[wingman -> rllib] Improved stats changes in AsyncSamplesOptimizer #3966
Conversation
bjg2
changed the title
bjg2
changed the title
|
Test FAILed. |
| @@ -85,7 +84,9 @@ def _init(self, | |||
| self.timers = {k: TimerStat() for k in ["train", "sample"]} | |||
There was a problem hiding this comment.
Does it also make sense to remove the timers?
There was a problem hiding this comment.
Well, I think so. Replaced self.timers with self._optimizer_step_timer.
|
Test FAILed. |
|
Test FAILed. |
|
Test PASSed. |
|
Test PASSed. |
|
Test FAILed. |
|
Pushed some lint fixes. Note that num remote_evaluators should be >0 still, since async samples optimizer doesn't have fallback code for 0 workers case like other optimizers do. |
|
Test FAILed. |
| "num_weight_syncs": self.num_weight_syncs, | ||
| "num_steps_replayed": self.num_replayed, | ||
| "timing_breakdown": timing, | ||
| "learner_queue": self.learner.learner_queue_size.stats(), | ||
| } | ||
| }, **self.get_mean_stats_and_reset()) |
There was a problem hiding this comment.
This syntax is needed for python 2.
|
Test FAILed. |
|
jenkins retest this please |
|
Test PASSed. |
|
Merged, thanks! |
Stats gathering in AsyncSamplesOptimizer is improved, as samples_throughput and train_throughput were giving bad values (mostly 0, with periodical peaks, instead of a steady graph showing throughput per second).