Custom trainer editor analytics

ml-agents-envs/mlagents_envs/environment.py

@@ -10,6 +10,7 @@
 
 from mlagents_envs.logging_util import get_logger
 from mlagents_envs.side_channel.side_channel import SideChannel
+from mlagents_envs.side_channel import DefaultTrainingAnalyticsSideChannel
 from mlagents_envs.side_channel.side_channel_manager import SideChannelManager
 from mlagents_envs import env_utils
 
@@ -186,6 +187,16 @@ def __init__(
         self._timeout_wait: int = timeout_wait
         self._communicator = self._get_communicator(worker_id, base_port, timeout_wait)
         self._worker_id = worker_id
+        if side_channels is None:
+            side_channels = []
+        default_training_side_channel: Optional[
+            DefaultTrainingAnalyticsSideChannel
+        ] = None
+        if DefaultTrainingAnalyticsSideChannel.CHANNEL_ID not in [
+            _.channel_id for _ in side_channels
+        ]:
+            default_training_side_channel = DefaultTrainingAnalyticsSideChannel()
+            side_channels.append(default_training_side_channel)
         self._side_channel_manager = SideChannelManager(side_channels)
         self._log_folder = log_folder
         self.academy_capabilities: UnityRLCapabilitiesProto = None  # type: ignore
@@ -246,6 +257,8 @@ def __init__(
         self._is_first_message = True
         self._update_behavior_specs(aca_output)
         self.academy_capabilities = aca_params.capabilities
+        if default_training_side_channel is not None:
+            default_training_side_channel.environment_initialized()
 
     @staticmethod
     def _get_communicator(worker_id, base_port, timeout_wait):

ml-agents-envs/mlagents_envs/side_channel/__init__.py

@@ -2,3 +2,6 @@
 from mlagents_envs.side_channel.outgoing_message import OutgoingMessage  # noqa
 
 from mlagents_envs.side_channel.side_channel import SideChannel  # noqa
+from mlagents_envs.side_channel.default_training_analytics_side_channel import (  # noqa
+    DefaultTrainingAnalyticsSideChannel,  # noqa
+)  # noqa

ml-agents-envs/mlagents_envs/side_channel/default_training_analytics_side_channel.py

@@ -0,0 +1,49 @@
+import sys
+import uuid
+import mlagents_envs
+
+from mlagents_envs.exception import UnityCommunicationException
+from mlagents_envs.side_channel import SideChannel, IncomingMessage, OutgoingMessage
+from mlagents_envs.communicator_objects.training_analytics_pb2 import (
+    TrainingEnvironmentInitialized,
+)
+from google.protobuf.any_pb2 import Any
+
+
+class DefaultTrainingAnalyticsSideChannel(SideChannel):
+    """
+    Side channel that sends information about the training to the Unity environment so it can be logged.
+    """
+
+    CHANNEL_ID = uuid.UUID("b664a4a9-d86f-5a5f-95cb-e8353a7e8356")
+
+    def __init__(self) -> None:
+        # >>> uuid.uuid5(uuid.NAMESPACE_URL, "com.unity.ml-agents/TrainingAnalyticsSideChannel")
+        # UUID('b664a4a9-d86f-5a5f-95cb-e8353a7e8356')
+        # We purposefully use the SAME side channel as the TrainingAnalyticsSideChannel
+
+        super().__init__(DefaultTrainingAnalyticsSideChannel.CHANNEL_ID)
+
+    def on_message_received(self, msg: IncomingMessage) -> None:
+        raise UnityCommunicationException(
+            "The DefaultTrainingAnalyticsSideChannel received a message from Unity, "
+            + "this should not have happened."
+        )
+
+    def environment_initialized(self) -> None:
+        # Tuple of (major, minor, patch)
+        vi = sys.version_info
+
+        msg = TrainingEnvironmentInitialized(
+            python_version=f"{vi[0]}.{vi[1]}.{vi[2]}",
+            mlagents_version="Custom",
+            mlagents_envs_version=mlagents_envs.__version__,
+            torch_version="Unknown",
+            torch_device_type="Unknown",
+        )
+        any_message = Any()
+        any_message.Pack(msg)
+
+        env_init_msg = OutgoingMessage()
+        env_init_msg.set_raw_bytes(any_message.SerializeToString())  # type: ignore
+        super().queue_message_to_send(env_init_msg)

ml-agents/mlagents/training_analytics_side_channel.py

@@ -1,12 +1,15 @@
 import sys
 from typing import Optional
-import uuid
 import mlagents_envs
 import mlagents.trainers
 from mlagents import torch_utils
 from mlagents.trainers.settings import RewardSignalType
 from mlagents_envs.exception import UnityCommunicationException
-from mlagents_envs.side_channel import SideChannel, IncomingMessage, OutgoingMessage
+from mlagents_envs.side_channel import (
+    IncomingMessage,
+    OutgoingMessage,
+    DefaultTrainingAnalyticsSideChannel,
+)
 from mlagents_envs.communicator_objects.training_analytics_pb2 import (
     TrainingEnvironmentInitialized,
     TrainingBehaviorInitialized,
@@ -16,21 +19,22 @@
 from mlagents.trainers.settings import TrainerSettings, RunOptions
 
 
-class TrainingAnalyticsSideChannel(SideChannel):
+class TrainingAnalyticsSideChannel(DefaultTrainingAnalyticsSideChannel):
     """
     Side channel that sends information about the training to the Unity environment so it can be logged.
     """
 
     def __init__(self) -> None:
         # >>> uuid.uuid5(uuid.NAMESPACE_URL, "com.unity.ml-agents/TrainingAnalyticsSideChannel")
         # UUID('b664a4a9-d86f-5a5f-95cb-e8353a7e8356')
-        super().__init__(uuid.UUID("b664a4a9-d86f-5a5f-95cb-e8353a7e8356"))
+        # Use the same uuid as the parent side channel
+        super().__init__()
         self.run_options: Optional[RunOptions] = None
 
     def on_message_received(self, msg: IncomingMessage) -> None:
         raise UnityCommunicationException(
             "The TrainingAnalyticsSideChannel received a message from Unity, "
-            + "this should not have happened."
+            "this should not have happened."
         )
 
     def environment_initialized(self, run_options: RunOptions) -> None: