[RLlib] AlgorithmConfig.update_from_dict needs to work for MultiCallbacks.

rllib/algorithms/algorithm_config.py

@@ -600,8 +600,10 @@ def update_from_dict(
             # correct methods to properly `.update()` those from given config dict
             # (to not lose any sub-keys).
             elif key == "callbacks_class":
-                # Resolve possible classpath.
-                value = deserialize_type(value, error=True)
+                # For backward compatibility reasons, only resolve possible
+                # classpath if value is a str type.
+                if isinstance(value, str):
+                    value = deserialize_type(value, error=True)
                 self.callbacks(callbacks_class=value)
             elif key == "env_config":
                 self.environment(env_config=value)
@@ -3142,7 +3144,11 @@ def items(self):
     @staticmethod
     def _serialize_dict(config):
         # Serialize classes to classpaths:
-        config["callbacks"] = serialize_type(config["callbacks"])
+        if isinstance(config.get("callbacks"), type):
+            config["callbacks"] = serialize_type(config["callbacks"])
+        else:
+            # TODO(sven): Figure out how to serialize MultiCallbacks
+            config["callbacks"] = NOT_SERIALIZABLE
         config["sample_collector"] = serialize_type(config["sample_collector"])
         if isinstance(config["env"], type):
             config["env"] = serialize_type(config["env"])

rllib/algorithms/tests/test_algorithm_config.py

@@ -3,6 +3,7 @@
 
 import ray
 from ray.rllib.algorithms.algorithm_config import AlgorithmConfig
+from ray.rllib.algorithms.callbacks import MultiCallbacks
 from ray.rllib.algorithms.ppo import PPO, PPOConfig
 from ray.rllib.algorithms.ppo.torch.ppo_torch_rl_module import PPOTorchRLModule
 from ray.rllib.core.rl_module.rl_module import SingleAgentRLModuleSpec
@@ -34,6 +35,24 @@ def test_running_specific_algo_with_generic_config(self):
         algo.train()
         algo.stop()
 
+    def test_update_from_dict_works_for_multi_callbacks(self):
+        """Test to make sure callbacks config dict works."""
+        config_dict = {
+            "callbacks": MultiCallbacks([])
+        }
+        config = (
+            AlgorithmConfig(algo_class=PPO)
+            .environment("CartPole-v0")
+            .training(lr=0.12345, train_batch_size=3000)
+        )
+        # This should work.
+        config.update_from_dict(config_dict)
+
+        serialized = config.serialize()
+
+        # For now, we don't support serializing MultiCallbacks.
+        self.assertEqual(serialized["callbacks"], "__not_serializable__")
+
     def test_freezing_of_algo_config(self):
         """Tests, whether freezing an AlgorithmConfig actually works as expected."""
         config = (