accelerate

accelerate copied to clipboard

Right now, it seems config is ignored when you accelerator.init_trackers(run, config) for Tensorboard tracker. Would be nice if the config is saved as hparams.yaml so it shows up in Tensorboard.

Wdyt @muellerzr ?

I think it makes sense to me, and still keeps in line with it being consistent throughout all the APIs for their respective trackers

@muellerzr can I work on this?

@Shreyz-max sure! 😄

So, I did some research and found out that according to this issue, the HParams dashboard currently requires at least one hparam and at least one metric defined in the hparams configuration but according to this line here we are leaving the metric as empty. So, should I save the configs as a yaml file in the tracking.py file itself?

@Shreyz-max that does seem like that'd be the way to do it then, so long as it can be shown in tensorboard correctly

@muellerzr cool, thanks for the clarification. I'll complete it then.