fix the custom_preprocessor issue arising from training for centraliz…

…ed paradigm

baselines/marl_benchmark/marl_benchmark/__init__.py

@@ -57,15 +57,12 @@ def gen_config(**kwargs):
                 "obs_space": gym.spaces.Tuple([obs_space] * agent_missions_count),
                 "act_space": gym.spaces.Tuple([act_space] * agent_missions_count),
                 "groups": {"group": agent_ids},
+                "model": config["policy"][-1],
             }
         )
-        tune_config.update({"model": config["policy"][-1]})
-
     policies = {}
     for k in agents:
-        policies[k] = config["policy"][:-1] + (
-            {**config["policy"][-1], "agent_id": k},
-        )
+        policies[k] = config["policy"][:-1] + ({**config["policy"][-1], "agent_id": k},)
     tune_config.update(
         {
             "multiagent": {

baselines/marl_benchmark/marl_benchmark/evaluate.py

@@ -86,11 +86,10 @@ def main(
         tune_config = config["run"]["config"]
         trainer_cls = config["trainer"]
         trainer_config = {"env_config": config["env_config"]}
-        if paradigm != "centralized":
-            trainer_config.update({"multiagent": tune_config["multiagent"]})
-        else:
-            trainer_config.update({"model": tune_config["model"]})
+        if paradigm == "centralized":
+            trainer_config["model"] = config["policy"][-1]
 
+        trainer_config.update({"multiagent": tune_config["multiagent"]})
         trainer = trainer_cls(env=tune_config["env"], config=trainer_config)
         trainer_config["evaluation_interval"] = True
         trainer.setup(trainer_config)

baselines/marl_benchmark/marl_benchmark/wrappers/rllib/frame_stack.py

@@ -25,6 +25,7 @@
 import gym
 import numpy as np
 from ray import logger
+from ray.rllib.models import ModelCatalog
 from ray.rllib.models import Preprocessor
 from ray.rllib.models.preprocessors import get_preprocessor
 from ray.rllib.utils.annotations import override
@@ -144,7 +145,8 @@ def stack_frames(frames):
 
     @staticmethod
     def get_preprocessor():
-        return TupleStackingPreprocessor
+        ModelCatalog.register_custom_preprocessor("my_prep", TupleStackingPreprocessor)
+        return "my_prep"
 
     def _get_observations(self, raw_frames):
         """Update frame stack with given single frames,

baselines/marl_benchmark/marl_benchmark/wrappers/rllib/group.py

@@ -21,7 +21,6 @@
 # THE SOFTWARE.
 import copy
 
-from ray.rllib.agents.dqn import DQNTrainer
 from ray.rllib.env.constants import GROUP_INFO, GROUP_REWARDS
 from ray.rllib.env.group_agents_wrapper import _GroupAgentsWrapper