How to early stop bad runs in sweeps to save time

Hello,
that’s my first topic in the community, so I hope I am posting that in the correct category :wink:

I started exploring sweeps last week for a university project, and it is incredible! As we also got a new PyTorch version with support for the new apple silicon, I wanted to try that on my M1 Pro. As this is not as powerful as, for example, using GoogleColab for a fraction of the time, I wanted to ask if it is somehow possible to stop bad runs after a few epochs.

As you can see in the report linked below, the run hopeful-sweep-2 does not look promising. It would be nice to cancel that run and start a new one instead.

Thanks,
Markus

Hi @markuskarner ,

Thank you for writing in with your question. We do support early termination of sweeps, this reference doc covers this. When the early stopping is triggered, the agent stops the current run and gets the next set of hyperparameters to try. Here is a link to an example sweep configuration for reference. If after setting up your configuration and your require review / feedback. Please do write back in this thread and we can review your work more closely.

Regards,

Mohammad

1 Like

Thanks a lot :smiley:
That was exactly what I was looking for, just searched for early stopping and not terminating and somehow didn’t find it :slight_smile:

Is there also a way to change the configuration of a running sweep? Or stop it and continue it with a new configuration?

Thanks & Regards,
Markus

@markuskarner,

You’re welcome, glad that helped.

Currently we do not allow the modification of a running sweep. A sweep is inherently tied to its config, so once it is set there is no way to change it. You can however, create a new sweep and seed it with existing runs, see here. Hope this helps.

Regards,

Mohammad

Thanks for your kind reply :slight_smile: