Merge branch 'internal-policy-ghost' into soccer-2v1

ml-agents/mlagents/trainers/ghost/trainer.py

     def save_model(self, name_behavior_id: str) -> None:
         """
         Forwarding call to wrapped trainers save_model
-        Loads the latest policy weights, saves it, then reloads
-        the current policy weights before resuming training.
-        policy = self.trainer.get_policy(brain_name)
-        reload_weights = policy.get_weights()
-        # save current snapshot to policy
-        policy.load_weights(self.current_policy_snapshot[brain_name])
-        self.trainer.save_model(name_behavior_id)
-        # reload
-        policy.load_weights(reload_weights)
+        self.trainer.save_model(brain_name)
-        First loads the latest snapshot.
-        policy = self.trainer.get_policy(brain_name)
-        policy.load_weights(self.current_policy_snapshot[brain_name])
         self.trainer.export_model(brain_name)
 
     def create_policy(self, brain_parameters: BrainParameters) -> TFPolicy:
         return self.trainer.create_policy(brain_parameters)
 
     def add_policy(
-        self, parsed_behavior_id: BehaviorIdentifiers, policy: TFPolicy
+        self, parsed_behavior_id: BehaviorIdentifiers, brain_parameters: BrainParameters
     ) -> None:
         """
         Adds policy to trainer. The first policy encountered sets the wrapped
         name_behavior_id = parsed_behavior_id.behavior_id
         team_id = parsed_behavior_id.team_id
         self.controller.subscribe_team_id(team_id, self)
+        policy = self.create_policy(brain_parameters)
+        policy.create_tf_graph()
+        policy.init_load_weights()
-        policy.create_tf_graph()
-        policy.init_load_weights()
+            # creates an internal trainer policy. This always contains the current learning policy
+            # parameterization and is the object the wrapped trainer uses to compute gradients.
+            self.trainer.add_policy(parsed_behavior_id, brain_parameters)
+            internal_trainer_policy = self.trainer.get_policy(
+                parsed_behavior_id.brain_name
+            )
+            internal_trainer_policy.create_tf_graph()
+            internal_trainer_policy.init_load_weights()
+
-            ] = policy.get_weights()
+            ] = internal_trainer_policy.get_weights()
+
+            # initialize ghost level policy to have the same weights
+            policy.load_weights(internal_trainer_policy.get_weights())
-            self.trainer.add_policy(parsed_behavior_id, policy)
             self._learning_team = self.controller.get_learning_team
             self.wrapped_trainer_team = team_id
 

ml-agents/mlagents/trainers/ppo/trainer.py

         self._check_param_keys()
         self.load = load
         self.seed = seed
-        self.policy: NNPolicy = None  # type: ignore
+        self.policy: TFPolicy = None  # type: ignore
 
     def _check_param_keys(self):
         super()._check_param_keys()
         return policy
 
     def add_policy(
-        self, parsed_behavior_id: BehaviorIdentifiers, policy: TFPolicy
+        self, parsed_behavior_id: BehaviorIdentifiers, brain_parameters: BrainParameters
     ) -> None:
         """
         Adds policy to trainer.
                     self.__class__.__name__
                 )
             )
-        if not isinstance(policy, NNPolicy):
-            raise RuntimeError("Non-NNPolicy passed to PPOTrainer.add_policy()")
-        self.policy = policy
+        self.policy = self.create_policy(brain_parameters)
-        self.step = policy.get_current_step()
+        self.step = self.policy.get_current_step()
         self.next_summary_step = self._get_next_summary_step()
 
     def get_policy(self, name_behavior_id: str) -> TFPolicy:

ml-agents/mlagents/trainers/sac/trainer.py

         self._check_param_keys()
         self.load = load
         self.seed = seed
-        self.policy: NNPolicy = None  # type: ignore
+        self.policy: TFPolicy = None  # type: ignore
         self.optimizer: SACOptimizer = None  # type: ignore
 
         self.step = 0
             self._stats_reporter.add_stat(stat, np.mean(stat_list))
 
     def add_policy(
-        self, parsed_behavior_id: BehaviorIdentifiers, policy: TFPolicy
+        self, parsed_behavior_id: BehaviorIdentifiers, brain_parameters: BrainParameters
     ) -> None:
         """
         Adds policy to trainer.
                     self.__class__.__name__
                 )
             )
-        if not isinstance(policy, NNPolicy):
-            raise RuntimeError("Non-SACPolicy passed to SACTrainer.add_policy()")
-        self.policy = policy
+        self.policy = self.create_policy(brain_parameters)
-        self.step = policy.get_current_step()
+        self.step = self.policy.get_current_step()
         self.next_summary_step = self._get_next_summary_step()
 
     def get_policy(self, name_behavior_id: str) -> TFPolicy:

ml-agents/mlagents/trainers/trainer/trainer.py

 
     @abc.abstractmethod
     def add_policy(
-        self, parsed_behavior_id: BehaviorIdentifiers, policy: TFPolicy
+        self, parsed_behavior_id: BehaviorIdentifiers, brain_parameters: BrainParameters
     ) -> None:
         """
         Adds policy to trainer.

ml-agents/mlagents/trainers/trainer_controller.py

             trainer = self.trainer_factory.generate(brain_name)
             self.trainers[brain_name] = trainer
 
-        policy = trainer.create_policy(env_manager.external_brains[name_behavior_id])
-        trainer.add_policy(parsed_behavior_id, policy)
+        trainer.add_policy(
+            parsed_behavior_id, env_manager.external_brains[name_behavior_id]
+        )
+        policy = trainer.get_policy(name_behavior_id)
 
         agent_manager = AgentManager(
             policy,