Hey, I am logging custom metrics from my ray tune run to tensorboard by overriding the on_episode_end function from DefaultCallbacks . For each value I have currently value_mean, value_max and value_min, but I only want to save the value_mean field. Is there a way to configure this? (edited)
1 Like
No, unfortunately, this is still pretty hard-coded to store mean/max/min values only.
Hi @sven1977, I have a similar question on this topic, but instead, I want to calculate and log the variance on a custom metric. Concretely, I have a few workers that will log some stats regarding the environment setting and I want to see how these stats vary across different workers. Is there a clean way to do this? I tried to look into _process_trial and _process_trial_result for some insights but I couldn’t figure out where does RLlib calculates those min/mean/max stats.
As I look into metric_ops.py deeper today I think I resolve my doubt. CollectMetrics stores a queue of episode history for the size of min_history=100, therefore it calculates a running mean with the window of 100 episodes.
I think perhaps this should be noted somewhere in the docs?