W&B: Allow for passing experiment into the WandbLogger (and logging semantics)

[colehurwitz]

Currently, the WandbLogger will automatically create a new internal experiment (run) whenever you create a new WandbLogger.

Issue

If I instantiate a wandb experiment outside of the logger, then I will have two experiments when I train my model since there is no way to set the internal experiment of the WandbLogger to my current external experiment.

Potential Solution

Allow for passing an experiment into the WandbLogger:

class WandbLogger(LightningLoggerBase):
    def __init__(self, name=None, save_dir=None, offline=False, id=None, anonymous=False,
                 version=None, project=None, tags=None, experiment=None):
              .
              .
              .
              self._experiment = experiment

Then I can do this:

experiment = wandb.init(.......)
wandb_logger = WandbLogger(experiment=experiment)

I made this change locally, however, I wasn't sure if this was something you also wanted to implement as well. It works for me.

Another small note

In the WandbLogger.log_metrics function, I would change:

self.experiment.history.add(metrics) --> self.experiment.log(metrics)

[williamFalcon]

@colehurwitz sounds good. feel free to submit PR!

[colehurwitz]

Will do!

[neergaard]

@colehurwitz Which version of wandb are you using? When I try logging metrics with self.logger.experiment.log(metric), I keep getting errors because the wandb Run object doesn't have a log method.

[colehurwitz]

Hi @neergaard, I am currently using wandb 0.8.21. Maybe this is a recent change?

If this is not backwards compatible with previously used versions, then I suggest we go back to:

self.experiment.log(metrics) ----> self.experiment.history.add(metrics)