Let's say, I am trying to optimize some hyperparameters on a goal metric by maximizing it. The selected sweep method is bayes. Also, that goal metric is defined by define_metric function with a summary flag, namely max. Basically, on the summary table, I am logging the max value of this goal metric, instead of the last logged value.
My question is, how does the Bayesian sweep work with define_metric method? Does the sweep optimize on the metric by maximizing the summary value (last logged/max/min etc.) ? Or does it neglect define_metric and takes a global maximum value into account ?
In the training script:
wandb.define_metric("goal_metric", summary='max')
...
wandb.log({"goal_metric": ...
In the sweep configuration:
method: bayes
metric:
goal: maximize
name: goal_metric