-
Notifications
You must be signed in to change notification settings - Fork 6
Open
Labels
documentationImprovements or additions to documentationImprovements or additions to documentationenhancementNew feature or requestNew feature or requestpriority: lowIssue with low priorityIssue with low priority
Description
Description:
Currently, there’s a filename clash between the config files generated by Weights & Biases (W&B) and PyTorch Lightning CLI:
- W&B automatically generates a file named
config.yaml
to store hyperparameters and run settings passed for a given run. - LightningCLI also generates a file named
config.yaml
(by default) containing the CLI/Lightning configuration for reproducibility.
Since both have similar default names, we changed LightningCLI’s config file name to avoid conflicts:
ChebaiCLI(
save_config_kwargs={"config_filename": "lightning_config.yaml"},
parser_kwargs={"parser_mode": "omegaconf"},
)
This prevents overwriting and allows both files to coexist.
However, while lightning_config.yaml
is saved locally (inside the W&B run’s files
directory), it is not uploaded to the W&B online run logs.
Why this matters:
The lightning_config.yaml
contains much more detailed information than W&B’s config.yml
— including:
- The classes and subclasses instantiated
- Detailed CLI arguments and defaults
- Model architecture instantiation details
- Extra metadata useful for experiment reproducibility
Having this file in the online W&B run logs would greatly improve:
- Traceability of runs
- Experiment reproducibility
- Collaboration across teams, where others can reconstruct the exact CLI command and configuration
- Also helpful when for debugging and reverse engineering to find certain inconsistencies across runs
Proposed solution:
- Ensure that
lightning_config.yaml
(or the filename set viasave_config_kwargs
) is automatically uploaded to W&B run logs, just like artifacts and logs.
Metadata
Metadata
Assignees
Labels
documentationImprovements or additions to documentationImprovements or additions to documentationenhancementNew feature or requestNew feature or requestpriority: lowIssue with low priorityIssue with low priority