dlup_lightning_mil/tools/train.py", line 10, in main
    cli = DLMCLI(model_class=pl.LightningModule,
  File "/users/dlup_lightning_mil/miniconda3/lib/python3.9/site-packages/pytorch_lightning/utilities/cli.py", line 489, in __init__
    self.parse_arguments(self.parser)
  File "/users/dlup_lightning_mil/miniconda3/lib/python3.9/site-packages/pytorch_lightning/utilities/cli.py", line 628, in parse_arguments
    self.config = parser.parse_args()
  File "/users/dlup_lightning_mil/miniconda3/lib/python3.9/site-packages/pytorch_lightning/utilities/cli.py", line 233, in parse_args
    return super().parse_args(*args, **kwargs)
  File "/users/dlup_lightning_mil/miniconda3/lib/python3.9/site-packages/jsonargparse/deprecated.py", line 112, in patched_parse
    cfg = parse_method(*args, _skip_check=_skip_check, **kwargs)
  File "/users/dlup_lightning_mil/miniconda3/lib/python3.9/site-packages/jsonargparse/core.py", line 366, in parse_args
    cfg, unk = self.parse_known_args(args=args)
  File "/users/dlup_lightning_mil/miniconda3/lib/python3.9/site-packages/jsonargparse/core.py", line 253, in parse_known_args
    namespace, args = self._parse_known_args(args, namespace)
  File "/users/dlup_lightning_mil/miniconda3/lib/python3.9/argparse.py", line 2062, in _parse_known_args
    start_index = consume_optional(start_index)
  File "/users/dlup_lightning_mil/miniconda3/lib/python3.9/argparse.py", line 2002, in consume_optional
    take_action(action, args, option_string)
  File "/users/dlup_lightning_mil/miniconda3/lib/python3.9/argparse.py", line 1930, in take_action
    action(self, namespace, argument_values, option_string)
  File "/users/dlup_lightning_mil/miniconda3/lib/python3.9/site-packages/jsonargparse/typehints.py", line 269, in __call__
    if cfg_dest.get('class_path') == val:
AttributeError: 'bool' object has no attribute 'get'

with the --trainer.logger.class_path pytorch_lightning.loggers.TensorBoardLogger option.

[slinnarsson]

I'm on 1.5.6 now, and probably was back then too.

[carmocca]

@YoniSchirris I believe that was caused by a bug in jsonargparse. Try pip install -U jsonargparse[signatures]