search

facebookresearch / how-to-autorl

Plug-and-play hydra sweepers for the EA-based multifidelity method DEHB and several population-based training variations, all proven to efficiently tune RL hyperparameters.
Apache License 2.0
66 stars 12 forks source link

Extended Weights and Biases Logging (Improvement) #20

Open LabChameleon opened 10 months ago

LabChameleon commented 10 months ago

Hi!

I think the Weights and Biases logging in the sweepers could be improved. During my work with the DEHB sweepers, I found the following extensions useful:

  1. Allowing to specify the Weights and Biases group and job type. This way one can have a group with the optimization having the job_type e.g. 'search'. Then all the evaluation runs of the found incumbent could be located in the same group with job_type e.g. 'eval'. This makes the whole setup cleaner if running multiple optimizations in the same Weights and Biases project.
  2. Sometimes my optimizations crash for various cluster hiccups. In these cases, it is useful to specify the Weights and Biases id of the crashed search and allow resuming them.

Consider the following PR for both improvements: #22