pre_episode and post_episode methods.As mentioned in the Platform Future Reality Tree - #13 by tslominski update, the work to remove pre_episode and post_episode methods from Learning Modules was completed.
As a consequence of migrating all the configuration to use Hydra a while back, we eliminated the use of inheritance hierarchy from configurations. This update removes (146 Platform configuration relies on inheritance hierarchy), as it is no longer true.
As mentioned in the Future Reality Tree Platform Future Reality Tree - #14 by tslominski update, the undesirable effect has been removed, thus eliminating the entire subtree from the Current Reality Tree.
@sknudstrup highlighted two undesirable effects. These are now added to the Current Reality Tree, but require further elaboration.
(385 Can’t load pickled models if classes change.)
(386 Can’t load pickled models without tbp.monty in the environment.)
In one of the recent meetings, I observed that we hesitate to update pretrained models because of the amount of toil required. This is adjacent and related to the general toil of running and updating benchmarks, but is its own activity.
For now, I documented the Undesirable Effect (387 Pretrained models are not often updated), which will require further elaboration.
