Skip to content

wandb logging seemingly broken #28

@Tshimanga

Description

@Tshimanga

@clwainwright and I have talked about this but there seems to be bug in the wandb code where the config fails to communicate wandb's presence downstream. It looks like this is the bit of code dropping the ball

            wandb.init(
                name=job_name, notes=run_notes, project=project, entity=entity,
                config=base_config)
            # Note that wandb config can contain different and/or new keys that
            # aren't in the command-line arguments. This is especially true for
            # wandb sweeps.
            config.update(wandb.config._items)

            # Save the environment type to the wandb summary data.
            # This allows env_type show up in the benchmark table.
            wandb.run.summary['env_type'] = config['env_type']

            if job_name is None:

In the above block which is only reachable when args.wandb=True it looks like specifically wandb.config._items is the offender.

print(wandb.config._items) yields,

{'algo': 'ppo', 'env_type': 'append-spawn', 'side_effect.baseline': 'starting-state', 'side_effect.penalty': 0.1, 'steps': 100, 'data_dir': None, 'run_type': 'train', 'seed': None, 'deterministic': False, '_
wandb': {}}

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions