Avoid Memory Leak in Rollout worker

ISSUE: When we restore checkpoints, we create new nodes in the
Tensorflow graph. This happens when we assign new value (op node) to
RefVariable in GlobalVariableSaver. With every restore the size of TF
graph increases as new nodes are created and old unused nodes are not
removed from the graph. This causes the memory leak in
restore_checkpoint codepath.

FIX: We reset the Tensorflow graph and recreate the Global, Online and
Target networks on every restore. This ensures that the old unused nodes
in TF graph is dropped.

rl_coach/agents/agent.py

@@ -953,6 +953,12 @@ class Agent(AgentInterface):
         self.input_filter.restore_state_from_checkpoint(checkpoint_dir, checkpoint_prefix)
         self.pre_network_filter.restore_state_from_checkpoint(checkpoint_dir, checkpoint_prefix)
 
+        if self.ap.task_parameters.framework_type == Frameworks.tensorflow:
+            import tensorflow as tf
+            tf.reset_default_graph()
+            # Recreate all the networks of the agent
+            self.networks = self.create_networks()
+
         # no output filters currently have an internal state to restore
         # self.output_filter.restore_state_from_checkpoint(checkpoint_dir)
 

rl_coach/graph_managers/graph_manager.py

@@ -150,7 +150,7 @@ class GraphManager(object):
 
         # create a session (it needs to be created after all the graph ops were created)
         self.sess = None
-        self.create_session(task_parameters=task_parameters)
+        self.restore_checkpoint()
 
         self._phase = self.phase = RunPhase.UNDEFINED
 
@@ -261,8 +261,6 @@ class GraphManager(object):
         self.checkpoint_saver = SaverCollection()
         for level in self.level_managers:
             self.checkpoint_saver.update(level.collect_savers())
-        # restore from checkpoint if given
-        self.restore_checkpoint()
 
     def save_graph(self) -> None:
         """
@@ -558,14 +556,20 @@ class GraphManager(object):
             else:
                 checkpoint = get_checkpoint_state(self.task_parameters.checkpoint_restore_dir)
 
+            # As part of this restore, Agent recreates the global, target and online networks
+            [manager.restore_checkpoint(self.task_parameters.checkpoint_restore_dir) for manager in self.level_managers]
+
+            # Recreate the session to use the new TF Graphs
+            self.create_session(self.task_parameters)
+
             if checkpoint is None:
                 screen.warning("No checkpoint to restore in: {}".format(self.task_parameters.checkpoint_restore_dir))
             else:
                 screen.log_title("Loading checkpoint: {}".format(checkpoint.model_checkpoint_path))
-                if not hasattr(self.agent_params.memory, 'memory_backend_params') or self.agent_params.memory.memory_backend_params.run_type != str(RunType.ROLLOUT_WORKER):
                 self.checkpoint_saver.restore(self.sess, checkpoint.model_checkpoint_path)
-
+        else:
-            [manager.restore_checkpoint(self.task_parameters.checkpoint_restore_dir) for manager in self.level_managers]
+            # Create the session to use the new TF Graphs
+            self.create_session(self.task_parameters)
 
     def _get_checkpoint_state_tf(self):
         import tensorflow as tf

rl_coach/tests/graph_managers/test_basic_rl_graph_manager.py

@@ -56,10 +56,6 @@ def test_basic_rl_graph_manager_with_cartpole_dqn_and_repeated_checkpoint_restor
     # graph_manager.save_checkpoint()
     #
     # graph_manager.task_parameters.checkpoint_restore_dir = "./experiments/test/checkpoint"
-    # graph_manager.agent_params.memory.register_var('memory_backend_params',
-    #                                                MemoryBackendParameters(store_type=None,
-    #                                                                        orchestrator_type=None,
-    #                                                                        run_type=str(RunType.ROLLOUT_WORKER)))
     # while True:
     #     graph_manager.restore_checkpoint()
     #     gc.collect()