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

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

         brain_name = parsed_behavior_id.brain_name
         self.trainer.export_model(brain_name)
 
-    def create_policy(self, brain_parameters: BrainParameters) -> TFPolicy:
+    def create_policy(
+        self, parsed_behavior_id: BehaviorIdentifiers, brain_parameters: BrainParameters
+    ) -> TFPolicy:
-        """
-        return self.trainer.create_policy(brain_parameters)
-
-    def add_policy(
-        self, parsed_behavior_id: BehaviorIdentifiers, brain_parameters: BrainParameters
-    ) -> None:
-        """
-        Adds policy to trainer. The first policy encountered sets the wrapped
+        The first policy encountered sets the wrapped
-        :param name_behavior_id: Behavior ID that the policy should belong to.
-        :param policy: Policy to associate with name_behavior_id.
-        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 = self.trainer.create_policy(parsed_behavior_id, brain_parameters)
-        self.policies[name_behavior_id] = policy
-
-        self._name_to_parsed_behavior_id[name_behavior_id] = parsed_behavior_id
-        # for saving/swapping snapshots
+        team_id = parsed_behavior_id.team_id
+        self.controller.subscribe_team_id(team_id, self)
-            # 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 = self.trainer.create_policy(
+                parsed_behavior_id, brain_parameters
-
-            # initialize ghost level policy to have the same weights
-
+            self.trainer.add_policy(parsed_behavior_id, internal_trainer_policy)
+        return policy
+
+    def add_policy(
+        self, parsed_behavior_id: BehaviorIdentifiers, policy: TFPolicy
+    ) -> None:
+        """
+        Adds policy to GhostTrainer.
+        :param parsed_behavior_id: Behavior ID that the policy should belong to.
+        :param policy: Policy to associate with name_behavior_id.
+        """
+        name_behavior_id = parsed_behavior_id.behavior_id
+        self._name_to_parsed_behavior_id[name_behavior_id] = parsed_behavior_id
+        self.policies[name_behavior_id] = policy
 
     def get_policy(self, name_behavior_id: str) -> TFPolicy:
         """

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

         self._check_param_keys()
         self.load = load
         self.seed = seed
-        self.policy: TFPolicy = None  # type: ignore
+        self.policy: NNPolicy = None  # type: ignore
 
     def _check_param_keys(self):
         super()._check_param_keys()
                 self._stats_reporter.add_stat(stat, val)
         self._clear_update_buffer()
 
-    def create_policy(self, brain_parameters: BrainParameters) -> TFPolicy:
+    def create_policy(
+        self, parsed_behavior_id: BehaviorIdentifiers, brain_parameters: BrainParameters
+    ) -> TFPolicy:
         """
         Creates a PPO policy to trainers list of policies.
         :param brain_parameters: specifications for policy construction
         return policy
 
     def add_policy(
-        self, parsed_behavior_id: BehaviorIdentifiers, brain_parameters: BrainParameters
+        self, parsed_behavior_id: BehaviorIdentifiers, policy: TFPolicy
     ) -> None:
         """
         Adds policy to trainer.
                     self.__class__.__name__
                 )
             )
-        self.policy = self.create_policy(brain_parameters)
+        if not isinstance(policy, NNPolicy):
+            raise RuntimeError("Non-NNPolicy passed to PPOTrainer.add_policy()")
+        self.policy = policy
-        self.step = self.policy.get_current_step()
+        self.step = 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: TFPolicy = None  # type: ignore
+        self.policy: NNPolicy = None  # type: ignore
         self.optimizer: SACOptimizer = None  # type: ignore
 
         self.step = 0
             self.update_sac_policy()
             self.update_reward_signals()
 
-    def create_policy(self, brain_parameters: BrainParameters) -> TFPolicy:
+    def create_policy(
+        self, parsed_behavior_id: BehaviorIdentifiers, brain_parameters: BrainParameters
+    ) -> TFPolicy:
         policy = NNPolicy(
             self.seed,
             brain_parameters,
             self._stats_reporter.add_stat(stat, np.mean(stat_list))
 
     def add_policy(
-        self, parsed_behavior_id: BehaviorIdentifiers, brain_parameters: BrainParameters
+        self, parsed_behavior_id: BehaviorIdentifiers, policy: TFPolicy
     ) -> None:
         """
         Adds policy to trainer.
                     self.__class__.__name__
                 )
             )
-        self.policy = self.create_policy(brain_parameters)
+        if not isinstance(policy, NNPolicy):
+            raise RuntimeError("Non-SACPolicy passed to SACTrainer.add_policy()")
+        self.policy = policy
-        self.step = self.policy.get_current_step()
+        self.step = 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/tests/test_ghost.py

     trainer_params = dummy_config
     trainer = PPOTrainer(mock_brain.brain_name, 0, trainer_params, True, False, 0, "0")
     trainer.seed = 1
-    policy = trainer.create_policy(mock_brain)
+    policy = trainer.create_policy(mock_brain.brain_name, mock_brain)
-    to_load_policy = trainer.create_policy(mock_brain)
+    to_load_policy = trainer.create_policy(mock_brain.brain_name, mock_brain)
     to_load_policy.create_tf_graph()
     to_load_policy.init_load_weights()
 
     parsed_behavior_id0 = BehaviorIdentifiers.from_name_behavior_id(
         brain_params_team0.brain_name
     )
-    trainer.add_policy(parsed_behavior_id0, brain_params_team0)
+    policy = trainer.create_policy(parsed_behavior_id0, brain_params_team0)
+    trainer.add_policy(parsed_behavior_id0, policy)
     trajectory_queue0 = AgentManagerQueue(brain_params_team0.brain_name)
     trainer.subscribe_trajectory_queue(trajectory_queue0)
 
     )
-    trainer.add_policy(parsed_behavior_id1, brain_params_team1)
+    policy = trainer.create_policy(parsed_behavior_id1, brain_params_team1)
+    trainer.add_policy(parsed_behavior_id1, policy)
     trajectory_queue1 = AgentManagerQueue(brain_params_team1.brain_name)
     trainer.subscribe_trajectory_queue(trajectory_queue1)
 
 
     # First policy encountered becomes policy trained by wrapped PPO
     # This queue should remain empty after swap snapshot
-    trainer.add_policy(parsed_behavior_id0, brain_params_team0)
+    policy = trainer.create_policy(parsed_behavior_id0, brain_params_team0)
+    trainer.add_policy(parsed_behavior_id0, policy)
     policy_queue0 = AgentManagerQueue(brain_params_team0.brain_name)
     trainer.publish_policy_queue(policy_queue0)
 
     )
-    trainer.add_policy(parsed_behavior_id1, brain_params_team1)
+    policy = trainer.create_policy(parsed_behavior_id1, brain_params_team1)
+    trainer.add_policy(parsed_behavior_id1, policy)
     policy_queue1 = AgentManagerQueue(brain_params_team1.brain_name)
     trainer.publish_policy_queue(policy_queue1)
 

ml-agents/mlagents/trainers/tests/test_ppo.py

     trainer_params["reward_signals"]["curiosity"]["encoding_size"] = 128
 
     trainer = PPOTrainer(mock_brain.brain_name, 0, trainer_params, True, False, 0, "0")
-    trainer.add_policy(mock_brain.brain_name, mock_brain)
+    policy = trainer.create_policy(mock_brain.brain_name, mock_brain)
+    trainer.add_policy(mock_brain.brain_name, policy)
     # Test update with sequence length smaller than batch size
     buffer = mb.simulate_rollout(BUFFER_INIT_SAMPLES, mock_brain)
     # Mock out reward signal eval
     dummy_config["summary_path"] = "./summaries/test_trainer_summary"
     dummy_config["model_path"] = "./models/test_trainer_models/TestModel"
     trainer = PPOTrainer(brain_params, 0, dummy_config, True, False, 0, "0")
-    trainer.add_policy(brain_params.brain_name, brain_params)
+    policy = trainer.create_policy(brain_params.brain_name, brain_params)
+    trainer.add_policy(brain_params.brain_name, policy)
     trajectory_queue = AgentManagerQueue("testbrain")
     trainer.subscribe_trajectory_queue(trajectory_queue)
     time_horizon = 15

ml-agents/mlagents/trainers/tests/test_sac.py

     trainer_params["model_path"] = str(tmpdir)
     trainer_params["save_replay_buffer"] = True
     trainer = SACTrainer(mock_brain.brain_name, 1, trainer_params, True, False, 0, 0)
-    trainer.add_policy(mock_brain.brain_name, mock_brain)
-    policy = trainer.get_policy(mock_brain.brain_name)
+    policy = trainer.create_policy(mock_brain.brain_name, mock_brain)
+    trainer.add_policy(mock_brain.brain_name, policy)
 
     trainer.update_buffer = mb.simulate_rollout(BUFFER_INIT_SAMPLES, policy.brain)
     buffer_len = trainer.update_buffer.num_experiences
     trainer2 = SACTrainer(mock_brain.brain_name, 1, trainer_params, True, True, 0, 0)
 
-    trainer2.add_policy(mock_brain.brain_name, mock_brain)
+    policy = trainer2.create_policy(mock_brain.brain_name, mock_brain)
+    trainer2.add_policy(mock_brain.brain_name, policy)
     assert trainer2.update_buffer.num_experiences == buffer_len
 
 
     dummy_config["summary_path"] = "./summaries/test_trainer_summary"
     dummy_config["model_path"] = "./models/test_trainer_models/TestModel"
     trainer = SACTrainer(brain_params, 0, dummy_config, True, False, 0, "0")
-    trainer.add_policy(brain_params.brain_name, brain_params)
+    policy = trainer.create_policy(brain_params.brain_name, brain_params)
+    trainer.add_policy(brain_params.brain_name, policy)
 
     trajectory_queue = AgentManagerQueue("testbrain")
     trainer.subscribe_trajectory_queue(trajectory_queue)

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

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

ml-agents/mlagents/trainers/trainer_controller.py

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