Closed kim-mskw closed 7 months ago
Attention: 39 lines
in your changes are missing coverage. Please review.
Comparison is base (
3328c38
) 81.90% compared to head (cad0911
) 81.83%.
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
I still get AttributeError: 'RLStrategy' object has no attribute 'actor'
when running:
assume -s example_02a -db sqlite:///./examples/local_db/test_mini_rl.db
Though I fixed some things which seem to have not looked right, merged main and made the tests pass :)
current main does also fail on load_actor_params
- seems that this issue exists longer..
There is world.learning_role.compare_and_save_policies()
which somehow does not what seems to be intended (it does not save policies and lacks the retrieval of the compared policies and has that weird self.rl_eval_rewards
which does not do anything.
I think it would be good to go through this and fix loading and comparing scenarios.
Then self.actor.optimizer.load_state_dict(params["actor_optimizer"])
says
AttributeError: 'Actor' object has no attribute 'optimizer'
which has to be checked too
Turns out that compare_and_save_policies
is a weird method which is never used but should be fixed somehow.
Instead of using this method of the learning_role - the correct function to use instead of learning_role.save_params
would have been learning_role.rl_algorithm.save_params
.
Now, everything works again :blush:
Turns out that
compare_and_save_policies
is a weird method which is never used but should be fixed somehow. Instead of using this method of the learning_role - the correct function to use instead oflearning_role.save_params
would have beenlearning_role.rl_algorithm.save_params
.Now, everything works again 😊
Actually, it does not. If continue_learning=True we should read the actor and critic params, which is currently not done. This is what I tried to fix for the redundant config variable load_model_path. I am on it though.