Mava
Mava copied to clipboard
Feature: Control number of vmapped envs in evaluator using `arch.num_envs`
What?
Modify the evaluator to limit the number of vmapped envs to arch.num_envs
when the total number of evaluation episodes arch.num_eval_episodes
exceeds this limit (instead of parallelise all the arch.num_eval_episodes
). In such cases, evaluations are conducted in sequential batches, with each batch containing arch.num_envs
parallel envs.
Why?
Limiting parallel evaluations to num_envs prevents out-of-memory issues by avoiding vmap over all episodes at once.