Cherry picking #5283 and #5519 into 2.0.1-verified-patch to fix gRPC for mac m1.

.yamato/training-backcompat-tests.yml

@@ -1,19 +1,19 @@
 
-test_mac_backcompat_2020.1:
-  {% capture editor_version %}2020.1{% endcapture %}
+test_mac_backcompat_2019.4:
+  {% capture editor_version %}2019.4{% endcapture %}
   {% capture csharp_backcompat_version %}1.0.0{% endcapture %}
   # This test has to run on mac because it requires the custom build of tensorflow without AVX
   # Test against 2020.1 because 2020.2 has to run against package version 1.2.0
   name: Test Mac Backcompat Training {{ editor_version }}
   agent:
     type: Unity::VM::osx
-    image: ml-agents/ml-agents-bokken-mac:0.1.4-492264
+    image: ml-agents/ml-agents-bokken-mac:0.1.5-853758
     flavor: b1.small
   variables:
     UNITY_VERSION: {{ editor_version }}
   commands:
     - |
-      python3 -m venv venv && source venv/bin/activate
+      python -m venv venv && source venv/bin/activate
       python -m pip install pyyaml --index-url https://artifactory.prd.it.unity3d.com/artifactory/api/pypi/pypi/simple
       python -m pip install unity-downloader-cli --index-url https://artifactory.prd.it.unity3d.com/artifactory/api/pypi/pypi/simple --upgrade
       unity-downloader-cli -u {{ editor_version }} -c editor --wait --fast
@@ -23,9 +23,9 @@ test_mac_backcompat_2020.1:
       python -u -m ml-agents.tests.yamato.standalone_build_tests --build-target=mac
       python -u -m ml-agents.tests.yamato.training_int_tests --csharp {{ csharp_backcompat_version }}
     - |
-      python3 -m venv venv_old && source venv_old/bin/activate
+      python -m venv venv_old && source venv_old/bin/activate
       python -m pip install pyyaml --index-url https://artifactory.prd.it.unity3d.com/artifactory/api/pypi/pypi/simple
-      python -u -m ml-agents.tests.yamato.training_int_tests --python 0.16.0
+      python -u -m ml-agents.tests.yamato.training_int_tests --python 0.24.0
   triggers:
     cancel_old_ci: true
     recurring:

com.unity.ml-agents/CHANGELOG.md

@@ -9,6 +9,9 @@ and this project adheres to
 ### Minor Changes
 #### com.unity.ml-agents (C#)
 - Upgrade to 2.0.1
+#### ml-agents / ml-agents-envs / gym-unity (Python)
+- Set gym version in gym-unity to gym release 0.20.0
+- Added minimal analytics collection to LL-API (#5511)
 
 ## [2.0.0] - 2021-09-01
 ### Minor Changes

com.unity.ml-agents/Runtime/Communicator/RpcCommunicator.cs

@@ -2,7 +2,7 @@
 #define MLA_SUPPORTED_TRAINING_PLATFORM
 #endif
 
-# if MLA_SUPPORTED_TRAINING_PLATFORM
+#if MLA_SUPPORTED_TRAINING_PLATFORM
 using Grpc.Core;
 #if UNITY_EDITOR
 using UnityEditor;
@@ -50,6 +50,7 @@ internal class RpcCommunicator : ICommunicator
 
         /// The Unity to External client.
         UnityToExternalProto.UnityToExternalProtoClient m_Client;
+        Channel m_Channel;
 
         /// <summary>
         /// Initializes a new instance of the RPCCommunicator class.
@@ -140,6 +141,7 @@ out input
                     Debug.Log($"Unexpected exception when trying to initialize communication: {ex}");
                 }
                 initParametersOut = new UnityRLInitParameters();
+                NotifyQuitAndShutDownChannel();
                 return false;
             }
 
@@ -181,6 +183,8 @@ out input
 
             UpdateEnvironmentWithInput(input.RlInput);
             initParametersOut = initializationInput.RlInitializationInput.ToUnityRLInitParameters();
+            // Be sure to shut down the grpc channel when the application is quitting.
+            Application.quitting += NotifyQuitAndShutDownChannel;
             return true;
 #else
             initParametersOut = new UnityRLInitParameters();
@@ -217,9 +221,9 @@ void UpdateEnvironmentWithInput(UnityRLInputProto rlInput)
         UnityInputProto Initialize(int port, UnityOutputProto unityOutput, out UnityInputProto unityInput)
         {
             m_IsOpen = true;
-            var channel = new Channel($"localhost:{port}", ChannelCredentials.Insecure);
+            m_Channel = new Channel($"localhost:{port}", ChannelCredentials.Insecure);
 
-            m_Client = new UnityToExternalProto.UnityToExternalProtoClient(channel);
+            m_Client = new UnityToExternalProto.UnityToExternalProtoClient(m_Channel);
             var result = m_Client.Exchange(WrapMessage(unityOutput, 200));
             var inputMessage = m_Client.Exchange(WrapMessage(null, 200));
             unityInput = inputMessage.UnityInput;
@@ -229,11 +233,24 @@ UnityInputProto Initialize(int port, UnityOutputProto unityOutput, out UnityInpu
             if (result.Header.Status != 200 || inputMessage.Header.Status != 200)
             {
                 m_IsOpen = false;
-                QuitCommandReceived?.Invoke();
+                NotifyQuitAndShutDownChannel();
             }
             return result.UnityInput;
         }
 
+        void NotifyQuitAndShutDownChannel()
+        {
+            QuitCommandReceived?.Invoke();
+            try
+            {
+                m_Channel.ShutdownAsync().Wait();
+            }
+            catch (Exception)
+            {
+                // do nothing
+            }
+        }
+
 #endregion
 
 #region Destruction
@@ -269,7 +286,7 @@ void SendCommandEvent(CommandProto command)
             {
                 case CommandProto.Quit:
                     {
-                        QuitCommandReceived?.Invoke();
+                        NotifyQuitAndShutDownChannel();
                         return;
                     }
                 case CommandProto.Reset:
@@ -456,7 +473,7 @@ UnityInputProto Exchange(UnityOutputProto unityOutput)
                 // Not sure if the quit command is actually sent when a
                 // non 200 message is received.  Notify that we are indeed
                 // quitting.
-                QuitCommandReceived?.Invoke();
+                NotifyQuitAndShutDownChannel();
                 return message.UnityInput;
             }
             catch (Exception ex)
@@ -488,7 +505,7 @@ UnityInputProto Exchange(UnityOutputProto unityOutput)
                 }
 
                 m_IsOpen = false;
-                QuitCommandReceived?.Invoke();
+                NotifyQuitAndShutDownChannel();
                 return null;
             }
         }

gym-unity/gym_unity/tests/test_gym.py

@@ -248,7 +248,7 @@ def create_mock_vector_steps(specs, num_agents=1, number_visual_observations=0):
     :BehaviorSpecs specs: The BehaviorSpecs for this mock
     :int num_agents: Number of "agents" to imitate in your BatchedStepResult values.
     """
-    obs = [np.array([num_agents * [1, 2, 3]]).reshape(num_agents, 3)]
+    obs = [np.array([num_agents * [1, 2, 3]], dtype=np.float32).reshape(num_agents, 3)]
     if number_visual_observations:
         obs += [
             np.zeros(shape=(num_agents, 8, 8, 3), dtype=np.float32)

gym-unity/setup.py

@@ -38,6 +38,6 @@ def run(self):
     author_email="ML-Agents@unity3d.com",
     url="https://github.com/Unity-Technologies/ml-agents",
     packages=find_packages(),
-    install_requires=["gym", f"mlagents_envs=={VERSION}"],
+    install_requires=["gym==0.20.0", f"mlagents_envs=={VERSION}"],
     cmdclass={"verify": VerifyVersionCommand},
 )

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,
@@ -16,15 +19,16 @@
 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:

ml-agents/setup.py

@@ -74,7 +74,7 @@ def run(self):
         # cattrs 1.1.0 dropped support for python 3.6, but 1.0.0 doesn't work for python 3.9
         # Since there's no version that supports both, we have to draw the line somwehere.
         "cattrs<1.1.0; python_version<'3.8'",
-        "cattrs>=1.1.0; python_version>='3.8'",
+        "cattrs>=1.1.0,<1.7; python_version>='3.8'",
         "attrs>=19.3.0",
         'pypiwin32==223;platform_system=="Windows"',
         "importlib_metadata; python_version<'3.8'",

ml-agents/tests/yamato/training_int_tests.py

@@ -5,7 +5,6 @@
 import sys
 import subprocess
 import time
-from typing import Any
 
 from .yamato_utils import (
     find_executables,
@@ -14,7 +13,6 @@
     run_standalone_build,
     init_venv,
     override_config_file,
-    override_legacy_config_file,
     checkout_csharp_version,
     undo_git_checkout,
 )
@@ -26,7 +24,7 @@ def run_training(python_version: str, csharp_version: str) -> bool:
     print(
         f"Running training with python={python_version or latest} and c#={csharp_version or latest}"
     )
-    output_dir = "models" if python_version else "results"
+    output_dir = "results"
     onnx_file_expected = f"./{output_dir}/{run_id}/3DBall.onnx"
 
     if os.path.exists(onnx_file_expected):
@@ -70,17 +68,11 @@ def run_training(python_version: str, csharp_version: str) -> bool:
     # Copy the default training config but override the max_steps parameter,
     # and reduce the batch_size and buffer_size enough to ensure an update step happens.
     yaml_out = "override.yaml"
-    if python_version:
-        overrides: Any = {"max_steps": 100, "batch_size": 10, "buffer_size": 10}
-        override_legacy_config_file(
-            python_version, "config/trainer_config.yaml", yaml_out, **overrides
-        )
-    else:
-        overrides = {
-            "hyperparameters": {"batch_size": 10, "buffer_size": 10},
-            "max_steps": 100,
-        }
-        override_config_file("config/ppo/3DBall.yaml", yaml_out, overrides)
+    overrides = {
+        "hyperparameters": {"batch_size": 10, "buffer_size": 10},
+        "max_steps": 100,
+    }
+    override_config_file("config/ppo/3DBall.yaml", yaml_out, overrides)
 
     log_output_path = f"{get_base_output_path()}/training.log"
     env_path = os.path.join(get_base_output_path(), standalone_player_path)

ml-agents/tests/yamato/yamato_utils.py

@@ -134,13 +134,10 @@ def init_venv(
     pip_commands = ["--upgrade pip", "--upgrade setuptools"]
     if mlagents_python_version:
         # install from pypi
-        if platform != "darwin":
-            raise RuntimeError("Yamato can only run tensorflow on mac platforms!")
         pip_commands += [
             f"mlagents=={mlagents_python_version}",
             f"gym-unity=={mlagents_python_version}",
             # TODO build these and publish to internal pypi
-            "~/tensorflow_pkg/tensorflow-2.0.0-cp37-cp37m-macosx_10_14_x86_64.whl",
             "tf2onnx==1.6.1",
         ]
     else: