Fix exporting again

docs/Training-ML-Agents.md

 ended.
 
 
+#### Curriculum
+
+To enable curriculum learning, you need to add a `curriculum` sub-section to your environment
+parameter. Here is one example with the environment parameter `my_environment_parameter` :
+
+```yml
+behaviors:
+  BehaviorY:
+    # < Same as above >
+
+# Add this section
+environment_parameters:
+  my_environment_parameter:
+    curriculum:
+      - name: MyFirstLesson # The '-' is important as this is a list
+        completion_criteria:
+          measure: progress
+          behavior: my_behavior
+          signal_smoothing: true
+          min_lesson_length: 100
+          threshold: 0.2
+        value: 0.0
+      - name: MySecondLesson # This is the start of the second lesson
+        completion_criteria:
+          measure: progress
+          behavior: my_behavior
+          signal_smoothing: true
+          min_lesson_length: 100
+          threshold: 0.6
+          require_reset: true
+        value:
+          sampler_type: uniform
+          sampler_parameters:
+            min_value: 4.0
+            max_value: 7.0
+      - name: MyLastLesson
+        value: 8.0
+```
+
+Note that this curriculum __only__ applies to `my_environment_parameter`. The `curriculum` section
+contains a list of `Lessons`. In the example, the lessons are named `MyFirstLesson`, `MySecondLesson`
+and `MyLastLesson`.
+Each `Lesson` has 3 fields :
+
+ - `name` which is a user defined name for the lesson (The name of the lesson will be displayed in
+ the console when the lesson changes)
+ - `completion_criteria` which determines what needs to happen in the simulation before the lesson
+ can be considered complete. When that condition is met, the curriculum moves on to the next
+ `Lesson`. Note that you do not need to specify a `completion_criteria` for the last `Lesson`
+ - `value` which is the value the environment parameter will take during the lesson. Note that this
+ can be a float or a sampler.
+
+ There are the different settings of the `completion_criteria` :
+
+
+| **Setting**         | **Description**                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 |
+| :------------------ | :---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `measure`           | What to measure learning progress, and advancement in lessons by.<br><br> `reward` uses a measure received reward, while `progress` uses the ratio of steps/max_steps.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          |
+| `behavior`        | Specifies which behavior is being tracked. There can be multiple behaviors with different names, each at different points of training. This setting allows the curriculum to track only one of them.                                                                                                                                                                                                                                                                                                                                                                                                                 |
+| `threshold`        | Determines at what point in value of `measure` the lesson should be increased.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  |
+| `min_lesson_length` | The minimum number of episodes that should be completed before the lesson can change. If `measure` is set to `reward`, the average cumulative reward of the last `min_lesson_length` episodes will be used to determine if the lesson should change. Must be nonnegative. <br><br> **Important**: the average reward that is compared to the thresholds is different than the mean reward that is logged to the console. For example, if `min_lesson_length` is `100`, the lesson will increment after the average cumulative reward of the last `100` episodes exceeds the current threshold. The mean reward logged to the console is dictated by the `summary_freq` parameter defined above. |
+| `signal_smoothing`  | Whether to weight the current progress measure by previous values.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              |
+| `require_reset`  | Whether changing lesson requires the environment to reset (default: false)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              |
+##### Training with a Curriculum
+
+Once we have specified our metacurriculum and curricula, we can launch
+`mlagents-learn` to point to the config file containing
+our curricula and PPO will train using Curriculum Learning. For example, to
+train agents in the Wall Jump environment with curriculum learning, we can run:
+
+```sh
+mlagents-learn config/ppo/WallJump_curriculum.yaml --run-id=wall-jump-curriculum
+```
+
+We can then keep track of the current lessons and progresses via TensorBoard. If you've terminated
+the run, you can resume it using `--resume` and lesson progress will start off where it
+ended.
+
+
 ### Training Using Concurrent Unity Instances
 
 In order to run concurrent Unity instances during training, set the number of

docs/Learning-Environment-Examples.md

   head, thighs, shins, feet, arms, forearms and hands.
 - Goal: The agents must move its body toward the goal direction without falling.
   - `WalkerDynamic`- Goal direction is randomized.
-  - `WalkerDynamicVariableSpeed`- Goal direction and walking speed are randomized. 
+  - `WalkerDynamicVariableSpeed`- Goal direction and walking speed are randomized.
   - `WalkerStatic` - Goal direction is always forward.
   - `WalkerStaticVariableSpeed` - Goal direction is always forward. Walking
      speed is randomized

ml-agents-envs/mlagents_envs/communicator_objects/unity_to_external_pb2_grpc.py

 from mlagents_envs.communicator_objects import unity_message_pb2 as mlagents__envs_dot_communicator__objects_dot_unity__message__pb2
 
 
-class UnityToExternalProtoStub(object):
+class UnityToExternalProtoStub:
   # missing associated documentation comment in .proto file
   pass
 
         )
 
 
-class UnityToExternalProtoServicer(object):
+class UnityToExternalProtoServicer:
   # missing associated documentation comment in .proto file
   pass
 

ml-agents/mlagents/trainers/buffer.py

             Adds a list of np.arrays to the end of the list of np.arrays.
             :param data: The np.array list to append.
             """
-            self += list(np.array(data))
+            self += list(np.array(data, dtype=np.float32))
 
         def set(self, data):
             """

ml-agents/mlagents/trainers/cli_utils.py

         action=DetectDefaultStoreTrue,
         help="Forces training using CPU only",
     )
+    argparser.add_argument(
+        "--torch",
+        default=False,
+        action=DetectDefaultStoreTrue,
+        help="(Experimental) Use the PyTorch framework instead of TensorFlow. Install PyTorch "
+        "before using this option",
+    )
 
     eng_conf = argparser.add_argument_group(title="Engine Configuration")
     eng_conf.add_argument(

ml-agents/mlagents/trainers/settings.py

         return _mapping[self]
 
 
+class FrameworkType(Enum):
+    TENSORFLOW: str = "tensorflow"
+    PYTORCH: str = "pytorch"
+
+
 @attr.s(auto_attribs=True)
 class TrainerSettings(ExportableSettings):
     trainer_type: TrainerType = TrainerType.PPO
     threaded: bool = True
     self_play: Optional[SelfPlaySettings] = None
     behavioral_cloning: Optional[BehavioralCloningSettings] = None
+    framework: FrameworkType = FrameworkType.TENSORFLOW
 
     cattr.register_structure_hook(
         Dict[RewardSignalType, RewardSignalSettings], RewardSignalSettings.structure
                     configured_dict["engine_settings"][key] = val
                 else:  # Base options
                     configured_dict[key] = val
-        return RunOptions.from_dict(configured_dict)
+
+        # Apply --torch retroactively
+        final_runoptions = RunOptions.from_dict(configured_dict)
+        if "torch" in DetectDefault.non_default_args:
+            for trainer_set in final_runoptions.behaviors.values():
+                trainer_set.framework = FrameworkType.PYTORCH
+        return final_runoptions
 
     @staticmethod
     def from_dict(options_dict: Dict[str, Any]) -> "RunOptions":

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

         self.vis_obs_size = sum(
             1 for shape in behavior_spec.observation_shapes if len(shape) == 3
         )
+        self.vis_obs_shape = (
+            [shape for shape in behavior_spec.observation_shapes if len(shape) == 3][0]
+            if self.vis_obs_size > 0
+            else None
+        )
         self.use_continuous_act = behavior_spec.is_action_continuous()
         self.num_branches = self.behavior_spec.action_size
         self.previous_action_dict: Dict[str, np.array] = {}

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

 from mlagents.trainers.optimizer import Optimizer
 from mlagents.trainers.buffer import AgentBuffer
 from mlagents.trainers.trainer import Trainer
-from mlagents.trainers.components.reward_signals import RewardSignalResult
+from mlagents.trainers.components.reward_signals import RewardSignalResult, RewardSignal
+from mlagents_envs.base_env import BehaviorSpec
+from mlagents.trainers.policy.policy import Policy
+from mlagents.trainers.policy.tf_policy import TFPolicy
+from mlagents.trainers.behavior_id_utils import BehaviorIdentifiers
-from mlagents.trainers.settings import TrainerSettings
+from mlagents.trainers.settings import TrainerSettings, FrameworkType
+from mlagents.trainers.exception import UnityTrainerException
+
+try:
+    from mlagents.trainers.policy.torch_policy import TorchPolicy
+    from mlagents.trainers.saver.torch_saver import TorchSaver
+except ModuleNotFoundError:
+    TorchPolicy = None  # type: ignore
 
 RewardSignalResults = Dict[str, RewardSignalResult]
 
         self._stats_reporter.add_property(
             StatsPropertyType.HYPERPARAMETERS, self.trainer_settings.as_dict()
         )
+        self.framework = self.trainer_settings.framework
+        logger.debug(f"Using framework {self.framework.value}")
+
-            self.trainer_settings, self.artifact_path, self.load
+            self.framework, self.trainer_settings, self.artifact_path, self.load
         )
 
     def end_episode(self) -> None:
             for agent_id in rewards:
                 rewards[agent_id] = 0
 
-    @staticmethod
-    def create_saver(
-        trainer_settings: TrainerSettings, model_path: str, load: bool
-    ) -> BaseSaver:
-        saver = TFSaver(trainer_settings, model_path, load)
-        return saver
-
     def _update_end_episode_stats(self, agent_id: str, optimizer: Optimizer) -> None:
         for name, rewards in self.collected_rewards.items():
             if name == "environment":
                 self.reward_buffer.appendleft(rewards.get(agent_id, 0))
                 rewards[agent_id] = 0
             else:
-                self.stats_reporter.add_stat(
-                    optimizer.reward_signals[name].stat_name, rewards.get(agent_id, 0)
-                )
+                if isinstance(optimizer.reward_signals[name], RewardSignal):
+                    self.stats_reporter.add_stat(
+                        optimizer.reward_signals[name].stat_name,
+                        rewards.get(agent_id, 0),
+                    )
+                else:
+                    self.stats_reporter.add_stat(
+                        f"Policy/{optimizer.reward_signals[name].name.capitalize()} Reward",
+                        rewards.get(agent_id, 0),
+                    )
                 rewards[agent_id] = 0
 
     def _clear_update_buffer(self) -> None:
         """
         return False
 
+    def create_policy(
+        self, parsed_behavior_id: BehaviorIdentifiers, behavior_spec: BehaviorSpec
+    ) -> Policy:
+        if self.framework == FrameworkType.PYTORCH and TorchPolicy is None:
+            raise UnityTrainerException(
+                "To use the experimental PyTorch backend, install the PyTorch Python package first."
+            )
+        elif self.framework == FrameworkType.PYTORCH:
+            return self.create_torch_policy(parsed_behavior_id, behavior_spec)
+        else:
+            return self.create_tf_policy(parsed_behavior_id, behavior_spec)
+
+    @abc.abstractmethod
+    def create_torch_policy(
+        self, parsed_behavior_id: BehaviorIdentifiers, behavior_spec: BehaviorSpec
+    ) -> TorchPolicy:
+        """
+        Create a Policy object that uses the PyTorch backend.
+        """
+        pass
+
+    @abc.abstractmethod
+    def create_tf_policy(
+        self, parsed_behavior_id: BehaviorIdentifiers, behavior_spec: BehaviorSpec
+    ) -> TFPolicy:
+        """
+        Create a Policy object that uses the TensorFlow backend.
+        """
+        pass
+
+    @staticmethod
+    def create_saver(
+        framework: str, trainer_settings: TrainerSettings, model_path: str, load: bool
+    ) -> BaseSaver:
+        if framework == FrameworkType.PYTORCH:
+            saver = TorchSaver(  # type: ignore
+                trainer_settings, model_path, load
+            )
+        else:
+            saver = TFSaver(  # type: ignore
+                trainer_settings, model_path, load
+            )
+        return saver
+
     def _policy_mean_reward(self) -> Optional[float]:
         """ Returns the mean episode reward for the current policy. """
         rewards = self.cumulative_returns_since_policy_update
             logger.warning(
                 "Trainer has multiple policies, but default behavior only saves the first."
             )
+        elif n_policies == 0:
+            logger.warning("Trainer has no policies, not saving anything.")
+            return
+
-
-        # Copy the checkpointed model files to the final output location
-
         final_checkpoint = attr.evolve(
             model_checkpoint, file_path=f"{self.saver.model_path}.nn"
         )

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

 from mlagents.trainers.trainer.rl_trainer import RLTrainer
 from mlagents.trainers.trajectory import Trajectory, SplitObservations
 from mlagents.trainers.behavior_id_utils import BehaviorIdentifiers
-from mlagents.trainers.settings import TrainerSettings, SACSettings
+from mlagents.trainers.settings import TrainerSettings, SACSettings, FrameworkType
+from mlagents.trainers.components.reward_signals import RewardSignal
+try:
+    from mlagents.trainers.policy.torch_policy import TorchPolicy
+    from mlagents.trainers.sac.optimizer_torch import TorchSACOptimizer
+except ModuleNotFoundError:
+    TorchPolicy = None  # type: ignore
+    TorchSACOptimizer = None  # type: ignore
 
 logger = get_logger(__name__)
 
             agent_buffer_trajectory["environment_rewards"]
         )
         for name, reward_signal in self.optimizer.reward_signals.items():
-            evaluate_result = reward_signal.evaluate_batch(
-                agent_buffer_trajectory
-            ).scaled_reward
+            if isinstance(reward_signal, RewardSignal):
+                evaluate_result = reward_signal.evaluate_batch(
+                    agent_buffer_trajectory
+                ).scaled_reward
+            else:
+                evaluate_result = (
+                    reward_signal.evaluate(agent_buffer_trajectory)
+                    * reward_signal.strength
+                )
             # Report the reward signals
             self.collected_rewards[name][agent_id] += np.sum(evaluate_result)
 
         )
         for name, v in value_estimates.items():
-            self._stats_reporter.add_stat(
-                self.optimizer.reward_signals[name].value_name, np.mean(v)
-            )
+            if isinstance(self.optimizer.reward_signals[name], RewardSignal):
+                self._stats_reporter.add_stat(
+                    self.optimizer.reward_signals[name].value_name, np.mean(v)
+                )
+            else:
+                self._stats_reporter.add_stat(
+                    f"Policy/{self.optimizer.reward_signals[name].name.capitalize()} Value",
+                    np.mean(v),
+                )
 
         # Bootstrap using the last step rather than the bootstrap step if max step is reached.
         # Set last element to duplicate obs and remove dones.
         self._update_reward_signals()
         return policy_was_updated
 
-    def create_policy(
-        self, parsed_behavior_id: BehaviorIdentifiers, behavior_spec: BehaviorSpec
-    ) -> TFPolicy:
-        policy = TFPolicy(
-            self.seed,
-            behavior_spec,
-            self.trainer_settings,
-            tanh_squash=True,
-            reparameterize=True,
-            create_tf_graph=False,
-        )
+    def maybe_load_replay_buffer(self):
         # Load the replay buffer if load
         if self.load and self.checkpoint_replay_buffer:
             try:
                 )
             )
 
+    def create_tf_policy(
+        self, parsed_behavior_id: BehaviorIdentifiers, behavior_spec: BehaviorSpec
+    ) -> TFPolicy:
+        policy = TFPolicy(
+            self.seed,
+            behavior_spec,
+            self.trainer_settings,
+            tanh_squash=True,
+            reparameterize=True,
+            create_tf_graph=False,
+        )
+        self.maybe_load_replay_buffer()
+        return policy
+
+    def create_torch_policy(
+        self, parsed_behavior_id: BehaviorIdentifiers, behavior_spec: BehaviorSpec
+    ) -> TorchPolicy:
+        """
+        Creates a PPO policy to trainers list of policies.
+        :param parsed_behavior_id:
+        :param behavior_spec: specifications for policy construction
+        :return policy
+        """
+        policy = TorchPolicy(
+            self.seed,
+            behavior_spec,
+            self.trainer_settings,
+            condition_sigma_on_obs=True,
+            tanh_squash=True,
+            separate_critic=True,
+        )
+        self.maybe_load_replay_buffer()
         return policy
 
     def _update_sac_policy(self) -> bool:
                 )
                 # Get rewards for each reward
                 for name, signal in self.optimizer.reward_signals.items():
-                    sampled_minibatch[f"{name}_rewards"] = signal.evaluate_batch(
-                        sampled_minibatch
-                    ).scaled_reward
+                    if isinstance(signal, RewardSignal):
+                        sampled_minibatch[f"{name}_rewards"] = signal.evaluate_batch(
+                            sampled_minibatch
+                        ).scaled_reward
+                    else:
+                        sampled_minibatch[f"{name}_rewards"] = (
+                            signal.evaluate(sampled_minibatch) * signal.strength
+                        )
 
                 update_stats = self.optimizer.update(sampled_minibatch, n_sequences)
                 for stat_name, value in update_stats.items():
             reward_signal_minibatches = {}
             for name, signal in self.optimizer.reward_signals.items():
                 logger.debug(f"Updating {name} at step {self.step}")
-                # Some signals don't need a minibatch to be sampled - so we don't!
-                if signal.update_dict:
-                    reward_signal_minibatches[name] = buffer.sample_mini_batch(
-                        self.hyperparameters.batch_size,
-                        sequence_length=self.policy.sequence_length,
-                    )
+                if isinstance(signal, RewardSignal):
+                    # Some signals don't need a minibatch to be sampled - so we don't!
+                    if signal.update_dict:
+                        reward_signal_minibatches[name] = buffer.sample_mini_batch(
+                            self.hyperparameters.batch_size,
+                            sequence_length=self.policy.sequence_length,
+                        )
             update_stats = self.optimizer.update_reward_signals(
                 reward_signal_minibatches, n_sequences
             )
                 self._stats_reporter.add_stat(stat, np.mean(stat_list))
 
     def create_sac_optimizer(self) -> SACOptimizer:
-        return SACOptimizer(cast(TFPolicy, self.policy), self.trainer_settings)
+        if self.framework == FrameworkType.PYTORCH:
+            return TorchSACOptimizer(  # type: ignore
+                cast(TorchPolicy, self.policy), self.trainer_settings  # type: ignore
+            )  # type: ignore
+        else:
+            return SACOptimizer(  # type: ignore
+                cast(TFPolicy, self.policy), self.trainer_settings  # type: ignore
+            )  # type: ignore
 
     def add_policy(
         self, parsed_behavior_id: BehaviorIdentifiers, policy: Policy

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

 
 from mlagents.trainers.trainer.rl_trainer import RLTrainer
 from mlagents.trainers.ppo.trainer import PPOTrainer, discount_rewards
-from mlagents.trainers.ppo.optimizer import PPOOptimizer
+from mlagents.trainers.ppo.optimizer_tf import PPOOptimizer
 from mlagents.trainers.policy.tf_policy import TFPolicy
 from mlagents.trainers.agent_processor import AgentManagerQueue
 from mlagents.trainers.tests import mock_brain as mb

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

 import mlagents.trainers.tests.mock_brain as mb
 from mlagents.trainers.policy.tf_policy import TFPolicy
 from mlagents.trainers.sac.optimizer import SACOptimizer
-from mlagents.trainers.ppo.optimizer import PPOOptimizer
+from mlagents.trainers.ppo.optimizer_tf import PPOOptimizer
 from mlagents.trainers.tests.test_simple_rl import PPO_CONFIG, SAC_CONFIG
 from mlagents.trainers.settings import (
     GAILSettings,

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

         mock_saver.save_checkpoint.side_effect = checkpoint_path
         self.saver = mock_saver
 
-    def create_policy(self):
+    def create_tf_policy(self, parsed_behavior_id, behavior_spec):
+        return mock.Mock()
+
+    def create_torch_policy(self, parsed_behavior_id, behavior_spec):
         return mock.Mock()
 
     def _process_trajectory(self, trajectory):

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

         0, mock_brain, trainer_settings, "test", False, create_tf_graph=False
     )
     optimizer = SACOptimizer(policy, trainer_settings)
-    policy.initialize()
+    optimizer.policy.initialize()
     return optimizer
 
 
     trainer.add_policy(behavior_id, policy)
     trainer.saver.initialize_or_load(policy)
     trainer.optimizer.update = mock.Mock()
+    trainer.saver.initialize_or_load(policy)
     trainer.optimizer.update_reward_signals = mock.Mock()
     trainer.optimizer.update_reward_signals.return_value = {}
     trainer.optimizer.update.return_value = {}

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

             name="old_probabilities",
         )
 
-        # Break old log probs into separate branches
+        # Break old log log_probs into separate branches
         old_log_prob_branches = ModelUtils.break_into_branches(
             self.all_old_log_probs, self.policy.act_size
         )

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

 from mlagents.trainers.trainer.rl_trainer import RLTrainer
 from mlagents.trainers.policy import Policy
 from mlagents.trainers.policy.tf_policy import TFPolicy
-from mlagents.trainers.ppo.optimizer import PPOOptimizer
+from mlagents.trainers.ppo.optimizer_tf import PPOOptimizer
-from mlagents.trainers.settings import TrainerSettings, PPOSettings
+from mlagents.trainers.settings import TrainerSettings, PPOSettings, FrameworkType
+from mlagents.trainers.components.reward_signals import RewardSignal
+
+try:
+    from mlagents.trainers.policy.torch_policy import TorchPolicy
+    from mlagents.trainers.ppo.optimizer_torch import TorchPPOOptimizer
+except ModuleNotFoundError:
+    TorchPolicy = None  # type: ignore
+    TorchPPOOptimizer = None  # type: ignore
 
 
 logger = get_logger(__name__)
             trajectory.next_obs,
             trajectory.done_reached and not trajectory.interrupted,
         )
+
-            self._stats_reporter.add_stat(
-                self.optimizer.reward_signals[name].value_name, np.mean(v)
-            )
+            if isinstance(self.optimizer.reward_signals[name], RewardSignal):
+                self._stats_reporter.add_stat(
+                    self.optimizer.reward_signals[name].value_name, np.mean(v)
+                )
+            else:
+                self._stats_reporter.add_stat(
+                    f"Policy/{self.optimizer.reward_signals[name].name.capitalize()} Value Estimate",
+                    np.mean(v),
+                )
 
         # Evaluate all reward functions
         self.collected_rewards["environment"][agent_id] += np.sum(
-            evaluate_result = reward_signal.evaluate_batch(
-                agent_buffer_trajectory
-            ).scaled_reward
+            if isinstance(reward_signal, RewardSignal):
+                evaluate_result = reward_signal.evaluate_batch(
+                    agent_buffer_trajectory
+                ).scaled_reward
+            else:
+                evaluate_result = (
+                    reward_signal.evaluate(agent_buffer_trajectory)
+                    * reward_signal.strength
+                )
             agent_buffer_trajectory[f"{name}_rewards"].extend(evaluate_result)
             # Report the reward signals
             self.collected_rewards[name][agent_id] += np.sum(evaluate_result)
             local_value_estimates = agent_buffer_trajectory[
                 f"{name}_value_estimates"
             ].get_batch()
+
             local_advantage = get_gae(
                 rewards=local_rewards,
                 value_estimates=local_value_estimates,
         self._clear_update_buffer()
         return True
 
-    def create_policy(
+    def create_tf_policy(
         self, parsed_behavior_id: BehaviorIdentifiers, behavior_spec: BehaviorSpec
     ) -> TFPolicy:
         """
             behavior_spec,
             self.trainer_settings,
             condition_sigma_on_obs=False,  # Faster training for PPO
-            create_tf_graph=False,  # We will create the TF graph in the Optimizer
+        return policy
+    def create_torch_policy(
+        self, parsed_behavior_id: BehaviorIdentifiers, behavior_spec: BehaviorSpec
+    ) -> TorchPolicy:
+        """
+        Creates a PPO policy to trainers list of policies.
+        :param parsed_behavior_id:
+        :param brain_parameters: specifications for policy construction
+        :return policy
+        """
+        policy = TorchPolicy(
+            self.seed,
+            behavior_spec,
+            self.trainer_settings,
+            condition_sigma_on_obs=False,  # Faster training for PPO
+            separate_critic=behavior_spec.is_action_continuous(),
+        )
-        return PPOOptimizer(cast(TFPolicy, self.policy), self.trainer_settings)
+        if self.framework == FrameworkType.PYTORCH:
+            return TorchPPOOptimizer(  # type: ignore
+                cast(TorchPolicy, self.policy), self.trainer_settings  # type: ignore
+            )  # type: ignore
+        else:
+            return PPOOptimizer(  # type: ignore
+                cast(TFPolicy, self.policy), self.trainer_settings  # type: ignore
+            )  # type: ignore
 
     def add_policy(
         self, parsed_behavior_id: BehaviorIdentifiers, policy: Policy
             )
         self.policy = policy
         self.policies[parsed_behavior_id.behavior_id] = policy
+
         self.optimizer = self.create_ppo_optimizer()
         for _reward_signal in self.optimizer.reward_signals.keys():
             self.collected_rewards[_reward_signal] = defaultdict(lambda: 0)

com.unity.ml-agents/Runtime/SensorHelper.cs.meta

 fileFormatVersion: 2
-guid: 11fe037a02b4a483cb9342c3454232cd
+guid: 7c1189c0af42c46f7b533350d49ad3e7
 MonoImporter:
   externalObjects: {}
   serializedVersion: 2

Project/Assets/ML-Agents/Examples/SharedAssets/Scripts/OrientationCubeController.cs.meta

 fileFormatVersion: 2
-guid: fcb7a51f0d5f8404db7b85bd35ecc1fb
+guid: 771e78c5e980e440e8cd19716b55075f
 MonoImporter:
   externalObjects: {}
   serializedVersion: 2

ml-agents/mlagents/trainers/tests/torch/test_layers.py

     linear_layer,
     lstm_layer,
     Initialization,
+    LSTM,
 )
 
 
             assert torch.all(
                 torch.eq(param.data[4:8], torch.ones_like(param.data[4:8]))
             )
+
+
+def test_lstm_class():
+    torch.manual_seed(0)
+    input_size = 12
+    memory_size = 64
+    batch_size = 8
+    seq_len = 16
+    lstm = LSTM(input_size, memory_size)
+
+    assert lstm.memory_size == memory_size
+
+    sample_input = torch.ones((batch_size, seq_len, input_size))
+    sample_memories = torch.ones((1, batch_size, memory_size))
+    out, mem = lstm(sample_input, sample_memories)
+    # Hidden size should be half of memory_size
+    assert out.shape == (batch_size, seq_len, memory_size // 2)
+    assert mem.shape == (1, batch_size, memory_size)

ml-agents/mlagents/trainers/tests/torch/test_networks.py

     if lstm:
         sample_obs = torch.ones((1, network_settings.memory.sequence_length, obs_size))
         memories = torch.ones(
-            (
-                1,
-                network_settings.memory.sequence_length,
-                network_settings.memory.memory_size,
-            )
+            (1, network_settings.memory.sequence_length, actor.memory_size)
         )
     else:
         sample_obs = torch.ones((1, obs_size))

ml-agents/mlagents/trainers/tests/torch/test_utils.py

     masks = torch.tensor([False, False, False, False, False])
     mean = ModelUtils.masked_mean(test_input, masks=masks)
     assert mean == 0.0
+
+    # Make sure it works with 2d arrays of shape (mask_length, N)
+    test_input = torch.tensor([1, 2, 3, 4, 5]).repeat(2, 1).T
+    masks = torch.tensor([False, False, True, True, True])
+    mean = ModelUtils.masked_mean(test_input, masks=masks)
+    assert mean == 4.0

ml-agents/mlagents/trainers/tf/models.py

         :param action_masks: The mask for the logits. Must be of dimension [None x total_number_of_action]
         :param action_size: A list containing the number of possible actions for each branch
         :return: The action output dimension [batch_size, num_branches], the concatenated
-            normalized probs (after softmax)
-        and the concatenated normalized log probs
+            normalized log_probs (after softmax)
+        and the concatenated normalized log log_probs
         """
         branch_masks = ModelUtils.break_into_branches(action_masks, action_size)
         raw_probs = [

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

         pass
 
     @abc.abstractmethod
-    def save_checkpoint(self, brain_name: str, step: int) -> str:
+    def save_checkpoint(self, behavior_name: str, step: int) -> str:
-        :param brain_name: Brain name of brain to be trained
+        :param behavior_name: Behavior name of behavior to be trained
-    def export(self, output_filepath: str, brain_name: str) -> None:
+    def export(self, output_filepath: str, behavior_name: str) -> None:
-        Saves the serialized model, given a path and brain name.
+        Saves the serialized model, given a path and behavior name.
-        :param brain_name: Brain name of brain to be trained.
+        :param behavior_name: Behavior name of behavior to be trained.
         """
         pass
 

ml-agents/mlagents/trainers/torch/layers.py

 import torch
+import abc
+from typing import Tuple
 from enum import Enum
 
 
                         forget_bias
                     )
     return lstm
+
+
+class MemoryModule(torch.nn.Module):
+    @abc.abstractproperty
+    def memory_size(self) -> int:
+        """
+        Size of memory that is required at the start of a sequence.
+        """
+        pass
+
+    @abc.abstractmethod
+    def forward(
+        self, input_tensor: torch.Tensor, memories: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Pass a sequence to the memory module.
+        :input_tensor: Tensor of shape (batch_size, seq_length, size) that represents the input.
+        :memories: Tensor of initial memories.
+        :return: Tuple of output, final memories.
+        """
+        pass
+
+
+class LSTM(MemoryModule):
+    """
+    Memory module that implements LSTM.
+    """
+
+    def __init__(
+        self,
+        input_size: int,
+        memory_size: int,
+        num_layers: int = 1,
+        forget_bias: float = 1.0,
+        kernel_init: Initialization = Initialization.XavierGlorotUniform,
+        bias_init: Initialization = Initialization.Zero,
+    ):
+        super().__init__()
+        # We set hidden size to half of memory_size since the initial memory
+        # will be divided between the hidden state and initial cell state.
+        self.hidden_size = memory_size // 2
+        self.lstm = lstm_layer(
+            input_size,
+            self.hidden_size,
+            num_layers,
+            True,
+            forget_bias,
+            kernel_init,
+            bias_init,
+        )
+
+    @property
+    def memory_size(self) -> int:
+        return 2 * self.hidden_size
+
+    def forward(
+        self, input_tensor: torch.Tensor, memories: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        h0, c0 = torch.split(memories, self.hidden_size, dim=-1)
+        hidden = (h0, c0)
+        lstm_out, hidden_out = self.lstm(input_tensor, hidden)
+        output_mem = torch.cat(hidden_out, dim=-1)
+        return lstm_out, output_mem

ml-agents/mlagents/trainers/torch/networks.py

 from typing import Callable, List, Dict, Tuple, Optional
-import attr
 import abc
 
 import torch
 from mlagents.trainers.settings import NetworkSettings
 from mlagents.trainers.torch.utils import ModelUtils
 from mlagents.trainers.torch.decoders import ValueHeads
-from mlagents.trainers.torch.layers import lstm_layer
+from mlagents.trainers.torch.layers import LSTM
 
 ActivationFunction = Callable[[torch.Tensor], torch.Tensor]
 EncoderFunction = Callable[
         )
 
         if self.use_lstm:
-            self.lstm = lstm_layer(self.h_size, self.m_size // 2, batch_first=True)
+            self.lstm = LSTM(self.h_size, self.m_size)
-            self.lstm = None
+            self.lstm = None  # type: ignore
 
     def update_normalization(self, vec_inputs: List[torch.Tensor]) -> None:
         for vec_input, vec_enc in zip(vec_inputs, self.vector_encoders):
             for n1, n2 in zip(self.vector_encoders, other_network.vector_encoders):
                 n1.copy_normalization(n2)
 
+    @property
+    def memory_size(self) -> int:
+        return self.lstm.memory_size if self.use_lstm else 0
+
     def forward(
         self,
         vec_inputs: List[torch.Tensor],
         sequence_length: int = 1,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
-        vec_encodes = []
+        encodes = []
         for idx, encoder in enumerate(self.vector_encoders):
             vec_input = vec_inputs[idx]
             if actions is not None:
-            vec_encodes.append(hidden)
+            encodes.append(hidden)
-        vis_encodes = []
-            vis_encodes.append(hidden)
+            encodes.append(hidden)
-        if len(vec_encodes) > 0 and len(vis_encodes) > 0:
-            vec_encodes_tensor = torch.stack(vec_encodes, dim=-1).sum(dim=-1)
-            vis_encodes_tensor = torch.stack(vis_encodes, dim=-1).sum(dim=-1)
-            encoding = torch.stack(
-                [vec_encodes_tensor, vis_encodes_tensor], dim=-1
-            ).sum(dim=-1)
-        elif len(vec_encodes) > 0:
-            encoding = torch.stack(vec_encodes, dim=-1).sum(dim=-1)
-        elif len(vis_encodes) > 0:
-            encoding = torch.stack(vis_encodes, dim=-1).sum(dim=-1)
-        else:
+        if len(encodes) == 0:
+        # Constants don't work in Barracuda
+        encoding = encodes[0]
+        if len(encodes) > 1:
+            for _enc in encodes[1:]:
+                encoding += _enc
+
-            memories = torch.split(memories, self.m_size // 2, dim=-1)
-            memories = torch.cat(memories, dim=-1)
         return encoding, memories
 
 
         else:
             encoding_size = network_settings.hidden_units
         self.value_heads = ValueHeads(stream_names, encoding_size, outputs_per_stream)
+
+    @property
+    def memory_size(self) -> int:
+        return self.network_body.memory_size
 
     def forward(
         self,
         """
         pass
 
+    @abc.abstractproperty
+    def memory_size(self):
+        """
+        Returns the size of the memory (same size used as input and output in the other
+        methods) used by this Actor.
+        """
+        pass
+
 
 class SimpleActor(nn.Module, Actor):
     def __init__(
         self.act_type = act_type
         self.act_size = act_size
         self.version_number = torch.nn.Parameter(torch.Tensor([2.0]))
-        self.memory_size = torch.nn.Parameter(torch.Tensor([0]))
         self.is_continuous_int = torch.nn.Parameter(
             torch.Tensor([int(act_type == ActionType.CONTINUOUS)])
         )
             self.encoding_size = network_settings.memory.memory_size // 2
         else:
             self.encoding_size = network_settings.hidden_units
+
         if self.act_type == ActionType.CONTINUOUS:
             self.distribution = GaussianDistribution(
                 self.encoding_size,
             self.distribution = MultiCategoricalDistribution(
                 self.encoding_size, act_size
             )
+
+    @property
+    def memory_size(self) -> int:
+        return self.network_body.memory_size
 
     def update_normalization(self, vector_obs: List[torch.Tensor]) -> None:
         self.network_body.update_normalization(vector_obs)
         )
         action_list = self.sample_action(dists)
         sampled_actions = torch.stack(action_list, dim=-1)
+        if self.act_type == ActionType.CONTINUOUS:
+            log_probs = dists[0].log_prob(sampled_actions)
+        else:
+            log_probs = dists[0].all_log_prob()
-            dists[0].pdf(sampled_actions),
+            log_probs,
-            self.memory_size,
+            torch.Tensor([self.network_body.memory_size]),
             self.is_continuous_int,
             self.act_size_vector,
         )
         # Give the Actor only half the memories. Note we previously validate
         # that memory_size must be a multiple of 4.
         self.use_lstm = network_settings.memory is not None
-        if network_settings.memory is not None:
-            self.half_mem_size = network_settings.memory.memory_size // 2
-            new_memory_settings = attr.evolve(
-                network_settings.memory, memory_size=self.half_mem_size
-            )
-            use_network_settings = attr.evolve(
-                network_settings, memory=new_memory_settings
-            )
-        else:
-            use_network_settings = network_settings
-            self.half_mem_size = 0
-            use_network_settings,
+            network_settings,
             act_type,
             act_size,
             conditional_sigma,
-        self.critic = ValueNetwork(
-            stream_names, observation_shapes, use_network_settings
-        )
+        self.critic = ValueNetwork(stream_names, observation_shapes, network_settings)
+
+    @property
+    def memory_size(self) -> int:
+        return self.network_body.memory_size + self.critic.memory_size
 
     def critic_pass(
         self,
         actor_mem, critic_mem = None, None
         if self.use_lstm:
             # Use only the back half of memories for critic
-            actor_mem, critic_mem = torch.split(memories, self.half_mem_size, -1)
+            actor_mem, critic_mem = torch.split(memories, self.memory_size // 2, -1)
         value_outputs, critic_mem_out = self.critic(
             vec_inputs, vis_inputs, memories=critic_mem, sequence_length=sequence_length
         )
     ) -> Tuple[List[DistInstance], Dict[str, torch.Tensor], torch.Tensor]:
         if self.use_lstm:
             # Use only the back half of memories for critic and actor
-            actor_mem, critic_mem = torch.split(memories, self.half_mem_size, dim=-1)
+            actor_mem, critic_mem = torch.split(memories, self.memory_size // 2, dim=-1)
         else:
             critic_mem = None
             actor_mem = None
 class GlobalSteps(nn.Module):
     def __init__(self):
         super().__init__()
-        self.global_step = torch.Tensor([0])
+        self.__global_step = nn.Parameter(torch.Tensor([0]), requires_grad=False)
+
+    @property
+    def current_step(self):
+        return int(self.__global_step.item())
+
+    @current_step.setter
+    def current_step(self, value):
+        self.__global_step[:] = value
-        self.global_step += value
+        self.__global_step += value
 
 
 class LearningRate(nn.Module):

ml-agents/mlagents/trainers/torch/utils.py

         :param tensor: Tensor which needs mean computation.
         :param masks: Boolean tensor of masks with same dimension as tensor.
         """
-        return (tensor * masks).sum() / torch.clamp(masks.float().sum(), min=1.0)
+        return (tensor.T * masks).sum() / torch.clamp(
+            (torch.ones_like(tensor.T) * masks).float().sum(), min=1.0
+        )

Project/Assets/csc.rsp

+-warnaserror+
+-warnaserror-:618

utils/validate_release_links.py

+#!/usr/bin/env python3
+
+import ast
+import sys
+import os
+import re
+import subprocess
+from typing import List, Optional, Pattern
+
+RELEASE_PATTERN = re.compile(r"release_[0-9]+(_docs)*")
+TRAINER_INIT_FILE = "ml-agents/mlagents/trainers/__init__.py"
+
+# Filename -> regex list to allow specific lines.
+# To allow everything in the file, use None for the value
+ALLOW_LIST = {
+    # Previous release table
+    "README.md": re.compile(r"\*\*Release [0-9]+\*\*"),
+    "docs/Versioning.md": None,
+    "com.unity.ml-agents/CHANGELOG.md": None,
+    "utils/make_readme_table.py": None,
+    "utils/validate_release_links.py": None,
+}
+
+
+def test_pattern():
+    # Just some sanity check that the regex works as expected.
+    assert RELEASE_PATTERN.search(
+        "https://github.com/Unity-Technologies/ml-agents/blob/release_4_docs/Food.md"
+    )
+    assert RELEASE_PATTERN.search(
+        "https://github.com/Unity-Technologies/ml-agents/blob/release_4/Foo.md"
+    )
+    assert RELEASE_PATTERN.search(
+        "git clone --branch release_4 https://github.com/Unity-Technologies/ml-agents.git"
+    )
+    assert RELEASE_PATTERN.search(
+        "https://github.com/Unity-Technologies/ml-agents/blob/release_123_docs/Foo.md"
+    )
+    assert RELEASE_PATTERN.search(
+        "https://github.com/Unity-Technologies/ml-agents/blob/release_123/Foo.md"
+    )
+    assert not RELEASE_PATTERN.search(
+        "https://github.com/Unity-Technologies/ml-agents/blob/latest_release/docs/Foo.md"
+    )
+    print("tests OK!")
+
+
+def git_ls_files() -> List[str]:
+    """
+    Run "git ls-files" and return a list with one entry per line.
+    This returns the list of all files tracked by git.
+    """
+    return subprocess.check_output(["git", "ls-files"], universal_newlines=True).split(
+        "\n"
+    )
+
+
+def get_release_tag() -> Optional[str]:
+    """
+    Returns the release tag for the mlagents python package.
+    This will be None on the master branch.
+    :return:
+    """
+    with open(TRAINER_INIT_FILE) as f:
+        for line in f:
+            if "__release_tag__" in line:
+                lhs, equals_string, rhs = line.strip().partition(" = ")
+                # Evaluate the right hand side of the expression
+                return ast.literal_eval(rhs)
+    # If we couldn't find the release tag, raise an exception
+    # (since we can't return None here)
+    raise RuntimeError("Can't determine release tag")
+
+
+def check_file(filename: str, global_allow_pattern: Pattern) -> List[str]:
+    """
+    Validate a single file and return any offending lines.
+    """
+    bad_lines = []
+    with open(filename) as f:
+        for line in f:
+            if not RELEASE_PATTERN.search(line):
+                continue
+
+            if global_allow_pattern.search(line):
+                continue
+
+            if filename in ALLOW_LIST:
+                if ALLOW_LIST[filename] is None or ALLOW_LIST[filename].search(line):
+                    continue
+
+            bad_lines.append(f"{filename}: {line.strip()}")
+    return bad_lines
+
+
+def check_all_files(allow_pattern: Pattern) -> List[str]:
+    """
+    Validate all files tracked by git.
+    :param allow_pattern:
+    """
+    bad_lines = []
+    file_types = {".py", ".md", ".cs"}
+    for file_name in git_ls_files():
+        if "localized" in file_name or os.path.splitext(file_name)[1] not in file_types:
+            continue
+        bad_lines += check_file(file_name, allow_pattern)
+    return bad_lines
+
+
+def main():
+    release_tag = get_release_tag()
+    if not release_tag:
+        print("Release tag is None, exiting")
+        sys.exit(0)
+
+    print(f"Release tag: {release_tag}")
+    allow_pattern = re.compile(f"{release_tag}(_docs)*")
+    bad_lines = check_all_files(allow_pattern)
+    if bad_lines:
+        print(
+            f"Found lines referring to previous release. Either update the files, or add an exclusion to {__file__}"
+        )
+        for line in bad_lines:
+            print(line)
+
+    sys.exit(1 if bad_lines else 0)
+
+
+if __name__ == "__main__":
+    if "--test" in sys.argv:
+        test_pattern()
+    main()

ml-agents/mlagents/trainers/optimizer/torch_optimizer.py

+from typing import Dict, Optional, Tuple, List
+import torch
+import numpy as np
+
+from mlagents.trainers.buffer import AgentBuffer
+from mlagents.trainers.trajectory import SplitObservations
+from mlagents.trainers.torch.components.bc.module import BCModule
+from mlagents.trainers.torch.components.reward_providers import create_reward_provider
+
+from mlagents.trainers.policy.torch_policy import TorchPolicy
+from mlagents.trainers.optimizer import Optimizer
+from mlagents.trainers.settings import TrainerSettings
+from mlagents.trainers.torch.utils import ModelUtils
+
+
+class TorchOptimizer(Optimizer):  # pylint: disable=W0223
+    def __init__(self, policy: TorchPolicy, trainer_settings: TrainerSettings):
+        super().__init__()
+        self.policy = policy
+        self.trainer_settings = trainer_settings
+        self.update_dict: Dict[str, torch.Tensor] = {}
+        self.value_heads: Dict[str, torch.Tensor] = {}
+        self.memory_in: torch.Tensor = None
+        self.memory_out: torch.Tensor = None
+        self.m_size: int = 0
+        self.global_step = torch.tensor(0)
+        self.bc_module: Optional[BCModule] = None
+        self.create_reward_signals(trainer_settings.reward_signals)
+        if trainer_settings.behavioral_cloning is not None:
+            self.bc_module = BCModule(
+                self.policy,
+                trainer_settings.behavioral_cloning,
+                policy_learning_rate=trainer_settings.hyperparameters.learning_rate,
+                default_batch_size=trainer_settings.hyperparameters.batch_size,
+                default_num_epoch=3,
+            )
+
+    def update(self, batch: AgentBuffer, num_sequences: int) -> Dict[str, float]:
+        pass
+
+    def create_reward_signals(self, reward_signal_configs):
+        """
+        Create reward signals
+        :param reward_signal_configs: Reward signal config.
+        """
+        for reward_signal, settings in reward_signal_configs.items():
+            # Name reward signals by string in case we have duplicates later
+            self.reward_signals[reward_signal.value] = create_reward_provider(
+                reward_signal, self.policy.behavior_spec, settings
+            )
+
+    def get_trajectory_value_estimates(
+        self, batch: AgentBuffer, next_obs: List[np.ndarray], done: bool
+    ) -> Tuple[Dict[str, np.ndarray], Dict[str, float]]:
+        vector_obs = [ModelUtils.list_to_tensor(batch["vector_obs"])]
+        if self.policy.use_vis_obs:
+            visual_obs = []
+            for idx, _ in enumerate(
+                self.policy.actor_critic.network_body.visual_encoders
+            ):
+                visual_ob = ModelUtils.list_to_tensor(batch["visual_obs%d" % idx])
+                visual_obs.append(visual_ob)
+        else:
+            visual_obs = []
+
+        memory = torch.zeros([1, 1, self.policy.m_size])
+
+        vec_vis_obs = SplitObservations.from_observations(next_obs)
+        next_vec_obs = [
+            ModelUtils.list_to_tensor(vec_vis_obs.vector_observations).unsqueeze(0)
+        ]
+        next_vis_obs = [
+            ModelUtils.list_to_tensor(_vis_ob).unsqueeze(0)
+            for _vis_ob in vec_vis_obs.visual_observations
+        ]
+
+        value_estimates, next_memory = self.policy.actor_critic.critic_pass(
+            vector_obs, visual_obs, memory, sequence_length=batch.num_experiences
+        )
+
+        next_value_estimate, _ = self.policy.actor_critic.critic_pass(
+            next_vec_obs, next_vis_obs, next_memory, sequence_length=1
+        )
+
+        for name, estimate in value_estimates.items():
+            value_estimates[name] = estimate.detach().cpu().numpy()
+            next_value_estimate[name] = next_value_estimate[name].detach().cpu().numpy()
+
+        if done:
+            for k in next_value_estimate:
+                if not self.reward_signals[k].ignore_done:
+                    next_value_estimate[k] = 0.0
+
+        return value_estimates, next_value_estimate

ml-agents/mlagents/trainers/policy/torch_policy.py

+from typing import Any, Dict, List, Tuple, Optional
+import numpy as np
+import torch
+
+from mlagents.trainers.action_info import ActionInfo
+from mlagents.trainers.behavior_id_utils import get_global_agent_id
+from mlagents.trainers.policy import Policy
+from mlagents_envs.base_env import DecisionSteps, BehaviorSpec
+from mlagents_envs.timers import timed
+
+from mlagents.trainers.settings import TrainerSettings
+from mlagents.trainers.trajectory import SplitObservations
+from mlagents.trainers.torch.networks import (
+    SharedActorCritic,
+    SeparateActorCritic,
+    GlobalSteps,
+)
+from mlagents.trainers.torch.utils import ModelUtils
+
+EPSILON = 1e-7  # Small value to avoid divide by zero
+
+
+class TorchPolicy(Policy):
+    def __init__(
+        self,
+        seed: int,
+        behavior_spec: BehaviorSpec,
+        trainer_settings: TrainerSettings,
+        tanh_squash: bool = False,
+        reparameterize: bool = False,
+        separate_critic: bool = True,
+        condition_sigma_on_obs: bool = True,
+    ):
+        """
+        Policy that uses a multilayer perceptron to map the observations to actions. Could
+        also use a CNN to encode visual input prior to the MLP. Supports discrete and
+        continuous action spaces, as well as recurrent networks.
+        :param seed: Random seed.
+        :param brain: Assigned BrainParameters object.
+        :param trainer_settings: Defined training parameters.
+        :param load: Whether a pre-trained model will be loaded or a new one created.
+        :param tanh_squash: Whether to use a tanh function on the continuous output,
+        or a clipped output.
+        :param reparameterize: Whether we are using the resampling trick to update the policy
+        in continuous output.
+        """
+        super().__init__(
+            seed,
+            behavior_spec,
+            trainer_settings,
+            tanh_squash,
+            reparameterize,
+            condition_sigma_on_obs,
+        )
+        self.global_step = (
+            GlobalSteps()
+        )  # could be much simpler if TorchPolicy is nn.Module
+        self.grads = None
+
+        torch.set_default_tensor_type(torch.FloatTensor)
+
+        reward_signal_configs = trainer_settings.reward_signals
+        reward_signal_names = [key.value for key, _ in reward_signal_configs.items()]
+
+        self.stats_name_to_update_name = {
+            "Losses/Value Loss": "value_loss",
+            "Losses/Policy Loss": "policy_loss",
+        }
+        if separate_critic:
+            ac_class = SeparateActorCritic
+        else:
+            ac_class = SharedActorCritic
+        self.actor_critic = ac_class(
+            observation_shapes=self.behavior_spec.observation_shapes,
+            network_settings=trainer_settings.network_settings,
+            act_type=behavior_spec.action_type,
+            act_size=self.act_size,
+            stream_names=reward_signal_names,
+            conditional_sigma=self.condition_sigma_on_obs,
+            tanh_squash=tanh_squash,
+        )
+        # Save the m_size needed for export
+        self._export_m_size = self.m_size
+        # m_size needed for training is determined by network, not trainer settings
+        self.m_size = self.actor_critic.memory_size
+
+        self.actor_critic.to("cpu")
+
+    @property
+    def export_memory_size(self) -> int:
+        """
+        Returns the memory size of the exported ONNX policy. This only includes the memory
+        of the Actor and not any auxillary networks.
+        """
+        return self._export_m_size
+
+    def _split_decision_step(
+        self, decision_requests: DecisionSteps
+    ) -> Tuple[SplitObservations, np.ndarray]:
+        vec_vis_obs = SplitObservations.from_observations(decision_requests.obs)
+        mask = None
+        if not self.use_continuous_act:
+            mask = torch.ones([len(decision_requests), np.sum(self.act_size)])
+            if decision_requests.action_mask is not None:
+                mask = torch.as_tensor(
+                    1 - np.concatenate(decision_requests.action_mask, axis=1)
+                )
+        return vec_vis_obs, mask
+
+    def update_normalization(self, vector_obs: np.ndarray) -> None:
+        """
+        If this policy normalizes vector observations, this will update the norm values in the graph.
+        :param vector_obs: The vector observations to add to the running estimate of the distribution.
+        """
+        vector_obs = [torch.as_tensor(vector_obs)]
+        if self.use_vec_obs and self.normalize:
+            self.actor_critic.update_normalization(vector_obs)
+
+    @timed
+    def sample_actions(
+        self,
+        vec_obs: List[torch.Tensor],
+        vis_obs: List[torch.Tensor],
+        masks: Optional[torch.Tensor] = None,
+        memories: Optional[torch.Tensor] = None,
+        seq_len: int = 1,
+        all_log_probs: bool = False,
+    ) -> Tuple[
+        torch.Tensor, torch.Tensor, torch.Tensor, Dict[str, torch.Tensor], torch.Tensor
+    ]:
+        """
+        :param all_log_probs: Returns (for discrete actions) a tensor of log probs, one for each action.
+        """
+        dists, value_heads, memories = self.actor_critic.get_dist_and_value(
+            vec_obs, vis_obs, masks, memories, seq_len
+        )
+        action_list = self.actor_critic.sample_action(dists)
+        log_probs, entropies, all_logs = ModelUtils.get_probs_and_entropy(
+            action_list, dists
+        )
+        actions = torch.stack(action_list, dim=-1)
+        if self.use_continuous_act:
+            actions = actions[:, :, 0]
+        else:
+            actions = actions[:, 0, :]
+
+        return (
+            actions,
+            all_logs if all_log_probs else log_probs,
+            entropies,
+            value_heads,
+            memories,
+        )
+
+    def evaluate_actions(
+        self,
+        vec_obs: torch.Tensor,
+        vis_obs: torch.Tensor,
+        actions: torch.Tensor,
+        masks: Optional[torch.Tensor] = None,
+        memories: Optional[torch.Tensor] = None,
+        seq_len: int = 1,
+    ) -> Tuple[torch.Tensor, torch.Tensor, Dict[str, torch.Tensor]]:
+        dists, value_heads, _ = self.actor_critic.get_dist_and_value(
+            vec_obs, vis_obs, masks, memories, seq_len
+        )
+        action_list = [actions[..., i] for i in range(actions.shape[-1])]
+        log_probs, entropies, _ = ModelUtils.get_probs_and_entropy(action_list, dists)
+
+        return log_probs, entropies, value_heads
+
+    @timed
+    def evaluate(
+        self, decision_requests: DecisionSteps, global_agent_ids: List[str]
+    ) -> Dict[str, Any]:
+        """
+        Evaluates policy for the agent experiences provided.
+        :param global_agent_ids:
+        :param decision_requests: DecisionStep object containing inputs.
+        :return: Outputs from network as defined by self.inference_dict.
+        """
+        vec_vis_obs, masks = self._split_decision_step(decision_requests)
+        vec_obs = [torch.as_tensor(vec_vis_obs.vector_observations)]
+        vis_obs = [
+            torch.as_tensor(vis_ob) for vis_ob in vec_vis_obs.visual_observations
+        ]
+        memories = torch.as_tensor(self.retrieve_memories(global_agent_ids)).unsqueeze(
+            0
+        )
+
+        run_out = {}
+        with torch.no_grad():
+            action, log_probs, entropy, value_heads, memories = self.sample_actions(
+                vec_obs, vis_obs, masks=masks, memories=memories
+            )
+        run_out["action"] = action.detach().cpu().numpy()
+        run_out["pre_action"] = action.detach().cpu().numpy()
+        # Todo - make pre_action difference
+        run_out["log_probs"] = log_probs.detach().cpu().numpy()
+        run_out["entropy"] = entropy.detach().cpu().numpy()
+        run_out["value_heads"] = {
+            name: t.detach().cpu().numpy() for name, t in value_heads.items()
+        }
+        run_out["value"] = np.mean(list(run_out["value_heads"].values()), 0)
+        run_out["learning_rate"] = 0.0
+        if self.use_recurrent:
+            run_out["memory_out"] = memories.detach().cpu().numpy().squeeze(0)
+        return run_out
+
+    def get_action(
+        self, decision_requests: DecisionSteps, worker_id: int = 0
+    ) -> ActionInfo:
+        """
+        Decides actions given observations information, and takes them in environment.
+        :param worker_id:
+        :param decision_requests: A dictionary of brain names and BrainInfo from environment.
+        :return: an ActionInfo containing action, memories, values and an object
+        to be passed to add experiences
+        """
+        if len(decision_requests) == 0:
+            return ActionInfo.empty()
+
+        global_agent_ids = [
+            get_global_agent_id(worker_id, int(agent_id))
+            for agent_id in decision_requests.agent_id
+        ]  # For 1-D array, the iterator order is correct.
+
+        run_out = self.evaluate(
+            decision_requests, global_agent_ids
+        )  # pylint: disable=assignment-from-no-return
+        self.save_memories(global_agent_ids, run_out.get("memory_out"))
+        return ActionInfo(
+            action=run_out.get("action"),
+            value=run_out.get("value"),
+            outputs=run_out,
+            agent_ids=list(decision_requests.agent_id),
+        )
+
+    @property
+    def use_vis_obs(self):
+        return self.vis_obs_size > 0
+
+    @property
+    def use_vec_obs(self):
+        return self.vec_obs_size > 0
+
+    def get_current_step(self):
+        """
+        Gets current model step.
+        :return: current model step.
+        """
+        return self.global_step.current_step
+
+    def set_step(self, step: int) -> int:
+        """
+        Sets current model step to step without creating additional ops.
+        :param step: Step to set the current model step to.
+        :return: The step the model was set to.
+        """
+        self.global_step.current_step = step
+        return step
+
+    def increment_step(self, n_steps):
+        """
+        Increments model step.
+        """
+        self.global_step.increment(n_steps)
+        return self.get_current_step()
+
+    def load_weights(self, values: List[np.ndarray]) -> None:
+        pass
+
+    def init_load_weights(self) -> None:
+        pass
+
+    def get_weights(self) -> List[np.ndarray]:
+        return []
+
+    def get_modules(self):
+        return {"Policy": self.actor_critic, "global_step": self.global_step}

ml-agents/mlagents/trainers/sac/optimizer_torch.py

+import numpy as np
+from typing import Dict, List, Mapping, cast, Tuple, Optional
+import torch
+from torch import nn
+import attr
+
+from mlagents_envs.logging_util import get_logger
+from mlagents_envs.base_env import ActionType
+from mlagents.trainers.optimizer.torch_optimizer import TorchOptimizer
+from mlagents.trainers.policy.torch_policy import TorchPolicy
+from mlagents.trainers.settings import NetworkSettings
+from mlagents.trainers.torch.networks import ValueNetwork
+from mlagents.trainers.torch.utils import ModelUtils
+from mlagents.trainers.buffer import AgentBuffer
+from mlagents_envs.timers import timed
+from mlagents.trainers.exception import UnityTrainerException
+from mlagents.trainers.settings import TrainerSettings, SACSettings
+
+EPSILON = 1e-6  # Small value to avoid divide by zero
+
+logger = get_logger(__name__)
+
+
+class TorchSACOptimizer(TorchOptimizer):
+    class PolicyValueNetwork(nn.Module):
+        def __init__(
+            self,
+            stream_names: List[str],
+            observation_shapes: List[Tuple[int, ...]],
+            network_settings: NetworkSettings,
+            act_type: ActionType,
+            act_size: List[int],
+        ):
+            super().__init__()
+            if act_type == ActionType.CONTINUOUS:
+                num_value_outs = 1
+                num_action_ins = sum(act_size)
+            else:
+                num_value_outs = sum(act_size)
+                num_action_ins = 0
+            self.q1_network = ValueNetwork(
+                stream_names,
+                observation_shapes,
+                network_settings,
+                num_action_ins,
+                num_value_outs,
+            )
+            self.q2_network = ValueNetwork(
+                stream_names,
+                observation_shapes,
+                network_settings,
+                num_action_ins,
+                num_value_outs,
+            )
+
+        def forward(
+            self,
+            vec_inputs: List[torch.Tensor],
+            vis_inputs: List[torch.Tensor],
+            actions: Optional[torch.Tensor] = None,
+            memories: Optional[torch.Tensor] = None,
+            sequence_length: int = 1,
+        ) -> Tuple[Dict[str, torch.Tensor], Dict[str, torch.Tensor]]:
+            q1_out, _ = self.q1_network(
+                vec_inputs,
+                vis_inputs,
+                actions=actions,
+                memories=memories,
+                sequence_length=sequence_length,
+            )
+            q2_out, _ = self.q2_network(
+                vec_inputs,
+                vis_inputs,
+                actions=actions,
+                memories=memories,
+                sequence_length=sequence_length,
+            )
+            return q1_out, q2_out
+
+    def __init__(self, policy: TorchPolicy, trainer_params: TrainerSettings):
+        super().__init__(policy, trainer_params)
+        hyperparameters: SACSettings = cast(SACSettings, trainer_params.hyperparameters)
+        self.tau = hyperparameters.tau
+        self.init_entcoef = hyperparameters.init_entcoef
+
+        self.policy = policy
+        self.act_size = policy.act_size
+        policy_network_settings = policy.network_settings
+
+        self.tau = hyperparameters.tau
+        self.burn_in_ratio = 0.0
+
+        # Non-exposed SAC parameters
+        self.discrete_target_entropy_scale = 0.2  # Roughly equal to e-greedy 0.05
+        self.continuous_target_entropy_scale = 1.0
+
+        self.stream_names = list(self.reward_signals.keys())
+        # Use to reduce "survivor bonus" when using Curiosity or GAIL.
+        self.gammas = [_val.gamma for _val in trainer_params.reward_signals.values()]
+        self.use_dones_in_backup = {
+            name: int(not self.reward_signals[name].ignore_done)
+            for name in self.stream_names
+        }
+
+        # Critics should have 1/2 of the memory of the policy
+        critic_memory = policy_network_settings.memory
+        if critic_memory is not None:
+            critic_memory = attr.evolve(
+                critic_memory, memory_size=critic_memory.memory_size // 2
+            )
+        value_network_settings = attr.evolve(
+            policy_network_settings, memory=critic_memory
+        )
+
+        self.value_network = TorchSACOptimizer.PolicyValueNetwork(
+            self.stream_names,
+            self.policy.behavior_spec.observation_shapes,
+            value_network_settings,
+            self.policy.behavior_spec.action_type,
+            self.act_size,
+        )
+
+        self.target_network = ValueNetwork(
+            self.stream_names,
+            self.policy.behavior_spec.observation_shapes,
+            value_network_settings,
+        )
+        self.soft_update(self.policy.actor_critic.critic, self.target_network, 1.0)
+
+        self._log_ent_coef = torch.nn.Parameter(
+            torch.log(torch.as_tensor([self.init_entcoef] * len(self.act_size))),
+            requires_grad=True,
+        )
+        if self.policy.use_continuous_act:
+            self.target_entropy = torch.as_tensor(
+                -1
+                * self.continuous_target_entropy_scale
+                * np.prod(self.act_size[0]).astype(np.float32)
+            )
+        else:
+            self.target_entropy = [
+                self.discrete_target_entropy_scale * np.log(i).astype(np.float32)
+                for i in self.act_size
+            ]
+
+        policy_params = list(self.policy.actor_critic.network_body.parameters()) + list(
+            self.policy.actor_critic.distribution.parameters()
+        )
+        value_params = list(self.value_network.parameters()) + list(
+            self.policy.actor_critic.critic.parameters()
+        )
+
+        logger.debug("value_vars")
+        for param in value_params:
+            logger.debug(param.shape)
+        logger.debug("policy_vars")
+        for param in policy_params:
+            logger.debug(param.shape)
+
+        self.decay_learning_rate = ModelUtils.DecayedValue(
+            hyperparameters.learning_rate_schedule,
+            hyperparameters.learning_rate,
+            1e-10,
+            self.trainer_settings.max_steps,
+        )
+        self.policy_optimizer = torch.optim.Adam(
+            policy_params, lr=hyperparameters.learning_rate
+        )
+        self.value_optimizer = torch.optim.Adam(
+            value_params, lr=hyperparameters.learning_rate
+        )
+        self.entropy_optimizer = torch.optim.Adam(
+            [self._log_ent_coef], lr=hyperparameters.learning_rate
+        )
+
+    def sac_q_loss(
+        self,
+        q1_out: Dict[str, torch.Tensor],
+        q2_out: Dict[str, torch.Tensor],
+        target_values: Dict[str, torch.Tensor],
+        dones: torch.Tensor,
+        rewards: Dict[str, torch.Tensor],
+        loss_masks: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        q1_losses = []
+        q2_losses = []
+        # Multiple q losses per stream
+        for i, name in enumerate(q1_out.keys()):
+            q1_stream = q1_out[name].squeeze()
+            q2_stream = q2_out[name].squeeze()
+            with torch.no_grad():
+                q_backup = rewards[name] + (
+                    (1.0 - self.use_dones_in_backup[name] * dones)
+                    * self.gammas[i]
+                    * target_values[name]
+                )
+            _q1_loss = 0.5 * ModelUtils.masked_mean(
+                torch.nn.functional.mse_loss(q_backup, q1_stream), loss_masks
+            )
+            _q2_loss = 0.5 * ModelUtils.masked_mean(
+                torch.nn.functional.mse_loss(q_backup, q2_stream), loss_masks
+            )
+
+            q1_losses.append(_q1_loss)
+            q2_losses.append(_q2_loss)
+        q1_loss = torch.mean(torch.stack(q1_losses))
+        q2_loss = torch.mean(torch.stack(q2_losses))
+        return q1_loss, q2_loss
+
+    def soft_update(self, source: nn.Module, target: nn.Module, tau: float) -> None:
+        for source_param, target_param in zip(source.parameters(), target.parameters()):
+            target_param.data.copy_(
+                target_param.data * (1.0 - tau) + source_param.data * tau
+            )
+
+    def sac_value_loss(
+        self,
+        log_probs: torch.Tensor,
+        values: Dict[str, torch.Tensor],
+        q1p_out: Dict[str, torch.Tensor],
+        q2p_out: Dict[str, torch.Tensor],
+        loss_masks: torch.Tensor,
+        discrete: bool,
+    ) -> torch.Tensor:
+        min_policy_qs = {}
+        with torch.no_grad():
+            _ent_coef = torch.exp(self._log_ent_coef)
+        for name in values.keys():
+            if not discrete:
+                min_policy_qs[name] = torch.min(q1p_out[name], q2p_out[name])
+            else:
+                action_probs = log_probs.exp()
+                _branched_q1p = ModelUtils.break_into_branches(
+                    q1p_out[name] * action_probs, self.act_size
+                )
+                _branched_q2p = ModelUtils.break_into_branches(
+                    q2p_out[name] * action_probs, self.act_size
+                )
+                _q1p_mean = torch.mean(
+                    torch.stack(
+                        [torch.sum(_br, dim=1, keepdim=True) for _br in _branched_q1p]
+                    ),
+                    dim=0,
+                )
+                _q2p_mean = torch.mean(
+                    torch.stack(
+                        [torch.sum(_br, dim=1, keepdim=True) for _br in _branched_q2p]
+                    ),
+                    dim=0,
+                )
+
+                min_policy_qs[name] = torch.min(_q1p_mean, _q2p_mean)
+
+        value_losses = []
+        if not discrete:
+            for name in values.keys():
+                with torch.no_grad():
+                    v_backup = min_policy_qs[name] - torch.sum(
+                        _ent_coef * log_probs, dim=1
+                    )
+                value_loss = 0.5 * ModelUtils.masked_mean(
+                    torch.nn.functional.mse_loss(values[name], v_backup), loss_masks
+                )
+                value_losses.append(value_loss)
+        else:
+            branched_per_action_ent = ModelUtils.break_into_branches(
+                log_probs * log_probs.exp(), self.act_size
+            )
+            # We have to do entropy bonus per action branch
+            branched_ent_bonus = torch.stack(
+                [
+                    torch.sum(_ent_coef[i] * _lp, dim=1, keepdim=True)
+                    for i, _lp in enumerate(branched_per_action_ent)
+                ]
+            )
+            for name in values.keys():
+                with torch.no_grad():
+                    v_backup = min_policy_qs[name] - torch.mean(
+                        branched_ent_bonus, axis=0
+                    )
+                value_loss = 0.5 * ModelUtils.masked_mean(
+                    torch.nn.functional.mse_loss(values[name], v_backup.squeeze()),
+                    loss_masks,
+                )
+                value_losses.append(value_loss)
+        value_loss = torch.mean(torch.stack(value_losses))
+        if torch.isinf(value_loss).any() or torch.isnan(value_loss).any():
+            raise UnityTrainerException("Inf found")
+        return value_loss
+
+    def sac_policy_loss(
+        self,
+        log_probs: torch.Tensor,
+        q1p_outs: Dict[str, torch.Tensor],
+        loss_masks: torch.Tensor,
+        discrete: bool,
+    ) -> torch.Tensor:
+        _ent_coef = torch.exp(self._log_ent_coef)
+        mean_q1 = torch.mean(torch.stack(list(q1p_outs.values())), axis=0)
+        if not discrete:
+            mean_q1 = mean_q1.unsqueeze(1)
+            batch_policy_loss = torch.mean(_ent_coef * log_probs - mean_q1, dim=1)
+            policy_loss = ModelUtils.masked_mean(batch_policy_loss, loss_masks)
+        else:
+            action_probs = log_probs.exp()
+            branched_per_action_ent = ModelUtils.break_into_branches(
+                log_probs * action_probs, self.act_size
+            )
+            branched_q_term = ModelUtils.break_into_branches(
+                mean_q1 * action_probs, self.act_size
+            )
+            branched_policy_loss = torch.stack(
+                [
+                    torch.sum(_ent_coef[i] * _lp - _qt, dim=1, keepdim=True)
+                    for i, (_lp, _qt) in enumerate(
+                        zip(branched_per_action_ent, branched_q_term)
+                    )
+                ]
+            )
+            batch_policy_loss = torch.squeeze(branched_policy_loss)
+        policy_loss = torch.mean(loss_masks * batch_policy_loss)
+        return policy_loss
+
+    def sac_entropy_loss(
+        self, log_probs: torch.Tensor, loss_masks: torch.Tensor, discrete: bool
+    ) -> torch.Tensor:
+        if not discrete:
+            with torch.no_grad():
+                target_current_diff = torch.sum(log_probs + self.target_entropy, dim=1)
+            entropy_loss = -torch.mean(
+                self._log_ent_coef * loss_masks * target_current_diff
+            )
+        else:
+            with torch.no_grad():
+                branched_per_action_ent = ModelUtils.break_into_branches(
+                    log_probs * log_probs.exp(), self.act_size
+                )
+                target_current_diff_branched = torch.stack(
+                    [
+                        torch.sum(_lp, axis=1, keepdim=True) + _te
+                        for _lp, _te in zip(
+                            branched_per_action_ent, self.target_entropy
+                        )
+                    ],
+                    axis=1,
+                )
+                target_current_diff = torch.squeeze(
+                    target_current_diff_branched, axis=2
+                )
+            entropy_loss = -1 * ModelUtils.masked_mean(
+                torch.mean(self._log_ent_coef * target_current_diff, axis=1), loss_masks
+            )
+
+        return entropy_loss
+
+    def _condense_q_streams(
+        self, q_output: Dict[str, torch.Tensor], discrete_actions: torch.Tensor
+    ) -> Dict[str, torch.Tensor]:
+        condensed_q_output = {}
+        onehot_actions = ModelUtils.actions_to_onehot(discrete_actions, self.act_size)
+        for key, item in q_output.items():
+            branched_q = ModelUtils.break_into_branches(item, self.act_size)
+            only_action_qs = torch.stack(
+                [
+                    torch.sum(_act * _q, dim=1, keepdim=True)
+                    for _act, _q in zip(onehot_actions, branched_q)
+                ]
+            )
+
+            condensed_q_output[key] = torch.mean(only_action_qs, dim=0)
+        return condensed_q_output
+
+    @timed
+    def update(self, batch: AgentBuffer, num_sequences: int) -> Dict[str, float]:
+        """
+        Updates model using buffer.
+        :param num_sequences: Number of trajectories in batch.
+        :param batch: Experience mini-batch.
+        :param update_target: Whether or not to update target value network
+        :param reward_signal_batches: Minibatches to use for updating the reward signals,
+            indexed by name. If none, don't update the reward signals.
+        :return: Output from update process.
+        """
+        rewards = {}
+        for name in self.reward_signals:
+            rewards[name] = ModelUtils.list_to_tensor(batch[f"{name}_rewards"])
+
+        vec_obs = [ModelUtils.list_to_tensor(batch["vector_obs"])]
+        next_vec_obs = [ModelUtils.list_to_tensor(batch["next_vector_in"])]
+        act_masks = ModelUtils.list_to_tensor(batch["action_mask"])
+        if self.policy.use_continuous_act:
+            actions = ModelUtils.list_to_tensor(batch["actions"]).unsqueeze(-1)
+        else:
+            actions = ModelUtils.list_to_tensor(batch["actions"], dtype=torch.long)
+
+        memories_list = [
+            ModelUtils.list_to_tensor(batch["memory"][i])
+            for i in range(0, len(batch["memory"]), self.policy.sequence_length)
+        ]
+        # LSTM shouldn't have sequence length <1, but stop it from going out of the index if true.
+        offset = 1 if self.policy.sequence_length > 1 else 0
+        next_memories_list = [
+            ModelUtils.list_to_tensor(
+                batch["memory"][i][self.policy.m_size // 2 :]
+            )  # only pass value part of memory to target network
+            for i in range(offset, len(batch["memory"]), self.policy.sequence_length)
+        ]
+
+        if len(memories_list) > 0:
+            memories = torch.stack(memories_list).unsqueeze(0)
+            next_memories = torch.stack(next_memories_list).unsqueeze(0)
+        else:
+            memories = None
+            next_memories = None
+        # Q network memories are 0'ed out, since we don't have them during inference.
+        q_memories = (
+            torch.zeros_like(next_memories) if next_memories is not None else None
+        )
+
+        vis_obs: List[torch.Tensor] = []
+        next_vis_obs: List[torch.Tensor] = []
+        if self.policy.use_vis_obs:
+            vis_obs = []
+            for idx, _ in enumerate(
+                self.policy.actor_critic.network_body.visual_encoders
+            ):
+                vis_ob = ModelUtils.list_to_tensor(batch["visual_obs%d" % idx])
+                vis_obs.append(vis_ob)
+                next_vis_ob = ModelUtils.list_to_tensor(
+                    batch["next_visual_obs%d" % idx]
+                )
+                next_vis_obs.append(next_vis_ob)
+
+        # Copy normalizers from policy
+        self.value_network.q1_network.network_body.copy_normalization(
+            self.policy.actor_critic.network_body
+        )
+        self.value_network.q2_network.network_body.copy_normalization(
+            self.policy.actor_critic.network_body
+        )
+        self.target_network.network_body.copy_normalization(
+            self.policy.actor_critic.network_body
+        )
+        (
+            sampled_actions,
+            log_probs,
+            entropies,
+            sampled_values,
+            _,
+        ) = self.policy.sample_actions(
+            vec_obs,
+            vis_obs,
+            masks=act_masks,
+            memories=memories,
+            seq_len=self.policy.sequence_length,
+            all_log_probs=not self.policy.use_continuous_act,
+        )
+        if self.policy.use_continuous_act:
+            squeezed_actions = actions.squeeze(-1)
+            q1p_out, q2p_out = self.value_network(
+                vec_obs,
+                vis_obs,
+                sampled_actions,
+                memories=q_memories,
+                sequence_length=self.policy.sequence_length,
+            )
+            q1_out, q2_out = self.value_network(
+                vec_obs,
+                vis_obs,
+                squeezed_actions,
+                memories=q_memories,
+                sequence_length=self.policy.sequence_length,
+            )
+            q1_stream, q2_stream = q1_out, q2_out
+        else:
+            with torch.no_grad():
+                q1p_out, q2p_out = self.value_network(
+                    vec_obs,
+                    vis_obs,
+                    memories=q_memories,
+                    sequence_length=self.policy.sequence_length,
+                )
+            q1_out, q2_out = self.value_network(
+                vec_obs,
+                vis_obs,
+                memories=q_memories,
+                sequence_length=self.policy.sequence_length,
+            )
+            q1_stream = self._condense_q_streams(q1_out, actions)
+            q2_stream = self._condense_q_streams(q2_out, actions)
+
+        with torch.no_grad():
+            target_values, _ = self.target_network(
+                next_vec_obs,
+                next_vis_obs,
+                memories=next_memories,
+                sequence_length=self.policy.sequence_length,
+            )
+        masks = ModelUtils.list_to_tensor(batch["masks"], dtype=torch.bool)
+        use_discrete = not self.policy.use_continuous_act
+        dones = ModelUtils.list_to_tensor(batch["done"])
+
+        q1_loss, q2_loss = self.sac_q_loss(
+            q1_stream, q2_stream, target_values, dones, rewards, masks
+        )
+        value_loss = self.sac_value_loss(
+            log_probs, sampled_values, q1p_out, q2p_out, masks, use_discrete
+        )
+        policy_loss = self.sac_policy_loss(log_probs, q1p_out, masks, use_discrete)
+        entropy_loss = self.sac_entropy_loss(log_probs, masks, use_discrete)
+
+        total_value_loss = q1_loss + q2_loss + value_loss
+
+        decay_lr = self.decay_learning_rate.get_value(self.policy.get_current_step())
+        ModelUtils.update_learning_rate(self.policy_optimizer, decay_lr)
+        self.policy_optimizer.zero_grad()
+        policy_loss.backward()
+        self.policy_optimizer.step()
+
+        ModelUtils.update_learning_rate(self.value_optimizer, decay_lr)
+        self.value_optimizer.zero_grad()
+        total_value_loss.backward()
+        self.value_optimizer.step()
+
+        ModelUtils.update_learning_rate(self.entropy_optimizer, decay_lr)
+        self.entropy_optimizer.zero_grad()
+        entropy_loss.backward()
+        self.entropy_optimizer.step()
+
+        # Update target network
+        self.soft_update(self.policy.actor_critic.critic, self.target_network, self.tau)
+        update_stats = {
+            "Losses/Policy Loss": abs(policy_loss.detach().cpu().numpy()),
+            "Losses/Value Loss": value_loss.detach().cpu().numpy(),
+            "Losses/Q1 Loss": q1_loss.detach().cpu().numpy(),
+            "Losses/Q2 Loss": q2_loss.detach().cpu().numpy(),
+            "Policy/Entropy Coeff": torch.exp(self._log_ent_coef)
+            .detach()
+            .cpu()
+            .numpy(),
+            "Policy/Learning Rate": decay_lr,
+        }
+
+        for signal in self.reward_signals.values():
+            signal.update(batch)
+
+        return update_stats
+
+    def update_reward_signals(
+        self, reward_signal_minibatches: Mapping[str, AgentBuffer], num_sequences: int
+    ) -> Dict[str, float]:
+        return {}
+
+    def get_modules(self):
+        return {
+            "Optimizer:value_network": self.value_network,
+            "Optimizer:target_network": self.target_network,
+            "Optimizer:policy_optimizer": self.policy_optimizer,
+            "Optimizer:value_optimizer": self.value_optimizer,
+            "Optimizer:entropy_optimizer": self.entropy_optimizer,
+        }

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

+import pytest
+
+from mlagents.trainers.tf.models import ModelUtils
+from mlagents.tf_utils import tf
+from mlagents_envs.base_env import BehaviorSpec, ActionType
+
+
+def create_behavior_spec(num_visual, num_vector, vector_size):
+    behavior_spec = BehaviorSpec(
+        [(84, 84, 3)] * int(num_visual) + [(vector_size,)] * int(num_vector),
+        ActionType.DISCRETE,
+        (1,),
+    )
+    return behavior_spec
+
+
+@pytest.mark.parametrize("num_visual", [1, 2, 4])
+@pytest.mark.parametrize("num_vector", [1, 2, 4])
+def test_create_input_placeholders(num_vector, num_visual):
+    vec_size = 8
+    name_prefix = "test123"
+    bspec = create_behavior_spec(num_visual, num_vector, vec_size)
+    vec_in, vis_in = ModelUtils.create_input_placeholders(
+        bspec.observation_shapes, name_prefix=name_prefix
+    )
+
+    assert isinstance(vis_in, list)
+    assert len(vis_in) == num_visual
+    assert isinstance(vec_in, tf.Tensor)
+    assert vec_in.get_shape().as_list()[1] == num_vector * 8
+
+    # Check names contain prefix and vis shapes are correct
+    for _vis in vis_in:
+        assert _vis.get_shape().as_list() == [None, 84, 84, 3]
+        assert _vis.name.startswith(name_prefix)
+    assert vec_in.name.startswith(name_prefix)

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

+import pytest
+from unittest import mock
+import os
+import unittest
+import tempfile
+
+import numpy as np
+from mlagents.tf_utils import tf
+from mlagents.trainers.saver.tf_saver import TFSaver
+from mlagents.trainers import __version__
+from mlagents.trainers.settings import TrainerSettings
+from mlagents.trainers.policy.tf_policy import TFPolicy
+from mlagents.trainers.tests import mock_brain as mb
+from mlagents.trainers.tests.test_nn_policy import create_policy_mock
+from mlagents.trainers.ppo.optimizer_tf import PPOOptimizer
+
+
+def test_register(tmp_path):
+    trainer_params = TrainerSettings()
+    saver = TFSaver(trainer_params, tmp_path)
+
+    opt = mock.Mock(spec=PPOOptimizer)
+    saver.register(opt)
+    assert saver.policy is None
+
+    trainer_params = TrainerSettings()
+    policy = create_policy_mock(trainer_params)
+    saver.register(policy)
+    assert saver.policy is not None
+
+
+class ModelVersionTest(unittest.TestCase):
+    def test_version_compare(self):
+        # Test write_stats
+        with self.assertLogs("mlagents.trainers", level="WARNING") as cm:
+            trainer_params = TrainerSettings()
+            mock_path = tempfile.mkdtemp()
+            policy = create_policy_mock(trainer_params)
+            saver = TFSaver(trainer_params, mock_path)
+            saver.register(policy)
+
+            saver._check_model_version(
+                "0.0.0"
+            )  # This is not the right version for sure
+            # Assert that 1 warning has been thrown with incorrect version
+            assert len(cm.output) == 1
+            saver._check_model_version(__version__)  # This should be the right version
+            # Assert that no additional warnings have been thrown wth correct ver
+            assert len(cm.output) == 1
+
+
+def test_load_save(tmp_path):
+    path1 = os.path.join(tmp_path, "runid1")
+    path2 = os.path.join(tmp_path, "runid2")
+    trainer_params = TrainerSettings()
+    policy = create_policy_mock(trainer_params)
+    saver = TFSaver(trainer_params, path1)
+    saver.register(policy)
+    saver.initialize_or_load(policy)
+    policy.set_step(2000)
+
+    mock_brain_name = "MockBrain"
+    saver.save_checkpoint(mock_brain_name, 2000)
+    assert len(os.listdir(tmp_path)) > 0
+
+    # Try load from this path
+    saver = TFSaver(trainer_params, path1, load=True)
+    policy2 = create_policy_mock(trainer_params)
+    saver.register(policy2)
+    saver.initialize_or_load(policy2)
+    _compare_two_policies(policy, policy2)
+    assert policy2.get_current_step() == 2000
+
+    # Try initialize from path 1
+    trainer_params.init_path = path1
+    saver = TFSaver(trainer_params, path2)
+    policy3 = create_policy_mock(trainer_params)
+    saver.register(policy3)
+    saver.initialize_or_load(policy3)
+
+    _compare_two_policies(policy2, policy3)
+    # Assert that the steps are 0.
+    assert policy3.get_current_step() == 0
+
+
+def _compare_two_policies(policy1: TFPolicy, policy2: TFPolicy) -> None:
+    """
+    Make sure two policies have the same output for the same input.
+    """
+    decision_step, _ = mb.create_steps_from_behavior_spec(
+        policy1.behavior_spec, num_agents=1
+    )
+    run_out1 = policy1.evaluate(decision_step, list(decision_step.agent_id))
+    run_out2 = policy2.evaluate(decision_step, list(decision_step.agent_id))
+
+    np.testing.assert_array_equal(run_out2["log_probs"], run_out1["log_probs"])
+
+
+@pytest.mark.parametrize("discrete", [True, False], ids=["discrete", "continuous"])
+@pytest.mark.parametrize("visual", [True, False], ids=["visual", "vector"])
+@pytest.mark.parametrize("rnn", [True, False], ids=["rnn", "no_rnn"])
+def test_checkpoint_conversion(tmpdir, rnn, visual, discrete):
+    tf.reset_default_graph()
+    dummy_config = TrainerSettings()
+    model_path = os.path.join(tmpdir, "Mock_Brain")
+    policy = create_policy_mock(
+        dummy_config, use_rnn=rnn, use_discrete=discrete, use_visual=visual
+    )
+    trainer_params = TrainerSettings()
+    saver = TFSaver(trainer_params, model_path)
+    saver.register(policy)
+    saver.save_checkpoint("Mock_Brain", 100)
+    assert os.path.isfile(model_path + "/Mock_Brain-100.nn")

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

+from typing import Dict, cast
+import torch
+
+from mlagents.trainers.buffer import AgentBuffer
+
+from mlagents_envs.timers import timed
+from mlagents.trainers.policy.torch_policy import TorchPolicy
+from mlagents.trainers.optimizer.torch_optimizer import TorchOptimizer
+from mlagents.trainers.settings import TrainerSettings, PPOSettings
+from mlagents.trainers.torch.utils import ModelUtils
+
+
+class TorchPPOOptimizer(TorchOptimizer):
+    def __init__(self, policy: TorchPolicy, trainer_settings: TrainerSettings):
+        """
+        Takes a Policy and a Dict of trainer parameters and creates an Optimizer around the policy.
+        The PPO optimizer has a value estimator and a loss function.
+        :param policy: A TFPolicy object that will be updated by this PPO Optimizer.
+        :param trainer_params: Trainer parameters dictionary that specifies the
+        properties of the trainer.
+        """
+        # Create the graph here to give more granular control of the TF graph to the Optimizer.
+
+        super().__init__(policy, trainer_settings)
+        params = list(self.policy.actor_critic.parameters())
+        self.hyperparameters: PPOSettings = cast(
+            PPOSettings, trainer_settings.hyperparameters
+        )
+        self.decay_learning_rate = ModelUtils.DecayedValue(
+            self.hyperparameters.learning_rate_schedule,
+            self.hyperparameters.learning_rate,
+            1e-10,
+            self.trainer_settings.max_steps,
+        )
+        self.decay_epsilon = ModelUtils.DecayedValue(
+            self.hyperparameters.learning_rate_schedule,
+            self.hyperparameters.epsilon,
+            0.1,
+            self.trainer_settings.max_steps,
+        )
+        self.decay_beta = ModelUtils.DecayedValue(
+            self.hyperparameters.learning_rate_schedule,
+            self.hyperparameters.beta,
+            1e-5,
+            self.trainer_settings.max_steps,
+        )
+
+        self.optimizer = torch.optim.Adam(
+            params, lr=self.trainer_settings.hyperparameters.learning_rate
+        )
+        self.stats_name_to_update_name = {
+            "Losses/Value Loss": "value_loss",
+            "Losses/Policy Loss": "policy_loss",
+        }
+
+        self.stream_names = list(self.reward_signals.keys())
+
+    def ppo_value_loss(
+        self,
+        values: Dict[str, torch.Tensor],
+        old_values: Dict[str, torch.Tensor],
+        returns: Dict[str, torch.Tensor],
+        epsilon: float,
+        loss_masks: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Evaluates value loss for PPO.
+        :param values: Value output of the current network.
+        :param old_values: Value stored with experiences in buffer.
+        :param returns: Computed returns.
+        :param epsilon: Clipping value for value estimate.
+        :param loss_mask: Mask for losses. Used with LSTM to ignore 0'ed out experiences.
+        """
+        value_losses = []
+        for name, head in values.items():
+            old_val_tensor = old_values[name]
+            returns_tensor = returns[name]
+            clipped_value_estimate = old_val_tensor + torch.clamp(
+                head - old_val_tensor, -1 * epsilon, epsilon
+            )
+            v_opt_a = (returns_tensor - head) ** 2
+            v_opt_b = (returns_tensor - clipped_value_estimate) ** 2
+            value_loss = ModelUtils.masked_mean(torch.max(v_opt_a, v_opt_b), loss_masks)
+            value_losses.append(value_loss)
+        value_loss = torch.mean(torch.stack(value_losses))
+        return value_loss
+
+    def ppo_policy_loss(
+        self,
+        advantages: torch.Tensor,
+        log_probs: torch.Tensor,
+        old_log_probs: torch.Tensor,
+        loss_masks: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Evaluate PPO policy loss.
+        :param advantages: Computed advantages.
+        :param log_probs: Current policy probabilities
+        :param old_log_probs: Past policy probabilities
+        :param loss_masks: Mask for losses. Used with LSTM to ignore 0'ed out experiences.
+        """
+        advantage = advantages.unsqueeze(-1)
+
+        decay_epsilon = self.hyperparameters.epsilon
+
+        r_theta = torch.exp(log_probs - old_log_probs)
+        p_opt_a = r_theta * advantage
+        p_opt_b = (
+            torch.clamp(r_theta, 1.0 - decay_epsilon, 1.0 + decay_epsilon) * advantage
+        )
+        policy_loss = -1 * ModelUtils.masked_mean(
+            torch.min(p_opt_a, p_opt_b), loss_masks
+        )
+        return policy_loss
+
+    @timed
+    def update(self, batch: AgentBuffer, num_sequences: int) -> Dict[str, float]:
+        """
+        Performs update on model.
+        :param batch: Batch of experiences.
+        :param num_sequences: Number of sequences to process.
+        :return: Results of update.
+        """
+        # Get decayed parameters
+        decay_lr = self.decay_learning_rate.get_value(self.policy.get_current_step())
+        decay_eps = self.decay_epsilon.get_value(self.policy.get_current_step())
+        decay_bet = self.decay_beta.get_value(self.policy.get_current_step())
+        returns = {}
+        old_values = {}
+        for name in self.reward_signals:
+            old_values[name] = ModelUtils.list_to_tensor(
+                batch[f"{name}_value_estimates"]
+            )
+            returns[name] = ModelUtils.list_to_tensor(batch[f"{name}_returns"])
+
+        vec_obs = [ModelUtils.list_to_tensor(batch["vector_obs"])]
+        act_masks = ModelUtils.list_to_tensor(batch["action_mask"])
+        if self.policy.use_continuous_act:
+            actions = ModelUtils.list_to_tensor(batch["actions"]).unsqueeze(-1)
+        else:
+            actions = ModelUtils.list_to_tensor(batch["actions"], dtype=torch.long)
+
+        memories = [
+            ModelUtils.list_to_tensor(batch["memory"][i])
+            for i in range(0, len(batch["memory"]), self.policy.sequence_length)
+        ]
+        if len(memories) > 0:
+            memories = torch.stack(memories).unsqueeze(0)
+
+        if self.policy.use_vis_obs:
+            vis_obs = []
+            for idx, _ in enumerate(
+                self.policy.actor_critic.network_body.visual_encoders
+            ):
+                vis_ob = ModelUtils.list_to_tensor(batch["visual_obs%d" % idx])
+                vis_obs.append(vis_ob)
+        else:
+            vis_obs = []
+        log_probs, entropy, values = self.policy.evaluate_actions(
+            vec_obs,
+            vis_obs,
+            masks=act_masks,
+            actions=actions,
+            memories=memories,
+            seq_len=self.policy.sequence_length,
+        )
+        loss_masks = ModelUtils.list_to_tensor(batch["masks"], dtype=torch.bool)
+        value_loss = self.ppo_value_loss(
+            values, old_values, returns, decay_eps, loss_masks
+        )
+        policy_loss = self.ppo_policy_loss(
+            ModelUtils.list_to_tensor(batch["advantages"]),
+            log_probs,
+            ModelUtils.list_to_tensor(batch["action_probs"]),
+            loss_masks,
+        )
+        loss = (
+            policy_loss
+            + 0.5 * value_loss
+            - decay_bet * ModelUtils.masked_mean(entropy, loss_masks)
+        )
+
+        # Set optimizer learning rate
+        ModelUtils.update_learning_rate(self.optimizer, decay_lr)
+        self.optimizer.zero_grad()
+        loss.backward()
+
+        self.optimizer.step()
+        update_stats = {
+            "Losses/Policy Loss": abs(policy_loss.detach().cpu().numpy()),
+            "Losses/Value Loss": value_loss.detach().cpu().numpy(),
+            "Policy/Learning Rate": decay_lr,
+            "Policy/Epsilon": decay_eps,
+            "Policy/Beta": decay_bet,
+        }
+
+        for reward_provider in self.reward_signals.values():
+            update_stats.update(reward_provider.update(batch))
+
+        return update_stats
+
+    def get_modules(self):
+        return {"Optimizer": self.optimizer}

com.unity.ml-agents.extensions/Runtime/Sensors/ArticulationBodyPoseExtractor.cs.meta

+fileFormatVersion: 2
+guid: 11fe037a02b4a483cb9342c3454232cd
+MonoImporter:
+  externalObjects: {}
+  serializedVersion: 2
+  defaultReferences: []
+  executionOrder: 0
+  icon: {instanceID: 0}
+  userData: 
+  assetBundleName: 
+  assetBundleVariant: 

com.unity.ml-agents.extensions/Runtime/Sensors/PhysicsSensorSettings.cs.meta

+fileFormatVersion: 2
+guid: fcb7a51f0d5f8404db7b85bd35ecc1fb
+MonoImporter:
+  externalObjects: {}
+  serializedVersion: 2
+  defaultReferences: []
+  executionOrder: 0
+  icon: {instanceID: 0}
+  userData: 
+  assetBundleName: 
+  assetBundleVariant: 

com.unity.ml-agents.extensions/Runtime/Sensors/PhysicsSensorSettings.cs

+using System;
+using Unity.MLAgents.Sensors;
+
+namespace Unity.MLAgents.Extensions.Sensors
+{
+    /// <summary>
+    /// Settings that define the observations generated for physics-based sensors.
+    /// </summary>
+    [Serializable]
+    public struct PhysicsSensorSettings
+    {
+        /// <summary>
+        /// Whether to use model space (relative to the root body) translations as observations.
+        /// </summary>
+        public bool UseModelSpaceTranslations;
+
+        /// <summary>
+        /// Whether to use model space (relative to the root body) rotations as observations.
+        /// </summary>
+        public bool UseModelSpaceRotations;
+
+        /// <summary>
+        /// Whether to use local space (relative to the parent body) translations as observations.
+        /// </summary>
+        public bool UseLocalSpaceTranslations;
+
+        /// <summary>
+        /// Whether to use local space (relative to the parent body) translations as observations.
+        /// </summary>
+        public bool UseLocalSpaceRotations;
+
+        /// <summary>
+        /// Whether to use model space (relative to the root body) linear velocities as observations.
+        /// </summary>
+        public bool UseModelSpaceLinearVelocity;
+
+        /// <summary>
+        /// Whether to use local space (relative to the parent body) linear velocities as observations.
+        /// </summary>
+        public bool UseLocalSpaceLinearVelocity;
+
+        /// <summary>
+        /// Whether to use joint-specific positions and angles as observations.
+        /// </summary>
+        public bool UseJointPositionsAndAngles;
+
+        /// <summary>
+        /// Whether to use the joint forces and torques that are applied by the solver as observations.
+        /// </summary>
+        public bool UseJointForces;
+
+        /// <summary>
+        /// Creates a PhysicsSensorSettings with reasonable default values.
+        /// </summary>
+        /// <returns></returns>
+        public static PhysicsSensorSettings Default()
+        {
+            return new PhysicsSensorSettings
+            {
+                UseModelSpaceTranslations = true,
+                UseModelSpaceRotations = true,
+            };
+        }
+
+        /// <summary>
+        /// Whether any model space observations are being used.
+        /// </summary>
+        public bool UseModelSpace
+        {
+            get { return UseModelSpaceTranslations || UseModelSpaceRotations || UseModelSpaceLinearVelocity; }
+        }
+
+        /// <summary>
+        /// Whether any local space observations are being used.
+        /// </summary>
+        public bool UseLocalSpace
+        {
+            get { return UseLocalSpaceTranslations || UseLocalSpaceRotations || UseLocalSpaceLinearVelocity; }
+        }
+    }
+
+    internal static class ObservationWriterPhysicsExtensions
+    {
+        /// <summary>
+        /// Utility method for writing a PoseExtractor to an ObservationWriter.
+        /// </summary>
+        /// <param name="writer"></param>
+        /// <param name="settings"></param>
+        /// <param name="poseExtractor"></param>
+        /// <param name="baseOffset">The offset into the ObservationWriter to start writing at.</param>
+        /// <returns>The number of observations written.</returns>
+        public static int WritePoses(this ObservationWriter writer, PhysicsSensorSettings settings, PoseExtractor poseExtractor, int baseOffset = 0)
+        {
+            var offset = baseOffset;
+            if (settings.UseModelSpace)
+            {
+                foreach (var pose in poseExtractor.GetEnabledModelSpacePoses())
+                {
+                    if (settings.UseModelSpaceTranslations)
+                    {
+                        writer.Add(pose.position, offset);
+                        offset += 3;
+                    }
+
+                    if (settings.UseModelSpaceRotations)
+                    {
+                        writer.Add(pose.rotation, offset);
+                        offset += 4;
+                    }
+                }
+
+                foreach(var vel in poseExtractor.GetEnabledModelSpaceVelocities())
+                {
+                    if (settings.UseModelSpaceLinearVelocity)
+                    {
+                        writer.Add(vel, offset);
+                        offset += 3;
+                    }
+                }
+            }
+
+            if (settings.UseLocalSpace)
+            {
+                foreach (var pose in poseExtractor.GetEnabledLocalSpacePoses())
+                {
+                    if (settings.UseLocalSpaceTranslations)
+                    {
+                        writer.Add(pose.position, offset);
+                        offset += 3;
+                    }
+
+                    if (settings.UseLocalSpaceRotations)
+                    {
+                        writer.Add(pose.rotation, offset);
+                        offset += 4;
+                    }
+                }
+
+                foreach(var vel in poseExtractor.GetEnabledLocalSpaceVelocities())
+                {
+                    if (settings.UseLocalSpaceLinearVelocity)
+                    {
+                        writer.Add(vel, offset);
+                        offset += 3;
+                    }
+                }
+            }
+
+            return offset - baseOffset;
+        }
+    }
+}

com.unity.ml-agents.extensions/Runtime/Sensors/ArticulationBodyPoseExtractor.cs

+#if UNITY_2020_1_OR_NEWER
+
+using System.Collections.Generic;
+using UnityEngine;
+
+namespace Unity.MLAgents.Extensions.Sensors
+{
+    /// <summary>
+    /// Utility class to track a hierarchy of ArticulationBodies.
+    /// </summary>
+    public class ArticulationBodyPoseExtractor : PoseExtractor
+    {
+        ArticulationBody[] m_Bodies;
+
+        public ArticulationBodyPoseExtractor(ArticulationBody rootBody)
+        {
+            if (rootBody == null)
+            {
+                return;
+            }
+
+            if (!rootBody.isRoot)
+            {
+                Debug.Log("Must pass ArticulationBody.isRoot");
+                return;
+            }
+
+            var bodies = rootBody.GetComponentsInChildren <ArticulationBody>();
+            if (bodies[0] != rootBody)
+            {
+                Debug.Log("Expected root body at index 0");
+                return;
+            }
+
+            var numBodies = bodies.Length;
+            m_Bodies = bodies;
+            int[] parentIndices = new int[numBodies];
+            parentIndices[0] = -1;
+
+            var bodyToIndex = new Dictionary<ArticulationBody, int>();
+            for (var i = 0; i < numBodies; i++)
+            {
+                bodyToIndex[m_Bodies[i]] = i;
+            }
+
+            for (var i = 1; i < numBodies; i++)
+            {
+                var currentArticBody = m_Bodies[i];
+                // Component.GetComponentInParent will consider the provided object as well.
+                // So start looking from the parent.
+                var currentGameObject = currentArticBody.gameObject;
+                var parentGameObject = currentGameObject.transform.parent;
+                var parentArticBody = parentGameObject.GetComponentInParent<ArticulationBody>();
+                parentIndices[i] = bodyToIndex[parentArticBody];
+            }
+
+            Setup(parentIndices);
+        }
+
+        /// <inheritdoc/>
+        protected internal override Vector3 GetLinearVelocityAt(int index)
+        {
+            return m_Bodies[index].velocity;
+        }
+
+        /// <inheritdoc/>
+        protected internal override Pose GetPoseAt(int index)
+        {
+            var body = m_Bodies[index];
+            var go = body.gameObject;
+            var t = go.transform;
+            return new Pose { rotation = t.rotation, position = t.position };
+        }
+
+        /// <inheritdoc/>
+        protected internal override Object GetObjectAt(int index)
+        {
+            return m_Bodies[index];
+        }
+
+        internal ArticulationBody[] Bodies => m_Bodies;
+
+        internal IEnumerable<ArticulationBody> GetEnabledArticulationBodies()
+        {
+            if (m_Bodies == null)
+            {
+                yield break;
+            }
+
+            for (var i = 0; i < m_Bodies.Length; i++)
+            {
+                var articBody = m_Bodies[i];
+                if (articBody == null)
+                {
+                    // Ignore a virtual root.
+                    continue;
+                }
+
+                if (IsPoseEnabled(i))
+                {
+                    yield return articBody;
+                }
+            }
+        }
+    }
+}
+#endif // UNITY_2020_1_OR_NEWER

com.unity.ml-agents.extensions/Runtime/Sensors/ArticulationBodySensorComponent.cs

+#if UNITY_2020_1_OR_NEWER
+using UnityEngine;
+using Unity.MLAgents.Sensors;
+
+namespace Unity.MLAgents.Extensions.Sensors
+{
+    public class ArticulationBodySensorComponent  : SensorComponent
+    {
+        public ArticulationBody RootBody;
+
+        [SerializeField]
+        public PhysicsSensorSettings Settings = PhysicsSensorSettings.Default();
+        public string sensorName;
+
+        /// <summary>
+        /// Creates a PhysicsBodySensor.
+        /// </summary>
+        /// <returns></returns>
+        public override ISensor CreateSensor()
+        {
+            return new PhysicsBodySensor(RootBody, Settings, sensorName);
+        }
+
+        /// <inheritdoc/>
+        public override int[] GetObservationShape()
+        {
+            if (RootBody == null)
+            {
+                return new[] { 0 };
+            }
+
+            // TODO static method in PhysicsBodySensor?
+            // TODO only update PoseExtractor when body changes?
+            var poseExtractor = new ArticulationBodyPoseExtractor(RootBody);
+            var numPoseObservations = poseExtractor.GetNumPoseObservations(Settings);
+            var numJointObservations = 0;
+
+            foreach(var articBody in poseExtractor.GetEnabledArticulationBodies())
+            {
+                numJointObservations += ArticulationBodyJointExtractor.NumObservations(articBody, Settings);
+            }
+            return new[] { numPoseObservations + numJointObservations };
+        }
+    }
+
+}
+#endif // UNITY_2020_1_OR_NEWER

com.unity.ml-agents.extensions/Runtime/Sensors/PhysicsBodySensor.cs

+using System.Collections.Generic;
+using UnityEngine;
+using Unity.MLAgents.Sensors;
+
+namespace Unity.MLAgents.Extensions.Sensors
+{
+    /// <summary>
+    /// ISensor implementation that generates observations for a group of Rigidbodies or ArticulationBodies.
+    /// </summary>
+    public class PhysicsBodySensor : ISensor
+    {
+        int[] m_Shape;
+        string m_SensorName;
+
+        PoseExtractor m_PoseExtractor;
+        List<IJointExtractor> m_JointExtractors;
+        PhysicsSensorSettings m_Settings;
+
+        /// <summary>
+        /// Construct a new PhysicsBodySensor
+        /// </summary>
+        /// <param name="poseExtractor"></param>
+        /// <param name="settings"></param>
+        /// <param name="sensorName"></param>
+        public PhysicsBodySensor(
+            RigidBodyPoseExtractor poseExtractor,
+            PhysicsSensorSettings settings,
+            string sensorName
+        )
+        {
+            m_PoseExtractor = poseExtractor;
+            m_SensorName = sensorName;
+            m_Settings = settings;
+
+            var numJointExtractorObservations = 0;
+            m_JointExtractors = new List<IJointExtractor>(poseExtractor.NumEnabledPoses);
+            foreach(var rb in poseExtractor.GetEnabledRigidbodies())
+            {
+                var jointExtractor = new RigidBodyJointExtractor(rb);
+                numJointExtractorObservations += jointExtractor.NumObservations(settings);
+                m_JointExtractors.Add(jointExtractor);
+            }
+
+            var numTransformObservations = m_PoseExtractor.GetNumPoseObservations(settings);
+            m_Shape = new[] { numTransformObservations + numJointExtractorObservations };
+        }
+
+#if UNITY_2020_1_OR_NEWER
+        public PhysicsBodySensor(ArticulationBody rootBody, PhysicsSensorSettings settings, string sensorName=null)
+        {
+            var poseExtractor = new ArticulationBodyPoseExtractor(rootBody);
+            m_PoseExtractor = poseExtractor;
+            m_SensorName = string.IsNullOrEmpty(sensorName) ? $"ArticulationBodySensor:{rootBody?.name}" : sensorName;
+            m_Settings = settings;
+
+            var numJointExtractorObservations = 0;
+            m_JointExtractors = new List<IJointExtractor>(poseExtractor.NumEnabledPoses);
+            foreach(var articBody in poseExtractor.GetEnabledArticulationBodies())
+            {
+                var jointExtractor = new ArticulationBodyJointExtractor(articBody);
+                numJointExtractorObservations += jointExtractor.NumObservations(settings);
+                m_JointExtractors.Add(jointExtractor);
+            }
+
+            var numTransformObservations = m_PoseExtractor.GetNumPoseObservations(settings);
+            m_Shape = new[] { numTransformObservations + numJointExtractorObservations };
+        }
+#endif
+
+        /// <inheritdoc/>
+        public int[] GetObservationShape()
+        {
+            return m_Shape;
+        }
+
+        /// <inheritdoc/>
+        public int Write(ObservationWriter writer)
+        {
+            var numWritten = writer.WritePoses(m_Settings, m_PoseExtractor);
+            foreach (var jointExtractor in m_JointExtractors)
+            {
+                numWritten += jointExtractor.Write(m_Settings, writer, numWritten);
+            }
+            return numWritten;
+        }
+
+        /// <inheritdoc/>
+        public byte[] GetCompressedObservation()
+        {
+            return null;
+        }
+
+        /// <inheritdoc/>
+        public void Update()
+        {
+            if (m_Settings.UseModelSpace)
+            {
+                m_PoseExtractor.UpdateModelSpacePoses();
+            }
+
+            if (m_Settings.UseLocalSpace)
+            {
+                m_PoseExtractor.UpdateLocalSpacePoses();
+            }
+        }
+
+        /// <inheritdoc/>
+        public void Reset() {}
+
+        /// <inheritdoc/>
+        public SensorCompressionType GetCompressionType()
+        {
+            return SensorCompressionType.None;
+        }
+
+        /// <inheritdoc/>
+        public string GetName()
+        {
+            return m_SensorName;
+        }
+    }
+}

com.unity.ml-agents.extensions/Runtime/Sensors/PoseExtractor.cs

+using System;
+using System.Collections.Generic;
+using UnityEngine;
+using Object = UnityEngine.Object;
+
+namespace Unity.MLAgents.Extensions.Sensors
+{
+    /// <summary>
+    /// Abstract class for managing the transforms of a hierarchy of objects.
+    /// This could be GameObjects or Monobehaviours in the scene graph, but this is
+    /// not a requirement; for example, the objects could be rigid bodies whose hierarchy
+    /// is defined by Joint configurations.
+    ///
+    /// Poses are either considered in model space, which is relative to a root body,
+    /// or in local space, which is relative to their parent.
+    /// </summary>
+    public abstract class PoseExtractor
+    {
+        int[] m_ParentIndices;
+        Pose[] m_ModelSpacePoses;
+        Pose[] m_LocalSpacePoses;
+
+        Vector3[] m_ModelSpaceLinearVelocities;
+        Vector3[] m_LocalSpaceLinearVelocities;
+
+        bool[] m_PoseEnabled;
+
+
+        /// <summary>
+        /// Read iterator for the enabled model space transforms.
+        /// </summary>
+        public IEnumerable<Pose> GetEnabledModelSpacePoses()
+        {
+            if (m_ModelSpacePoses == null)
+            {
+                yield break;
+            }
+
+            for (var i = 0; i < m_ModelSpacePoses.Length; i++)
+            {
+                if (m_PoseEnabled[i])
+                {
+                    yield return m_ModelSpacePoses[i];
+                }
+            }
+        }
+
+        /// <summary>
+        /// Read iterator for the enabled local space transforms.
+        /// </summary>
+        public IEnumerable<Pose> GetEnabledLocalSpacePoses()
+        {
+            if (m_LocalSpacePoses == null)
+            {
+                yield break;
+            }
+
+            for (var i = 0; i < m_LocalSpacePoses.Length; i++)
+            {
+                if (m_PoseEnabled[i])
+                {
+                    yield return m_LocalSpacePoses[i];
+                }
+            }
+        }
+
+        /// <summary>
+        /// Read iterator for the enabled model space linear velocities.
+        /// </summary>
+        public IEnumerable<Vector3> GetEnabledModelSpaceVelocities()
+        {
+            if (m_ModelSpaceLinearVelocities == null)
+            {
+                yield break;
+            }
+
+            for (var i = 0; i < m_ModelSpaceLinearVelocities.Length; i++)
+            {
+                if (m_PoseEnabled[i])
+                {
+                    yield return m_ModelSpaceLinearVelocities[i];
+                }
+            }
+        }
+
+        /// <summary>
+        /// Read iterator for the enabled local space linear velocities.
+        /// </summary>
+        public IEnumerable<Vector3> GetEnabledLocalSpaceVelocities()
+        {
+            if (m_LocalSpaceLinearVelocities == null)
+            {
+                yield break;
+            }
+
+            for (var i = 0; i < m_LocalSpaceLinearVelocities.Length; i++)
+            {
+                if (m_PoseEnabled[i])
+                {
+                    yield return m_LocalSpaceLinearVelocities[i];
+                }
+            }
+        }
+
+        /// <summary>
+        /// Number of enabled poses in the hierarchy (read-only).
+        /// </summary>
+        public int NumEnabledPoses
+        {
+            get
+            {
+                if (m_PoseEnabled == null)
+                {
+                    return 0;
+                }
+
+                var numEnabled = 0;
+                for (var i = 0; i < m_PoseEnabled.Length; i++)
+                {
+                    numEnabled += m_PoseEnabled[i] ? 1 : 0;
+                }
+
+                return numEnabled;
+            }
+        }
+
+        /// <summary>
+        /// Number of total poses in the hierarchy (read-only).
+        /// </summary>
+        public int NumPoses
+        {
+            get { return m_ModelSpacePoses?.Length ?? 0; }
+        }
+
+        /// <summary>
+        /// Get the parent index of the body at the specified index.
+        /// </summary>
+        /// <param name="index"></param>
+        /// <returns></returns>
+        public int GetParentIndex(int index)
+        {
+            if (m_ParentIndices == null)
+            {
+                throw new NullReferenceException("No parent indices set");
+            }
+
+            return m_ParentIndices[index];
+        }
+
+        /// <summary>
+        /// Set whether the pose at the given index is enabled or disabled for observations.
+        /// </summary>
+        /// <param name="index"></param>
+        /// <param name="val"></param>
+        public void SetPoseEnabled(int index, bool val)
+        {
+            m_PoseEnabled[index] = val;
+        }
+
+        public bool IsPoseEnabled(int index)
+        {
+            return m_PoseEnabled[index];
+        }
+
+        /// <summary>
+        /// Initialize with the mapping of parent indices.
+        /// The 0th element is assumed to be -1, indicating that it's the root.
+        /// </summary>
+        /// <param name="parentIndices"></param>
+        protected void Setup(int[] parentIndices)
+        {
+#if DEBUG
+            if (parentIndices[0] != -1)
+            {
+                throw new UnityAgentsException($"Expected parentIndices[0] to be -1, got {parentIndices[0]}");
+            }
+#endif
+            m_ParentIndices = parentIndices;
+            var numPoses = parentIndices.Length;
+            m_ModelSpacePoses = new Pose[numPoses];
+            m_LocalSpacePoses = new Pose[numPoses];
+
+            m_ModelSpaceLinearVelocities = new Vector3[numPoses];
+            m_LocalSpaceLinearVelocities = new Vector3[numPoses];
+
+            m_PoseEnabled = new bool[numPoses];
+            // All poses are enabled by default. Generally we'll want to disable the root though.
+            for (var i = 0; i < numPoses; i++)
+            {
+                m_PoseEnabled[i] = true;
+            }
+        }
+
+        /// <summary>
+        /// Return the world space Pose of the i'th object.
+        /// </summary>
+        /// <param name="index"></param>
+        /// <returns></returns>
+        protected internal abstract Pose GetPoseAt(int index);
+
+        /// <summary>
+        /// Return the world space linear velocity of the i'th object.
+        /// </summary>
+        /// <param name="index"></param>
+        /// <returns></returns>
+        protected internal abstract Vector3 GetLinearVelocityAt(int index);
+
+        /// <summary>
+        /// Return the underlying object at the given index. This is only
+        /// used for display in the inspector.
+        /// </summary>
+        /// <param name="index"></param>
+        /// <returns></returns>
+        protected internal virtual Object GetObjectAt(int index)
+        {
+            return null;
+        }
+
+
+        /// <summary>
+        /// Update the internal model space transform storage based on the underlying system.
+        /// </summary>
+        public void UpdateModelSpacePoses()
+        {
+            using (TimerStack.Instance.Scoped("UpdateModelSpacePoses"))
+            {
+                if (m_ModelSpacePoses == null)
+                {
+                    return;
+                }
+
+                var rootWorldTransform = GetPoseAt(0);
+                var worldToModel = rootWorldTransform.Inverse();
+                var rootLinearVel = GetLinearVelocityAt(0);
+
+                for (var i = 0; i < m_ModelSpacePoses.Length; i++)
+                {
+                    var currentWorldSpacePose = GetPoseAt(i);
+                    var currentModelSpacePose = worldToModel.Multiply(currentWorldSpacePose);
+                    m_ModelSpacePoses[i] = currentModelSpacePose;
+
+                    var currentBodyLinearVel = GetLinearVelocityAt(i);
+                    var relativeVelocity = currentBodyLinearVel - rootLinearVel;
+                    m_ModelSpaceLinearVelocities[i] = worldToModel.rotation * relativeVelocity;
+                }
+            }
+        }
+
+        /// <summary>
+        /// Update the internal model space transform storage based on the underlying system.
+        /// </summary>
+        public void UpdateLocalSpacePoses()
+        {
+            using (TimerStack.Instance.Scoped("UpdateLocalSpacePoses"))
+            {
+                if (m_LocalSpacePoses == null)
+                {
+                    return;
+                }
+
+                for (var i = 0; i < m_LocalSpacePoses.Length; i++)
+                {
+                    if (m_ParentIndices[i] != -1)
+                    {
+                        var parentTransform = GetPoseAt(m_ParentIndices[i]);
+                        // This is slightly inefficient, since for a body with multiple children, we'll end up inverting
+                        // the transform multiple times. Might be able to trade space for perf here.
+                        var invParent = parentTransform.Inverse();
+                        var currentTransform = GetPoseAt(i);
+                        m_LocalSpacePoses[i] = invParent.Multiply(currentTransform);
+
+                        var parentLinearVel = GetLinearVelocityAt(m_ParentIndices[i]);
+                        var currentLinearVel = GetLinearVelocityAt(i);
+                        m_LocalSpaceLinearVelocities[i] = invParent.rotation * (currentLinearVel - parentLinearVel);
+                    }
+                    else
+                    {
+                        m_LocalSpacePoses[i] = Pose.identity;
+                        m_LocalSpaceLinearVelocities[i] = Vector3.zero;
+                    }
+                }
+            }
+        }
+
+        /// <summary>
+        /// Compute the number of floats needed to represent the poses for the given PhysicsSensorSettings.
+        /// </summary>
+        /// <param name="settings"></param>
+        /// <returns></returns>
+        public int GetNumPoseObservations(PhysicsSensorSettings settings)
+        {
+            int obsPerPose = 0;
+            obsPerPose += settings.UseModelSpaceTranslations ? 3 : 0;
+            obsPerPose += settings.UseModelSpaceRotations ? 4 : 0;
+            obsPerPose += settings.UseLocalSpaceTranslations ? 3 : 0;
+            obsPerPose += settings.UseLocalSpaceRotations ? 4 : 0;
+
+            obsPerPose += settings.UseModelSpaceLinearVelocity ? 3 : 0;
+            obsPerPose += settings.UseLocalSpaceLinearVelocity ? 3 : 0;
+
+            return NumEnabledPoses * obsPerPose;
+        }
+
+        internal void DrawModelSpace(Vector3 offset)
+        {
+            UpdateLocalSpacePoses();
+            UpdateModelSpacePoses();
+
+            var pose = m_ModelSpacePoses;
+            var localPose = m_LocalSpacePoses;
+            for (var i = 0; i < pose.Length; i++)
+            {
+                var current = pose[i];
+                if (m_ParentIndices[i] == -1)
+                {
+                    continue;
+                }
+
+                var parent = pose[m_ParentIndices[i]];
+                Debug.DrawLine(current.position + offset, parent.position + offset, Color.cyan);
+                var localUp = localPose[i].rotation * Vector3.up;
+                var localFwd = localPose[i].rotation * Vector3.forward;
+                var localRight = localPose[i].rotation * Vector3.right;
+                Debug.DrawLine(current.position+offset, current.position+offset+.1f*localUp, Color.red);
+                Debug.DrawLine(current.position+offset, current.position+offset+.1f*localFwd, Color.green);
+                Debug.DrawLine(current.position+offset, current.position+offset+.1f*localRight, Color.blue);
+            }
+        }
+
+        /// <summary>
+        /// Simplified representation of the a node in the hierarchy for display.
+        /// </summary>
+        internal struct DisplayNode
+        {
+            /// <summary>
+            /// Underlying object in the hierarchy. Pass to EditorGUIUtility.ObjectContent() for display.
+            /// </summary>
+            public Object NodeObject;
+
+            /// <summary>
+            /// Whether the poses for the object are enabled.
+            /// </summary>
+            public bool Enabled;
+
+            /// <summary>
+            /// Depth in the hierarchy, used for adjusting the indent level.
+            /// </summary>
+            public int Depth;
+
+            /// <summary>
+            /// The index of the corresponding object in the PoseExtractor.
+            /// </summary>
+            public int OriginalIndex;
+        }
+
+        /// <summary>
+        /// Get a list of display nodes in depth-first order.
+        /// </summary>
+        /// <returns></returns>
+        internal IList<DisplayNode> GetDisplayNodes()
+        {
+            if (NumPoses == 0)
+            {
+                return Array.Empty<DisplayNode>();
+            }
+            var nodesOut = new List<DisplayNode>(NumPoses);
+
+            // List of children for each node
+            var tree = new Dictionary<int, List<int>>();
+            for (var i = 0; i < NumPoses; i++)
+            {
+                var parent = GetParentIndex(i);
+                if (i == -1)
+                {
+                    continue;
+                }
+
+                if (!tree.ContainsKey(parent))
+                {
+                    tree[parent] = new List<int>();
+                }
+                tree[parent].Add(i);
+            }
+
+            // Store (index, depth) in the stack
+            var stack = new Stack<(int, int)>();
+            stack.Push((0, 0));
+
+            while (stack.Count != 0)
+            {
+                var (current, depth) = stack.Pop();
+                var obj = GetObjectAt(current);
+
+                var node = new DisplayNode
+                {
+                    NodeObject = obj,
+                    Enabled = IsPoseEnabled(current),
+                    OriginalIndex = current,
+                    Depth = depth
+                };
+                nodesOut.Add(node);
+
+                // Add children
+                if (tree.ContainsKey(current))
+                {
+                    // Push to the stack in reverse order
+                    var children = tree[current];
+                    for (var childIdx = children.Count-1; childIdx >= 0; childIdx--)
+                    {
+                        stack.Push((children[childIdx], depth+1));
+                    }
+                }
+
+                // Safety check
+                // This shouldn't even happen, but in case we have a cycle in the graph
+                // exit instead of looping forever and eating up all the memory.
+                if (nodesOut.Count > NumPoses)
+                {
+                    return nodesOut;
+                }
+            }
+
+            return nodesOut;
+        }
+
+    }
+
+    /// <summary>
+    /// Extension methods for the Pose struct, in order to improve the readability of some math.
+    /// </summary>
+    public static class PoseExtensions
+    {
+        /// <summary>
+        /// Compute the inverse of a Pose. For any Pose P,
+        ///   P.Inverse() * P
+        /// will equal the identity pose (within tolerance).
+        /// </summary>
+        /// <param name="pose"></param>
+        /// <returns></returns>
+        public static Pose Inverse(this Pose pose)
+        {
+            var rotationInverse = Quaternion.Inverse(pose.rotation);
+            var translationInverse = -(rotationInverse * pose.position);
+            return new Pose { rotation = rotationInverse, position = translationInverse };
+        }
+
+        /// <summary>
+        /// This is equivalent to Pose.GetTransformedBy(), but keeps the order more intuitive.
+        /// </summary>
+        /// <param name="pose"></param>
+        /// <param name="rhs"></param>
+        /// <returns></returns>
+        public static Pose Multiply(this Pose pose, Pose rhs)
+        {
+            return rhs.GetTransformedBy(pose);
+        }
+
+        /// <summary>
+        /// Transform the vector by the pose. Conceptually this is equivalent to treating the Pose
+        /// as a 4x4 matrix and multiplying the augmented vector.
+        /// See https://en.wikipedia.org/wiki/Affine_transformation#Augmented_matrix for more details.
+        /// </summary>
+        /// <param name="pose"></param>
+        /// <param name="rhs"></param>
+        /// <returns></returns>
+        public static Vector3 Multiply(this Pose pose, Vector3 rhs)
+        {
+            return pose.rotation * rhs + pose.position;
+        }
+
+        // TODO optimize inv(A)*B?
+    }
+}

com.unity.ml-agents.extensions/Runtime/Sensors/RigidBodyPoseExtractor.cs

+using System.Collections.Generic;
+using UnityEngine;
+
+namespace Unity.MLAgents.Extensions.Sensors
+{
+
+    /// <summary>
+    /// Utility class to track a hierarchy of RigidBodies. These are assumed to have a root node,
+    /// and child nodes are connect to their parents via Joints.
+    /// </summary>
+    public class RigidBodyPoseExtractor : PoseExtractor
+    {
+        Rigidbody[] m_Bodies;
+
+        /// <summary>
+        /// Optional game object used to determine the root of the poses, separate from the actual Rigidbodies
+        /// in the hierarchy. For locomotion
+        /// </summary>
+        GameObject m_VirtualRoot;
+
+        /// <summary>
+        /// Initialize given a root RigidBody.
+        /// </summary>
+        /// <param name="rootBody">The root Rigidbody. This has no Joints on it (but other Joints may connect to it).</param>
+        /// <param name="rootGameObject">Optional GameObject used to find Rigidbodies in the hierarchy.</param>
+        /// <param name="virtualRoot">Optional GameObject used to determine the root of the poses,
+        /// separate from the actual Rigidbodies in the hierarchy. For locomotion tasks, with ragdolls, this provides
+        /// a stabilized reference frame, which can improve learning.</param>
+        /// <param name="enableBodyPoses">Optional mapping of whether a body's psoe should be enabled or not.</param>
+        public RigidBodyPoseExtractor(Rigidbody rootBody, GameObject rootGameObject = null,
+            GameObject virtualRoot = null, Dictionary<Rigidbody, bool> enableBodyPoses = null)
+        {
+            if (rootBody == null)
+            {
+                return;
+            }
+
+            Rigidbody[] rbs;
+            Joint[] joints;
+            if (rootGameObject == null)
+            {
+                rbs = rootBody.GetComponentsInChildren<Rigidbody>();
+                joints = rootBody.GetComponentsInChildren <Joint>();
+            }
+            else
+            {
+                rbs = rootGameObject.GetComponentsInChildren<Rigidbody>();
+                joints = rootGameObject.GetComponentsInChildren<Joint>();
+            }
+
+            if (rbs == null || rbs.Length == 0)
+            {
+                Debug.Log("No rigid bodies found!");
+                return;
+            }
+
+            if (rbs[0] != rootBody)
+            {
+                Debug.Log("Expected root body at index 0");
+                return;
+            }
+
+            // Adjust the array if we have a virtual root.
+            // This will be at index 0, and the "real" root will be parented to it.
+            if (virtualRoot != null)
+            {
+                var extendedRbs = new Rigidbody[rbs.Length + 1];
+                for (var i = 0; i < rbs.Length; i++)
+                {
+                    extendedRbs[i + 1] = rbs[i];
+                }
+
+                rbs = extendedRbs;
+            }
+
+            var bodyToIndex = new Dictionary<Rigidbody, int>(rbs.Length);
+            var parentIndices = new int[rbs.Length];
+            parentIndices[0] = -1;
+
+            for (var i = 0; i < rbs.Length; i++)
+            {
+                if(rbs[i] != null)
+                {
+                    bodyToIndex[rbs[i]] = i;
+                }
+            }
+
+            foreach (var j in joints)
+            {
+                var parent = j.connectedBody;
+                var child = j.GetComponent<Rigidbody>();
+
+                var parentIndex = bodyToIndex[parent];
+                var childIndex = bodyToIndex[child];
+                parentIndices[childIndex] = parentIndex;
+            }
+
+            if (virtualRoot != null)
+            {
+                // Make sure the original root treats the virtual root as its parent.
+                parentIndices[1] = 0;
+                m_VirtualRoot = virtualRoot;
+            }
+
+            m_Bodies = rbs;
+            Setup(parentIndices);
+
+            // By default, ignore the root
+            SetPoseEnabled(0, false);
+
+            if (enableBodyPoses != null)
+            {
+                foreach (var pair in enableBodyPoses)
+                {
+                    var rb = pair.Key;
+                    if (bodyToIndex.TryGetValue(rb, out var index))
+                    {
+                        SetPoseEnabled(index, pair.Value);
+                    }
+                }
+            }
+        }
+
+        /// <inheritdoc/>
+        protected internal override Vector3 GetLinearVelocityAt(int index)
+        {
+            if (index == 0 && m_VirtualRoot != null)
+            {
+                // No velocity on the virtual root
+                return Vector3.zero;
+            }
+            return m_Bodies[index].velocity;
+        }
+
+        /// <inheritdoc/>
+        protected internal override Pose GetPoseAt(int index)
+        {
+            if (index == 0 && m_VirtualRoot != null)
+            {
+                // Use the GameObject's world transform
+                return new Pose
+                {
+                    rotation = m_VirtualRoot.transform.rotation,
+                    position = m_VirtualRoot.transform.position
+                };
+            }
+
+            var body = m_Bodies[index];
+            return new Pose { rotation = body.rotation, position = body.position };
+        }
+
+        /// <inheritdoc/>
+        protected internal override Object GetObjectAt(int index)
+        {
+            if (index == 0 && m_VirtualRoot != null)
+            {
+                return m_VirtualRoot;
+            }
+            return m_Bodies[index];
+        }
+
+        internal Rigidbody[] Bodies => m_Bodies;
+
+        /// <summary>
+        /// Get a dictionary indicating which Rigidbodies' poses are enabled or disabled.
+        /// </summary>
+        /// <returns></returns>
+        internal Dictionary<Rigidbody, bool> GetBodyPosesEnabled()
+        {
+            var bodyPosesEnabled = new Dictionary<Rigidbody, bool>(m_Bodies.Length);
+            for (var i = 0; i < m_Bodies.Length; i++)
+            {
+                var rb = m_Bodies[i];
+                if (rb == null)
+                {
+                    continue; // skip virtual root
+                }
+
+                bodyPosesEnabled[rb] = IsPoseEnabled(i);
+            }
+
+            return bodyPosesEnabled;
+        }
+
+        internal IEnumerable<Rigidbody> GetEnabledRigidbodies()
+        {
+            if (m_Bodies == null)
+            {
+                yield break;
+            }
+
+            for (var i = 0; i < m_Bodies.Length; i++)
+            {
+                var rb = m_Bodies[i];
+                if (rb == null)
+                {
+                    // Ignore a virtual root.
+                    continue;
+                }
+
+                if (IsPoseEnabled(i))
+                {
+                    yield return rb;
+                }
+            }
+        }
+    }
+
+}

com.unity.ml-agents.extensions/Runtime/Sensors/RigidBodySensorComponent.cs

+using System.Collections.Generic;
+using UnityEngine;
+using Unity.MLAgents.Sensors;
+
+namespace Unity.MLAgents.Extensions.Sensors
+{
+    /// <summary>
+    /// Editor component that creates a PhysicsBodySensor for the Agent.
+    /// </summary>
+    public class RigidBodySensorComponent  : SensorComponent
+    {
+        /// <summary>
+        /// The root Rigidbody of the system.
+        /// </summary>
+        public Rigidbody RootBody;
+
+        /// <summary>
+        /// Optional GameObject used to determine the root of the poses.
+        /// </summary>
+        public GameObject VirtualRoot;
+
+        /// <summary>
+        /// Settings defining what types of observations will be generated.
+        /// </summary>
+        [SerializeField]
+        public PhysicsSensorSettings Settings = PhysicsSensorSettings.Default();
+
+        /// <summary>
+        /// Optional sensor name. This must be unique for each Agent.
+        /// </summary>
+        [SerializeField]
+        public string sensorName;
+
+        [SerializeField]
+        [HideInInspector]
+        RigidBodyPoseExtractor m_PoseExtractor;
+
+        /// <summary>
+        /// Creates a PhysicsBodySensor.
+        /// </summary>
+        /// <returns></returns>
+        public override ISensor CreateSensor()
+        {
+            var _sensorName = string.IsNullOrEmpty(sensorName) ? $"PhysicsBodySensor:{RootBody?.name}" : sensorName;
+            return new PhysicsBodySensor(GetPoseExtractor(), Settings, _sensorName);
+        }
+
+        /// <inheritdoc/>
+        public override int[] GetObservationShape()
+        {
+            if (RootBody == null)
+            {
+                return new[] { 0 };
+            }
+
+            var poseExtractor = GetPoseExtractor();
+            var numPoseObservations = poseExtractor.GetNumPoseObservations(Settings);
+
+            var numJointObservations = 0;
+            foreach(var rb in poseExtractor.GetEnabledRigidbodies())
+            {
+                var joint = rb.GetComponent<Joint>();
+                numJointObservations += RigidBodyJointExtractor.NumObservations(rb, joint, Settings);
+            }
+            return new[] { numPoseObservations + numJointObservations };
+        }
+
+        /// <summary>
+        /// Get the DisplayNodes of the hierarchy.
+        /// </summary>
+        /// <returns></returns>
+        internal IList<PoseExtractor.DisplayNode> GetDisplayNodes()
+        {
+            return GetPoseExtractor().GetDisplayNodes();
+        }
+
+        /// <summary>
+        /// Lazy construction of the PoseExtractor.
+        /// </summary>
+        /// <returns></returns>
+        RigidBodyPoseExtractor GetPoseExtractor()
+        {
+            if (m_PoseExtractor == null)
+            {
+                ResetPoseExtractor();
+            }
+
+            return m_PoseExtractor;
+        }
+
+        /// <summary>
+        /// Reset the pose extractor, trying to keep the enabled state of the corresponding poses the same.
+        /// </summary>
+        internal void ResetPoseExtractor()
+        {
+            // Get the current enabled state of each body, so that we can reinitialize with them.
+            Dictionary<Rigidbody, bool> bodyPosesEnabled = null;
+            if (m_PoseExtractor != null)
+            {
+                bodyPosesEnabled = m_PoseExtractor.GetBodyPosesEnabled();
+            }
+            m_PoseExtractor = new RigidBodyPoseExtractor(RootBody, gameObject, VirtualRoot, bodyPosesEnabled);
+        }
+
+        /// <summary>
+        /// Toggle the pose at the given index.
+        /// </summary>
+        /// <param name="index"></param>
+        /// <param name="enabled"></param>
+        internal void SetPoseEnabled(int index, bool enabled)
+        {
+            GetPoseExtractor().SetPoseEnabled(index, enabled);
+        }
+    }
+
+}

com.unity.ml-agents.extensions/Tests/Editor/Sensors/ArticulationBodySensorTests.cs

+#if UNITY_2020_1_OR_NEWER
+using UnityEngine;
+using NUnit.Framework;
+using Unity.MLAgents.Extensions.Sensors;
+
+
+namespace Unity.MLAgents.Extensions.Tests.Sensors
+{
+
+    public class ArticulationBodySensorTests
+    {
+        [Test]
+        public void TestNullRootBody()
+        {
+            var gameObj = new GameObject();
+
+            var sensorComponent = gameObj.AddComponent<ArticulationBodySensorComponent>();
+            var sensor = sensorComponent.CreateSensor();
+            SensorTestHelper.CompareObservation(sensor, new float[0]);
+        }
+
+        [Test]
+        public void TestSingleBody()
+        {
+            var gameObj = new GameObject();
+            var articulationBody = gameObj.AddComponent<ArticulationBody>();
+            var sensorComponent = gameObj.AddComponent<ArticulationBodySensorComponent>();
+            sensorComponent.RootBody = articulationBody;
+            sensorComponent.Settings = new PhysicsSensorSettings
+            {
+                UseModelSpaceLinearVelocity = true,
+                UseLocalSpaceTranslations = true,
+                UseLocalSpaceRotations = true
+            };
+
+            var sensor = sensorComponent.CreateSensor();
+            sensor.Update();
+            var expected = new[]
+            {
+                0f, 0f, 0f, // ModelSpaceLinearVelocity
+                0f, 0f, 0f, // LocalSpaceTranslations
+                0f, 0f, 0f, 1f // LocalSpaceRotations
+            };
+            SensorTestHelper.CompareObservation(sensor, expected);
+            Assert.AreEqual(expected.Length, sensorComponent.GetObservationShape()[0]);
+        }
+
+        [Test]
+        public void TestBodiesWithJoint()
+        {
+            var rootObj = new GameObject();
+            var rootArticBody = rootObj.AddComponent<ArticulationBody>();
+
+            var middleGamObj = new GameObject();
+            var middleArticBody = middleGamObj.AddComponent<ArticulationBody>();
+            middleArticBody.AddForce(new Vector3(0f, 1f, 0f));
+            middleGamObj.transform.SetParent(rootObj.transform);
+            middleGamObj.transform.localPosition = new Vector3(13.37f, 0f, 0f);
+            middleArticBody.jointType = ArticulationJointType.RevoluteJoint;
+
+            var leafGameObj = new GameObject();
+            var leafArticBody = leafGameObj.AddComponent<ArticulationBody>();
+            leafGameObj.transform.SetParent(middleGamObj.transform);
+            leafGameObj.transform.localPosition = new Vector3(4.2f, 0f, 0f);
+            leafArticBody.jointType = ArticulationJointType.PrismaticJoint;
+            leafArticBody.linearLockZ = ArticulationDofLock.LimitedMotion;
+            leafArticBody.zDrive = new ArticulationDrive
+            {
+                lowerLimit = -3,
+                upperLimit = 1
+            };
+
+
+#if UNITY_2020_2_OR_NEWER
+            // ArticulationBody.velocity is read-only in 2020.1
+            rootArticBody.velocity = new Vector3(1f, 0f, 0f);
+            middleArticBody.velocity = new Vector3(0f, 1f, 0f);
+            leafArticBody.velocity = new Vector3(0f, 0f, 1f);
+#endif
+
+            var sensorComponent = rootObj.AddComponent<ArticulationBodySensorComponent>();
+            sensorComponent.RootBody = rootArticBody;
+            sensorComponent.Settings = new PhysicsSensorSettings
+            {
+                UseModelSpaceTranslations = true,
+                UseLocalSpaceTranslations = true,
+#if UNITY_2020_2_OR_NEWER
+                UseLocalSpaceLinearVelocity = true
+#endif
+            };
+
+            var sensor = sensorComponent.CreateSensor();
+            sensor.Update();
+            var expected = new[]
+            {
+                // Model space
+                0f, 0f, 0f, // Root pos
+                13.37f, 0f, 0f, // Middle pos
+                leafGameObj.transform.position.x, 0f, 0f, // Leaf pos
+
+                // Local space
+                0f, 0f, 0f, // Root pos
+                13.37f, 0f, 0f, // Attached pos
+                4.2f, 0f, 0f, // Leaf pos
+
+#if UNITY_2020_2_OR_NEWER
+                0f, 0f, 0f, // Root vel
+                -1f, 1f, 0f, // Attached vel
+                0f, -1f, 1f // Leaf vel
+#endif
+            };
+            SensorTestHelper.CompareObservation(sensor, expected);
+            Assert.AreEqual(expected.Length, sensorComponent.GetObservationShape()[0]);
+
+            // Update the settings to only process joint observations
+            sensorComponent.Settings = new PhysicsSensorSettings
+            {
+                UseJointForces = true,
+                UseJointPositionsAndAngles = true,
+            };
+
+            sensor = sensorComponent.CreateSensor();
+            sensor.Update();
+
+            expected = new[]
+            {
+                // revolute
+                0f, 1f, // joint1.position (sin and cos)
+                0f, // joint1.force
+
+                // prismatic
+                0.5f, // joint2.position (interpolate between limits)
+                0f, // joint2.force
+            };
+            SensorTestHelper.CompareObservation(sensor, expected);
+            Assert.AreEqual(expected.Length, sensorComponent.GetObservationShape()[0]);
+        }
+    }
+}
+#endif // #if UNITY_2020_1_OR_NEWER

com.unity.ml-agents.extensions/Tests/Editor/Sensors/RigidBodySensorTests.cs

+using UnityEngine;
+using NUnit.Framework;
+using Unity.MLAgents.Sensors;
+using Unity.MLAgents.Extensions.Sensors;
+
+
+namespace Unity.MLAgents.Extensions.Tests.Sensors
+{
+
+    public static class SensorTestHelper
+    {
+        public static void CompareObservation(ISensor sensor, float[] expected)
+        {
+            string errorMessage;
+            bool isOK = SensorHelper.CompareObservation(sensor, expected, out errorMessage);
+            Assert.IsTrue(isOK, errorMessage);
+        }
+    }
+
+    public class RigidBodySensorTests
+    {
+        [Test]
+        public void TestNullRootBody()
+        {
+            var gameObj = new GameObject();
+
+            var sensorComponent = gameObj.AddComponent<RigidBodySensorComponent>();
+            var sensor = sensorComponent.CreateSensor();
+            SensorTestHelper.CompareObservation(sensor, new float[0]);
+        }
+
+        [Test]
+        public void TestSingleRigidbody()
+        {
+            var gameObj = new GameObject();
+            var rootRb = gameObj.AddComponent<Rigidbody>();
+            var sensorComponent = gameObj.AddComponent<RigidBodySensorComponent>();
+            sensorComponent.RootBody = rootRb;
+            sensorComponent.Settings = new PhysicsSensorSettings
+            {
+                UseModelSpaceLinearVelocity = true,
+                UseLocalSpaceTranslations = true,
+                UseLocalSpaceRotations = true
+            };
+
+            var sensor = sensorComponent.CreateSensor();
+            sensor.Update();
+
+            // The root body is ignored since it always generates identity values
+            // and there are no other bodies to generate observations.
+            var expected = new float[0];
+            Assert.AreEqual(expected.Length, sensorComponent.GetObservationShape()[0]);
+            SensorTestHelper.CompareObservation(sensor, expected);
+        }
+
+        [Test]
+        public void TestBodiesWithJoint()
+        {
+            var rootObj = new GameObject();
+            var rootRb = rootObj.AddComponent<Rigidbody>();
+            rootRb.velocity = new Vector3(1f, 0f, 0f);
+
+            var middleGamObj = new GameObject();
+            var middleRb = middleGamObj.AddComponent<Rigidbody>();
+            middleRb.velocity = new Vector3(0f, 1f, 0f);
+            middleGamObj.transform.SetParent(rootObj.transform);
+            middleGamObj.transform.localPosition = new Vector3(13.37f, 0f, 0f);
+            var joint = middleGamObj.AddComponent<ConfigurableJoint>();
+            joint.connectedBody = rootRb;
+
+            var leafGameObj = new GameObject();
+            var leafRb = leafGameObj.AddComponent<Rigidbody>();
+            leafRb.velocity = new Vector3(0f, 0f, 1f);
+            leafGameObj.transform.SetParent(middleGamObj.transform);
+            leafGameObj.transform.localPosition = new Vector3(4.2f, 0f, 0f);
+            var joint2 = leafGameObj.AddComponent<ConfigurableJoint>();
+            joint2.connectedBody = middleRb;
+
+            var virtualRoot = new GameObject();
+
+            var sensorComponent = rootObj.AddComponent<RigidBodySensorComponent>();
+            sensorComponent.RootBody = rootRb;
+            sensorComponent.Settings = new PhysicsSensorSettings
+            {
+                UseModelSpaceTranslations = true,
+                UseLocalSpaceTranslations = true,
+                UseLocalSpaceLinearVelocity = true
+            };
+            sensorComponent.VirtualRoot = virtualRoot;
+
+            var sensor = sensorComponent.CreateSensor();
+            sensor.Update();
+
+            // Note that the VirtualRoot is ignored from the observations
+            var expected = new[]
+            {
+                // Model space
+                0f, 0f, 0f, // Root pos
+                13.37f, 0f, 0f, // Middle pos
+                leafGameObj.transform.position.x, 0f, 0f, // Leaf pos
+
+                // Local space
+                0f, 0f, 0f, // Root pos
+                13.37f, 0f, 0f, // Attached pos
+                4.2f, 0f, 0f, // Leaf pos
+
+                1f, 0f, 0f, // Root vel (relative to virtual root)
+                -1f, 1f, 0f, // Attached vel
+                0f, -1f, 1f // Leaf vel
+            };
+            Assert.AreEqual(expected.Length, sensorComponent.GetObservationShape()[0]);
+            SensorTestHelper.CompareObservation(sensor, expected);
+
+            // Update the settings to only process joint observations
+            sensorComponent.Settings = new PhysicsSensorSettings
+            {
+                UseJointPositionsAndAngles = true,
+                UseJointForces = true,
+            };
+
+            sensor = sensorComponent.CreateSensor();
+            sensor.Update();
+
+            expected = new[]
+            {
+                0f, 0f, 0f, // joint1.force
+                0f, 0f, 0f, // joint1.torque
+                0f, 0f, 0f, // joint2.force
+                0f, 0f, 0f, // joint2.torque
+            };
+            SensorTestHelper.CompareObservation(sensor, expected);
+            Assert.AreEqual(expected.Length, sensorComponent.GetObservationShape()[0]);
+
+        }
+    }
+}

com.unity.ml-agents.extensions/Tests/Editor/Sensors/PoseExtractorTests.cs

+using System;
+using UnityEngine;
+using NUnit.Framework;
+using Unity.MLAgents.Extensions.Sensors;
+
+namespace Unity.MLAgents.Extensions.Tests.Sensors
+{
+    public class PoseExtractorTests
+    {
+
+        class BasicPoseExtractor : PoseExtractor
+        {
+            protected internal override Pose GetPoseAt(int index)
+            {
+                return Pose.identity;
+            }
+
+            protected  internal override Vector3 GetLinearVelocityAt(int index)
+            {
+                return Vector3.zero;
+            }
+        }
+
+        class UselessPoseExtractor : BasicPoseExtractor
+        {
+            public void Init(int[] parentIndices)
+            {
+                Setup(parentIndices);
+            }
+        }
+
+        [Test]
+        public void TestEmptyExtractor()
+        {
+            var poseExtractor = new UselessPoseExtractor();
+
+            // These should be no-ops
+            poseExtractor.UpdateLocalSpacePoses();
+            poseExtractor.UpdateModelSpacePoses();
+
+            Assert.AreEqual(0, poseExtractor.NumPoses);
+
+            // Iterating through poses and velocities should be an empty loop
+            foreach (var pose in poseExtractor.GetEnabledModelSpacePoses())
+            {
+                throw new UnityAgentsException("This shouldn't happen");
+            }
+
+            foreach (var pose in poseExtractor.GetEnabledLocalSpacePoses())
+            {
+                throw new UnityAgentsException("This shouldn't happen");
+            }
+
+            foreach (var vel in poseExtractor.GetEnabledModelSpaceVelocities())
+            {
+                throw new UnityAgentsException("This shouldn't happen");
+            }
+
+            foreach (var vel in poseExtractor.GetEnabledLocalSpaceVelocities())
+            {
+                throw new UnityAgentsException("This shouldn't happen");
+            }
+
+            // Getting a parent index should throw an index exception
+            Assert.Throws <NullReferenceException>(
+                () => poseExtractor.GetParentIndex(0)
+            );
+
+            // DisplayNodes should be empty
+            var displayNodes = poseExtractor.GetDisplayNodes();
+            Assert.AreEqual(0, displayNodes.Count);
+        }
+
+        [Test]
+        public void TestSimpleExtractor()
+        {
+            var poseExtractor = new UselessPoseExtractor();
+            var parentIndices = new[] { -1, 0 };
+            poseExtractor.Init(parentIndices);
+            Assert.AreEqual(2, poseExtractor.NumPoses);
+        }
+
+
+        /// <summary>
+        /// A simple "chain" hierarchy, where each object is parented to the one before it.
+        ///   0 <- 1 <- 2 <- ...
+        /// </summary>
+        class ChainPoseExtractor : PoseExtractor
+        {
+            public Vector3 offset;
+            public ChainPoseExtractor(int size)
+            {
+                var parents = new int[size];
+                for (var i = 0; i < size; i++)
+                {
+                    parents[i] = i - 1;
+                }
+                Setup(parents);
+            }
+
+            protected internal override Pose GetPoseAt(int index)
+            {
+                var rotation = Quaternion.identity;
+                var translation = offset + new Vector3(index, index, index);
+                return new Pose
+                {
+                    rotation = rotation,
+                    position = translation
+                };
+            }
+
+            protected  internal override Vector3 GetLinearVelocityAt(int index)
+            {
+                return Vector3.zero;
+            }
+
+        }
+
+        [Test]
+        public void TestChain()
+        {
+            var size = 4;
+            var chain = new ChainPoseExtractor(size);
+            chain.offset = new Vector3(.5f, .75f, .333f);
+
+            chain.UpdateModelSpacePoses();
+            chain.UpdateLocalSpacePoses();
+
+
+            var modelPoseIndex = 0;
+            foreach (var modelSpace in chain.GetEnabledModelSpacePoses())
+            {
+                if (modelPoseIndex == 0)
+                {
+                    // Root transforms are currently always the identity.
+                    Assert.IsTrue(modelSpace == Pose.identity);
+                }
+                else
+                {
+                    var expectedModelTranslation = new Vector3(modelPoseIndex, modelPoseIndex, modelPoseIndex);
+                    Assert.IsTrue(expectedModelTranslation == modelSpace.position);
+
+                }
+                modelPoseIndex++;
+            }
+            Assert.AreEqual(size, modelPoseIndex);
+
+            var localPoseIndex = 0;
+            foreach (var localSpace in chain.GetEnabledLocalSpacePoses())
+            {
+                if (localPoseIndex == 0)
+                {
+                    // Root transforms are currently always the identity.
+                    Assert.IsTrue(localSpace == Pose.identity);
+                }
+                else
+                {
+                    var expectedLocalTranslation = new Vector3(1, 1, 1);
+                    Assert.IsTrue(expectedLocalTranslation == localSpace.position, $"{expectedLocalTranslation} != {localSpace.position}");
+                }
+
+                localPoseIndex++;
+            }
+            Assert.AreEqual(size, localPoseIndex);
+        }
+
+        [Test]
+        public void TestChainDisplayNodes()
+        {
+            var size = 4;
+            var chain = new ChainPoseExtractor(size);
+
+            var displayNodes = chain.GetDisplayNodes();
+            Assert.AreEqual(size, displayNodes.Count);
+
+            for (var i = 0; i < size; i++)
+            {
+                var displayNode = displayNodes[i];
+                Assert.AreEqual(i, displayNode.OriginalIndex);
+                Assert.AreEqual(null, displayNode.NodeObject);
+                Assert.AreEqual(i, displayNode.Depth);
+                Assert.AreEqual(true, displayNode.Enabled);
+            }
+        }
+
+        [Test]
+        public void TestDisplayNodesLoop()
+        {
+            // Degenerate case with a loop
+            var poseExtractor = new UselessPoseExtractor();
+            poseExtractor.Init(new[] {-1, 2, 1});
+
+            // This just shouldn't blow up
+            poseExtractor.GetDisplayNodes();
+
+            // Self-loop
+            poseExtractor.Init(new[] {-1, 1});
+
+            // This just shouldn't blow up
+            poseExtractor.GetDisplayNodes();
+        }
+
+        class BadPoseExtractor : BasicPoseExtractor
+        {
+            public BadPoseExtractor()
+            {
+                var size = 2;
+                var parents = new int[size];
+                // Parents are intentionally invalid - expect -1 at root
+                for (var i = 0; i < size; i++)
+                {
+                    parents[i] = i;
+                }
+                Setup(parents);
+            }
+        }
+
+        [Test]
+        public void TestExpectedRoot()
+        {
+            Assert.Throws<UnityAgentsException>(() =>
+            {
+                var bad = new BadPoseExtractor();
+            });
+        }
+
+    }
+
+    public class PoseExtensionTests
+    {
+        [Test]
+        public void TestInverse()
+        {
+            Pose t = new Pose
+            {
+                rotation = Quaternion.AngleAxis(23.0f, new Vector3(1, 1, 1).normalized),
+                position = new Vector3(-1.0f, 2.0f, 3.0f)
+            };
+
+            var inverseT = t.Inverse();
+            var product = inverseT.Multiply(t);
+            Assert.IsTrue(Vector3.zero == product.position);
+            Assert.IsTrue(Quaternion.identity == product.rotation);
+
+            Assert.IsTrue(Pose.identity == product);
+        }
+
+    }
+}

com.unity.ml-agents.extensions/Tests/Editor/Sensors/RigidBodyPoseExtractorTests.cs

+using UnityEngine;
+using NUnit.Framework;
+using Unity.MLAgents.Extensions.Sensors;
+using UnityEditor;
+
+namespace Unity.MLAgents.Extensions.Tests.Sensors
+{
+    public class RigidBodyPoseExtractorTests
+    {
+        [TearDown]
+        public void RemoveGameObjects()
+        {
+            var objects = GameObject.FindObjectsOfType<GameObject>();
+            foreach (var o in objects)
+            {
+                UnityEngine.Object.DestroyImmediate(o);
+            }
+        }
+
+        [Test]
+        public void TestNullRoot()
+        {
+            var poseExtractor = new RigidBodyPoseExtractor(null);
+            // These should be no-ops
+            poseExtractor.UpdateLocalSpacePoses();
+            poseExtractor.UpdateModelSpacePoses();
+
+            Assert.AreEqual(0, poseExtractor.NumPoses);
+        }
+
+        [Test]
+        public void TestSingleBody()
+        {
+            var go = new GameObject();
+            var rootRb = go.AddComponent<Rigidbody>();
+            var poseExtractor = new RigidBodyPoseExtractor(rootRb);
+            Assert.AreEqual(1, poseExtractor.NumPoses);
+
+            // Also pass the GameObject
+            poseExtractor = new RigidBodyPoseExtractor(rootRb, go);
+            Assert.AreEqual(1, poseExtractor.NumPoses);
+        }
+
+        [Test]
+        public void TestNoBodiesFound()
+        {
+            // Check that if we can't find any bodies under the game object, we get an empty extractor
+            var gameObj = new GameObject();
+            var rootRb = gameObj.AddComponent<Rigidbody>();
+            var otherGameObj = new GameObject();
+            var poseExtractor = new RigidBodyPoseExtractor(rootRb, otherGameObj);
+            Assert.AreEqual(0, poseExtractor.NumPoses);
+
+            // Add an RB under the other GameObject. Constructor will find a rigid body, but not the root.
+            var otherRb = otherGameObj.AddComponent<Rigidbody>();
+            poseExtractor = new RigidBodyPoseExtractor(rootRb, otherGameObj);
+            Assert.AreEqual(0, poseExtractor.NumPoses);
+        }
+
+        [Test]
+        public void TestTwoBodies()
+        {
+            // * rootObj
+            //   - rb1
+            //   * go2
+            //     - rb2
+            //     - joint
+            var rootObj = new GameObject();
+            var rb1 = rootObj.AddComponent<Rigidbody>();
+
+            var go2 = new GameObject();
+            var rb2 = go2.AddComponent<Rigidbody>();
+            go2.transform.SetParent(rootObj.transform);
+
+            var joint = go2.AddComponent<ConfigurableJoint>();
+            joint.connectedBody = rb1;
+
+            var poseExtractor = new RigidBodyPoseExtractor(rb1);
+            Assert.AreEqual(2, poseExtractor.NumPoses);
+
+            rb1.position = new Vector3(1, 0, 0);
+            rb1.rotation = Quaternion.Euler(0, 13.37f, 0);
+            rb1.velocity = new Vector3(2, 0, 0);
+
+            Assert.AreEqual(rb1.position, poseExtractor.GetPoseAt(0).position);
+            Assert.IsTrue(rb1.rotation == poseExtractor.GetPoseAt(0).rotation);
+            Assert.AreEqual(rb1.velocity, poseExtractor.GetLinearVelocityAt(0));
+
+            // Check DisplayNodes gives expected results
+            var displayNodes = poseExtractor.GetDisplayNodes();
+            Assert.AreEqual(2, displayNodes.Count);
+            Assert.AreEqual(rb1, displayNodes[0].NodeObject);
+            Assert.AreEqual(false, displayNodes[0].Enabled);
+
+            Assert.AreEqual(rb2, displayNodes[1].NodeObject);
+            Assert.AreEqual(true, displayNodes[1].Enabled);
+        }
+
+        [Test]
+        public void TestTwoBodiesVirtualRoot()
+        {
+            // * virtualRoot
+            // * rootObj
+            //   - rb1
+            //   * go2
+            //     - rb2
+            //     - joint
+            var virtualRoot = new GameObject("I am vroot");
+
+            var rootObj = new GameObject();
+            var rb1 = rootObj.AddComponent<Rigidbody>();
+
+            var go2 = new GameObject();
+            var rb2 = go2.AddComponent<Rigidbody>();
+            go2.transform.SetParent(rootObj.transform);
+
+            var joint = go2.AddComponent<ConfigurableJoint>();
+            joint.connectedBody = rb1;
+
+            var poseExtractor = new RigidBodyPoseExtractor(rb1, null, virtualRoot);
+            Assert.AreEqual(3, poseExtractor.NumPoses);
+
+            // "body" 0 has no parent
+            Assert.AreEqual(-1, poseExtractor.GetParentIndex(0));
+
+            // body 1 has parent 0
+            Assert.AreEqual(0, poseExtractor.GetParentIndex(1));
+
+            var virtualRootPos = new Vector3(0,2,0);
+            var virtualRootRot = Quaternion.Euler(0, 42, 0);
+            virtualRoot.transform.position = virtualRootPos;
+            virtualRoot.transform.rotation = virtualRootRot;
+
+            Assert.AreEqual(virtualRootPos, poseExtractor.GetPoseAt(0).position);
+            Assert.IsTrue(virtualRootRot == poseExtractor.GetPoseAt(0).rotation);
+            Assert.AreEqual(Vector3.zero, poseExtractor.GetLinearVelocityAt(0));
+
+            // Same as above test, but using index 1
+            rb1.position = new Vector3(1, 0, 0);
+            rb1.rotation = Quaternion.Euler(0, 13.37f, 0);
+            rb1.velocity = new Vector3(2, 0, 0);
+
+            Assert.AreEqual(rb1.position, poseExtractor.GetPoseAt(1).position);
+            Assert.IsTrue(rb1.rotation == poseExtractor.GetPoseAt(1).rotation);
+            Assert.AreEqual(rb1.velocity, poseExtractor.GetLinearVelocityAt(1));
+        }
+
+        [Test]
+        public void TestBodyPosesEnabledDictionary()
+        {
+            // * rootObj
+            //   - rb1
+            //   * go2
+            //     - rb2
+            //     - joint
+            var rootObj = new GameObject();
+            var rb1 = rootObj.AddComponent<Rigidbody>();
+
+            var go2 = new GameObject();
+            var rb2 = go2.AddComponent<Rigidbody>();
+            go2.transform.SetParent(rootObj.transform);
+
+            var joint = go2.AddComponent<ConfigurableJoint>();
+            joint.connectedBody = rb1;
+
+            var poseExtractor = new RigidBodyPoseExtractor(rb1);
+
+            // Expect the root body disabled and the attached one enabled.
+            Assert.IsFalse(poseExtractor.IsPoseEnabled(0));
+            Assert.IsTrue(poseExtractor.IsPoseEnabled(1));
+            var bodyPosesEnabled = poseExtractor.GetBodyPosesEnabled();
+            Assert.IsFalse(bodyPosesEnabled[rb1]);
+            Assert.IsTrue(bodyPosesEnabled[rb2]);
+
+            // Swap the values
+            bodyPosesEnabled[rb1] = true;
+            bodyPosesEnabled[rb2] = false;
+
+            var poseExtractor2 = new RigidBodyPoseExtractor(rb1, null, null, bodyPosesEnabled);
+            Assert.IsTrue(poseExtractor2.IsPoseEnabled(0));
+            Assert.IsFalse(poseExtractor2.IsPoseEnabled(1));
+
+
+        }
+    }
+}

ml-agents/mlagents/trainers/tf/model_serialization.py

+from distutils.util import strtobool
+import os
+from typing import Any, List, Set
+from distutils.version import LooseVersion
+
+try:
+    from tf2onnx.tfonnx import process_tf_graph, tf_optimize
+    from tf2onnx import optimizer
+
+    ONNX_EXPORT_ENABLED = True
+except ImportError:
+    # Either onnx and tf2onnx not installed, or they're not compatible with the version of tensorflow
+    ONNX_EXPORT_ENABLED = False
+    pass
+
+from mlagents.tf_utils import tf
+
+from tensorflow.python.platform import gfile
+from tensorflow.python.framework import graph_util
+
+from mlagents_envs.logging_util import get_logger
+from mlagents.trainers.settings import SerializationSettings
+from mlagents.trainers.tf import tensorflow_to_barracuda as tf2bc
+
+if LooseVersion(tf.__version__) < LooseVersion("1.12.0"):
+    # ONNX is only tested on 1.12.0 and later
+    ONNX_EXPORT_ENABLED = False
+
+logger = get_logger(__name__)
+
+
+POSSIBLE_INPUT_NODES = frozenset(
+    [
+        "action_masks",
+        "epsilon",
+        "prev_action",
+        "recurrent_in",
+        "sequence_length",
+        "vector_observation",
+    ]
+)
+
+POSSIBLE_OUTPUT_NODES = frozenset(
+    ["action", "action_probs", "recurrent_out", "value_estimate"]
+)
+
+MODEL_CONSTANTS = frozenset(
+    [
+        "action_output_shape",
+        "is_continuous_control",
+        "memory_size",
+        "version_number",
+        "trainer_major_version",
+        "trainer_minor_version",
+        "trainer_patch_version",
+    ]
+)
+VISUAL_OBSERVATION_PREFIX = "visual_observation_"
+
+
+def export_policy_model(
+    model_path: str,
+    output_filepath: str,
+    behavior_name: str,
+    graph: tf.Graph,
+    sess: tf.Session,
+) -> None:
+    """
+    Exports a TF graph for a Policy to .nn and/or .onnx format for Unity embedding.
+
+    :param output_filepath: file path to output the model (without file suffix)
+    :param behavior_name: behavior name of the trained model
+    :param graph: Tensorflow Graph for the policy
+    :param sess: Tensorflow session for the policy
+    """
+    frozen_graph_def = _make_frozen_graph(behavior_name, graph, sess)
+    if not os.path.exists(output_filepath):
+        os.makedirs(output_filepath)
+    # Save frozen graph
+    frozen_graph_def_path = model_path + "/frozen_graph_def.pb"
+    with gfile.GFile(frozen_graph_def_path, "wb") as f:
+        f.write(frozen_graph_def.SerializeToString())
+
+    # Convert to barracuda
+    if SerializationSettings.convert_to_barracuda:
+        tf2bc.convert(frozen_graph_def_path, f"{output_filepath}.nn")
+        logger.info(f"Exported {output_filepath}.nn")
+
+    # Save to onnx too (if we were able to import it)
+    if ONNX_EXPORT_ENABLED:
+        if SerializationSettings.convert_to_onnx:
+            try:
+                onnx_graph = convert_frozen_to_onnx(behavior_name, frozen_graph_def)
+                onnx_output_path = f"{output_filepath}.onnx"
+                with open(onnx_output_path, "wb") as f:
+                    f.write(onnx_graph.SerializeToString())
+                logger.info(f"Converting to {onnx_output_path}")
+            except Exception:
+                # Make conversion errors fatal depending on environment variables (only done during CI)
+                if _enforce_onnx_conversion():
+                    raise
+                logger.exception(
+                    "Exception trying to save ONNX graph. Please report this error on "
+                    "https://github.com/Unity-Technologies/ml-agents/issues and "
+                    "attach a copy of frozen_graph_def.pb"
+                )
+
+    else:
+        if _enforce_onnx_conversion():
+            raise RuntimeError(
+                "ONNX conversion enforced, but couldn't import dependencies."
+            )
+
+
+def _make_frozen_graph(
+    behavior_name: str, graph: tf.Graph, sess: tf.Session
+) -> tf.GraphDef:
+    with graph.as_default():
+        target_nodes = ",".join(_process_graph(behavior_name, graph))
+        graph_def = graph.as_graph_def()
+        output_graph_def = graph_util.convert_variables_to_constants(
+            sess, graph_def, target_nodes.replace(" ", "").split(",")
+        )
+    return output_graph_def
+
+
+def convert_frozen_to_onnx(behavior_name: str, frozen_graph_def: tf.GraphDef) -> Any:
+    # This is basically https://github.com/onnx/tensorflow-onnx/blob/master/tf2onnx/convert.py
+
+    inputs = _get_input_node_names(frozen_graph_def)
+    outputs = _get_output_node_names(frozen_graph_def)
+    logger.info(f"onnx export - inputs:{inputs} outputs:{outputs}")
+
+    frozen_graph_def = tf_optimize(
+        inputs, outputs, frozen_graph_def, fold_constant=True
+    )
+
+    with tf.Graph().as_default() as tf_graph:
+        tf.import_graph_def(frozen_graph_def, name="")
+    with tf.Session(graph=tf_graph):
+        g = process_tf_graph(
+            tf_graph,
+            input_names=inputs,
+            output_names=outputs,
+            opset=SerializationSettings.onnx_opset,
+        )
+
+    onnx_graph = optimizer.optimize_graph(g)
+    model_proto = onnx_graph.make_model(behavior_name)
+
+    return model_proto
+
+
+def _get_input_node_names(frozen_graph_def: Any) -> List[str]:
+    """
+    Get the list of input node names from the graph.
+    Names are suffixed with ":0"
+    """
+    node_names = _get_frozen_graph_node_names(frozen_graph_def)
+    input_names = node_names & POSSIBLE_INPUT_NODES
+
+    # Check visual inputs sequentially, and exit as soon as we don't find one
+    vis_index = 0
+    while True:
+        vis_node_name = f"{VISUAL_OBSERVATION_PREFIX}{vis_index}"
+        if vis_node_name in node_names:
+            input_names.add(vis_node_name)
+        else:
+            break
+        vis_index += 1
+    # Append the port
+    return [f"{n}:0" for n in input_names]
+
+
+def _get_output_node_names(frozen_graph_def: Any) -> List[str]:
+    """
+    Get the list of output node names from the graph.
+    Also include constants, so that they will be readable by the
+    onnx importer.
+    Names are suffixed with ":0"
+    """
+    node_names = _get_frozen_graph_node_names(frozen_graph_def)
+    output_names = node_names & (POSSIBLE_OUTPUT_NODES | MODEL_CONSTANTS)
+    # Append the port
+    return [f"{n}:0" for n in output_names]
+
+
+def _get_frozen_graph_node_names(frozen_graph_def: Any) -> Set[str]:
+    """
+    Get all the node names from the graph.
+    """
+    names = set()
+    for node in frozen_graph_def.node:
+        names.add(node.name)
+    return names
+
+
+def _process_graph(behavior_name: str, graph: tf.Graph) -> List[str]:
+    """
+    Gets the list of the output nodes present in the graph for inference
+    :return: list of node names
+    """
+    all_nodes = [x.name for x in graph.as_graph_def().node]
+    nodes = [x for x in all_nodes if x in POSSIBLE_OUTPUT_NODES | MODEL_CONSTANTS]
+    logger.info("List of nodes to export for behavior :" + behavior_name)
+    for n in nodes:
+        logger.info("\t" + n)
+    return nodes
+
+
+def _enforce_onnx_conversion() -> bool:
+    env_var_name = "TEST_ENFORCE_ONNX_CONVERSION"
+    if env_var_name not in os.environ:
+        return False
+
+    val = os.environ[env_var_name]
+    try:
+        # This handles e.g. "false" converting reasonably to False
+        return strtobool(val)
+    except Exception:
+        return False

ml-agents/mlagents/trainers/tests/torch/test_reward_providers/test_curiosity.py

+import numpy as np
+import pytest
+import torch
+from mlagents.trainers.torch.components.reward_providers import (
+    CuriosityRewardProvider,
+    create_reward_provider,
+)
+from mlagents_envs.base_env import BehaviorSpec, ActionType
+from mlagents.trainers.settings import CuriositySettings, RewardSignalType
+from mlagents.trainers.tests.torch.test_reward_providers.utils import (
+    create_agent_buffer,
+)
+
+SEED = [42]
+
+
+@pytest.mark.parametrize(
+    "behavior_spec",
+    [
+        BehaviorSpec([(10,)], ActionType.CONTINUOUS, 5),
+        BehaviorSpec([(10,)], ActionType.DISCRETE, (2, 3)),
+    ],
+)
+def test_construction(behavior_spec: BehaviorSpec) -> None:
+    curiosity_settings = CuriositySettings(32, 0.01)
+    curiosity_settings.strength = 0.1
+    curiosity_rp = CuriosityRewardProvider(behavior_spec, curiosity_settings)
+    assert curiosity_rp.strength == 0.1
+    assert curiosity_rp.name == "Curiosity"
+
+
+@pytest.mark.parametrize(
+    "behavior_spec",
+    [
+        BehaviorSpec([(10,)], ActionType.CONTINUOUS, 5),
+        BehaviorSpec([(10,), (64, 66, 3), (84, 86, 1)], ActionType.CONTINUOUS, 5),
+        BehaviorSpec([(10,), (64, 66, 1)], ActionType.DISCRETE, (2, 3)),
+        BehaviorSpec([(10,)], ActionType.DISCRETE, (2,)),
+    ],
+)
+def test_factory(behavior_spec: BehaviorSpec) -> None:
+    curiosity_settings = CuriositySettings(32, 0.01)
+    curiosity_rp = create_reward_provider(
+        RewardSignalType.CURIOSITY, behavior_spec, curiosity_settings
+    )
+    assert curiosity_rp.name == "Curiosity"
+
+
+@pytest.mark.parametrize("seed", SEED)
+@pytest.mark.parametrize(
+    "behavior_spec",
+    [
+        BehaviorSpec([(10,), (64, 66, 3), (24, 26, 1)], ActionType.CONTINUOUS, 5),
+        BehaviorSpec([(10,)], ActionType.DISCRETE, (2, 3)),
+        BehaviorSpec([(10,)], ActionType.DISCRETE, (2,)),
+    ],
+)
+def test_reward_decreases(behavior_spec: BehaviorSpec, seed: int) -> None:
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    curiosity_settings = CuriositySettings(32, 0.01)
+    curiosity_rp = CuriosityRewardProvider(behavior_spec, curiosity_settings)
+    buffer = create_agent_buffer(behavior_spec, 5)
+    curiosity_rp.update(buffer)
+    reward_old = curiosity_rp.evaluate(buffer)[0]
+    for _ in range(10):
+        curiosity_rp.update(buffer)
+        reward_new = curiosity_rp.evaluate(buffer)[0]
+        assert reward_new < reward_old
+        reward_old = reward_new
+
+
+@pytest.mark.parametrize("seed", SEED)
+@pytest.mark.parametrize(
+    "behavior_spec", [BehaviorSpec([(10,)], ActionType.CONTINUOUS, 5)]
+)
+def test_continuous_action_prediction(behavior_spec: BehaviorSpec, seed: int) -> None:
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    curiosity_settings = CuriositySettings(32, 0.1)
+    curiosity_rp = CuriosityRewardProvider(behavior_spec, curiosity_settings)
+    buffer = create_agent_buffer(behavior_spec, 5)
+    for _ in range(200):
+        curiosity_rp.update(buffer)
+    prediction = curiosity_rp._network.predict_action(buffer)[0].detach()
+    target = buffer["actions"][0]
+    error = float(torch.mean((prediction - target) ** 2))
+    assert error < 0.001
+
+
+@pytest.mark.parametrize("seed", SEED)
+@pytest.mark.parametrize(
+    "behavior_spec",
+    [
+        BehaviorSpec([(10,), (64, 66, 3)], ActionType.CONTINUOUS, 5),
+        BehaviorSpec([(10,)], ActionType.DISCRETE, (2, 3)),
+        BehaviorSpec([(10,)], ActionType.DISCRETE, (2,)),
+    ],
+)
+def test_next_state_prediction(behavior_spec: BehaviorSpec, seed: int) -> None:
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    curiosity_settings = CuriositySettings(32, 0.1)
+    curiosity_rp = CuriosityRewardProvider(behavior_spec, curiosity_settings)
+    buffer = create_agent_buffer(behavior_spec, 5)
+    for _ in range(100):
+        curiosity_rp.update(buffer)
+    prediction = curiosity_rp._network.predict_next_state(buffer)[0]
+    target = curiosity_rp._network.get_next_state(buffer)[0]
+    error = float(torch.mean((prediction - target) ** 2).detach())
+    assert error < 0.001

ml-agents/mlagents/trainers/tests/torch/test_reward_providers/test_extrinsic.py

+import pytest
+from mlagents.trainers.torch.components.reward_providers import (
+    ExtrinsicRewardProvider,
+    create_reward_provider,
+)
+from mlagents_envs.base_env import BehaviorSpec, ActionType
+from mlagents.trainers.settings import RewardSignalSettings, RewardSignalType
+from mlagents.trainers.tests.torch.test_reward_providers.utils import (
+    create_agent_buffer,
+)
+
+
+@pytest.mark.parametrize(
+    "behavior_spec",
+    [
+        BehaviorSpec([(10,)], ActionType.CONTINUOUS, 5),
+        BehaviorSpec([(10,)], ActionType.DISCRETE, (2, 3)),
+    ],
+)
+def test_construction(behavior_spec: BehaviorSpec) -> None:
+    settings = RewardSignalSettings()
+    settings.gamma = 0.2
+    extrinsic_rp = ExtrinsicRewardProvider(behavior_spec, settings)
+    assert extrinsic_rp.gamma == 0.2
+    assert extrinsic_rp.name == "Extrinsic"
+
+
+@pytest.mark.parametrize(
+    "behavior_spec",
+    [
+        BehaviorSpec([(10,)], ActionType.CONTINUOUS, 5),
+        BehaviorSpec([(10,)], ActionType.DISCRETE, (2, 3)),
+    ],
+)
+def test_factory(behavior_spec: BehaviorSpec) -> None:
+    settings = RewardSignalSettings()
+    extrinsic_rp = create_reward_provider(
+        RewardSignalType.EXTRINSIC, behavior_spec, settings
+    )
+    assert extrinsic_rp.name == "Extrinsic"
+
+
+@pytest.mark.parametrize("reward", [2.0, 3.0, 4.0])
+@pytest.mark.parametrize(
+    "behavior_spec",
+    [
+        BehaviorSpec([(10,)], ActionType.CONTINUOUS, 5),
+        BehaviorSpec([(10,)], ActionType.DISCRETE, (2, 3)),
+    ],
+)
+def test_reward(behavior_spec: BehaviorSpec, reward: float) -> None:
+    buffer = create_agent_buffer(behavior_spec, 1000, reward)
+    settings = RewardSignalSettings()
+    extrinsic_rp = ExtrinsicRewardProvider(behavior_spec, settings)
+    generated_rewards = extrinsic_rp.evaluate(buffer)
+    assert (generated_rewards == reward).all()

ml-agents/mlagents/trainers/tests/torch/test_reward_providers/test_gail.py

+from typing import Any
+import numpy as np
+import pytest
+from unittest.mock import patch
+import torch
+import os
+from mlagents.trainers.torch.components.reward_providers import (
+    GAILRewardProvider,
+    create_reward_provider,
+)
+from mlagents_envs.base_env import BehaviorSpec, ActionType
+from mlagents.trainers.settings import GAILSettings, RewardSignalType
+from mlagents.trainers.tests.torch.test_reward_providers.utils import (
+    create_agent_buffer,
+)
+from mlagents.trainers.torch.components.reward_providers.gail_reward_provider import (
+    DiscriminatorNetwork,
+)
+
+CONTINUOUS_PATH = (
+    os.path.join(os.path.dirname(os.path.abspath(__file__)), os.pardir, os.pardir)
+    + "/test.demo"
+)
+DISCRETE_PATH = (
+    os.path.join(os.path.dirname(os.path.abspath(__file__)), os.pardir, os.pardir)
+    + "/testdcvis.demo"
+)
+SEED = [42]
+
+
+@pytest.mark.parametrize(
+    "behavior_spec", [BehaviorSpec([(8,)], ActionType.CONTINUOUS, 2)]
+)
+def test_construction(behavior_spec: BehaviorSpec) -> None:
+    gail_settings = GAILSettings(demo_path=CONTINUOUS_PATH)
+    gail_rp = GAILRewardProvider(behavior_spec, gail_settings)
+    assert gail_rp.name == "GAIL"
+
+
+@pytest.mark.parametrize(
+    "behavior_spec", [BehaviorSpec([(8,)], ActionType.CONTINUOUS, 2)]
+)
+def test_factory(behavior_spec: BehaviorSpec) -> None:
+    gail_settings = GAILSettings(demo_path=CONTINUOUS_PATH)
+    gail_rp = create_reward_provider(
+        RewardSignalType.GAIL, behavior_spec, gail_settings
+    )
+    assert gail_rp.name == "GAIL"
+
+
+@pytest.mark.parametrize("seed", SEED)
+@pytest.mark.parametrize(
+    "behavior_spec",
+    [
+        BehaviorSpec([(8,), (24, 26, 1)], ActionType.CONTINUOUS, 2),
+        BehaviorSpec([(50,)], ActionType.DISCRETE, (2, 3, 3, 3)),
+        BehaviorSpec([(10,)], ActionType.DISCRETE, (20,)),
+    ],
+)
+@pytest.mark.parametrize("use_actions", [False, True])
+@patch(
+    "mlagents.trainers.torch.components.reward_providers.gail_reward_provider.demo_to_buffer"
+)
+def test_reward_decreases(
+    demo_to_buffer: Any, use_actions: bool, behavior_spec: BehaviorSpec, seed: int
+) -> None:
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    buffer_expert = create_agent_buffer(behavior_spec, 1000)
+    buffer_policy = create_agent_buffer(behavior_spec, 1000)
+    demo_to_buffer.return_value = None, buffer_expert
+    gail_settings = GAILSettings(
+        demo_path="", learning_rate=0.05, use_vail=False, use_actions=use_actions
+    )
+    gail_rp = create_reward_provider(
+        RewardSignalType.GAIL, behavior_spec, gail_settings
+    )
+
+    init_reward_expert = gail_rp.evaluate(buffer_expert)[0]
+    init_reward_policy = gail_rp.evaluate(buffer_policy)[0]
+
+    for _ in range(10):
+        gail_rp.update(buffer_policy)
+        reward_expert = gail_rp.evaluate(buffer_expert)[0]
+        reward_policy = gail_rp.evaluate(buffer_policy)[0]
+        assert reward_expert >= 0  # GAIL / VAIL reward always positive
+        assert reward_policy >= 0
+    reward_expert = gail_rp.evaluate(buffer_expert)[0]
+    reward_policy = gail_rp.evaluate(buffer_policy)[0]
+    assert reward_expert > reward_policy  # Expert reward greater than non-expert reward
+    assert (
+        reward_expert > init_reward_expert
+    )  # Expert reward getting better as network trains
+    assert (
+        reward_policy < init_reward_policy
+    )  # Non-expert reward getting worse as network trains
+
+
+@pytest.mark.parametrize("seed", SEED)
+@pytest.mark.parametrize(
+    "behavior_spec",
+    [
+        BehaviorSpec([(8,)], ActionType.CONTINUOUS, 2),
+        BehaviorSpec([(10,)], ActionType.DISCRETE, (2, 3, 3, 3)),
+        BehaviorSpec([(10,)], ActionType.DISCRETE, (20,)),
+    ],
+)
+@pytest.mark.parametrize("use_actions", [False, True])
+@patch(
+    "mlagents.trainers.torch.components.reward_providers.gail_reward_provider.demo_to_buffer"
+)
+def test_reward_decreases_vail(
+    demo_to_buffer: Any, use_actions: bool, behavior_spec: BehaviorSpec, seed: int
+) -> None:
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    buffer_expert = create_agent_buffer(behavior_spec, 1000)
+    buffer_policy = create_agent_buffer(behavior_spec, 1000)
+    demo_to_buffer.return_value = None, buffer_expert
+    gail_settings = GAILSettings(
+        demo_path="", learning_rate=0.005, use_vail=True, use_actions=use_actions
+    )
+    DiscriminatorNetwork.initial_beta = 0.0
+    # we must set the initial value of beta to 0 for testing
+    # If we do not, the kl-loss will dominate early and will block the estimator
+    gail_rp = create_reward_provider(
+        RewardSignalType.GAIL, behavior_spec, gail_settings
+    )
+
+    for _ in range(100):
+        gail_rp.update(buffer_policy)
+        reward_expert = gail_rp.evaluate(buffer_expert)[0]
+        reward_policy = gail_rp.evaluate(buffer_policy)[0]
+        assert reward_expert >= 0  # GAIL / VAIL reward always positive
+        assert reward_policy >= 0
+    reward_expert = gail_rp.evaluate(buffer_expert)[0]
+    reward_policy = gail_rp.evaluate(buffer_policy)[0]
+    assert reward_expert > reward_policy  # Expert reward greater than non-expert reward

ml-agents/mlagents/trainers/tests/torch/test_reward_providers/utils.py

+import numpy as np
+from mlagents.trainers.buffer import AgentBuffer
+from mlagents_envs.base_env import BehaviorSpec
+from mlagents.trainers.trajectory import SplitObservations
+
+
+def create_agent_buffer(
+    behavior_spec: BehaviorSpec, number: int, reward: float = 0.0
+) -> AgentBuffer:
+    buffer = AgentBuffer()
+    curr_observations = [
+        np.random.normal(size=shape) for shape in behavior_spec.observation_shapes
+    ]
+    next_observations = [
+        np.random.normal(size=shape) for shape in behavior_spec.observation_shapes
+    ]
+    action = behavior_spec.create_random_action(1)[0, :]
+    for _ in range(number):
+        curr_split_obs = SplitObservations.from_observations(curr_observations)
+        next_split_obs = SplitObservations.from_observations(next_observations)
+        for i, _ in enumerate(curr_split_obs.visual_observations):
+            buffer["visual_obs%d" % i].append(curr_split_obs.visual_observations[i])
+            buffer["next_visual_obs%d" % i].append(
+                next_split_obs.visual_observations[i]
+            )
+        buffer["vector_obs"].append(curr_split_obs.vector_observations)
+        buffer["next_vector_in"].append(next_split_obs.vector_observations)
+        buffer["actions"].append(action)
+        buffer["done"].append(np.zeros(1, dtype=np.float32))
+        buffer["reward"].append(np.ones(1, dtype=np.float32) * reward)
+        buffer["masks"].append(np.ones(1, dtype=np.float32))
+    return buffer

ml-agents/mlagents/trainers/tests/torch/test_bcmodule.py

+from unittest.mock import MagicMock
+import pytest
+import mlagents.trainers.tests.mock_brain as mb
+
+import numpy as np
+import os
+
+from mlagents.trainers.policy.torch_policy import TorchPolicy
+from mlagents.trainers.torch.components.bc.module import BCModule
+from mlagents.trainers.settings import (
+    TrainerSettings,
+    BehavioralCloningSettings,
+    NetworkSettings,
+)
+
+
+def create_bc_module(mock_behavior_specs, bc_settings, use_rnn, tanhresample):
+    # model_path = env.external_brain_names[0]
+    trainer_config = TrainerSettings()
+    trainer_config.network_settings.memory = (
+        NetworkSettings.MemorySettings() if use_rnn else None
+    )
+    policy = TorchPolicy(
+        0, mock_behavior_specs, trainer_config, tanhresample, tanhresample
+    )
+    bc_module = BCModule(
+        policy,
+        settings=bc_settings,
+        policy_learning_rate=trainer_config.hyperparameters.learning_rate,
+        default_batch_size=trainer_config.hyperparameters.batch_size,
+        default_num_epoch=3,
+    )
+    return bc_module
+
+
+# Test default values
+def test_bcmodule_defaults():
+    # See if default values match
+    mock_specs = mb.create_mock_3dball_behavior_specs()
+    bc_settings = BehavioralCloningSettings(
+        demo_path=os.path.dirname(os.path.abspath(__file__)) + "/" + "test.demo"
+    )
+    bc_module = create_bc_module(mock_specs, bc_settings, False, False)
+    assert bc_module.num_epoch == 3
+    assert bc_module.batch_size == TrainerSettings().hyperparameters.batch_size
+    # Assign strange values and see if it overrides properly
+    bc_settings = BehavioralCloningSettings(
+        demo_path=os.path.dirname(os.path.abspath(__file__)) + "/" + "test.demo",
+        num_epoch=100,
+        batch_size=10000,
+    )
+    bc_module = create_bc_module(mock_specs, bc_settings, False, False)
+    assert bc_module.num_epoch == 100
+    assert bc_module.batch_size == 10000
+
+
+# Test with continuous control env and vector actions
+@pytest.mark.parametrize("is_sac", [True, False], ids=["sac", "ppo"])
+def test_bcmodule_update(is_sac):
+    mock_specs = mb.create_mock_3dball_behavior_specs()
+    bc_settings = BehavioralCloningSettings(
+        demo_path=os.path.dirname(os.path.abspath(__file__)) + "/" + "test.demo"
+    )
+    bc_module = create_bc_module(mock_specs, bc_settings, False, is_sac)
+    stats = bc_module.update()
+    for _, item in stats.items():
+        assert isinstance(item, np.float32)
+
+
+# Test with constant pretraining learning rate
+@pytest.mark.parametrize("is_sac", [True, False], ids=["sac", "ppo"])
+def test_bcmodule_constant_lr_update(is_sac):
+    mock_specs = mb.create_mock_3dball_behavior_specs()
+    bc_settings = BehavioralCloningSettings(
+        demo_path=os.path.dirname(os.path.abspath(__file__)) + "/" + "test.demo",
+        steps=0,
+    )
+    bc_module = create_bc_module(mock_specs, bc_settings, False, is_sac)
+    stats = bc_module.update()
+    for _, item in stats.items():
+        assert isinstance(item, np.float32)
+    old_learning_rate = bc_module.current_lr
+
+    _ = bc_module.update()
+    assert old_learning_rate == bc_module.current_lr
+
+
+# Test with constant pretraining learning rate
+@pytest.mark.parametrize("is_sac", [True, False], ids=["sac", "ppo"])
+def test_bcmodule_linear_lr_update(is_sac):
+    mock_specs = mb.create_mock_3dball_behavior_specs()
+    bc_settings = BehavioralCloningSettings(
+        demo_path=os.path.dirname(os.path.abspath(__file__)) + "/" + "test.demo",
+        steps=100,
+    )
+    bc_module = create_bc_module(mock_specs, bc_settings, False, is_sac)
+    # Should decay by 10/100 * 0.0003 = 0.00003
+    bc_module.policy.get_current_step = MagicMock(return_value=10)
+    old_learning_rate = bc_module.current_lr
+    _ = bc_module.update()
+    assert old_learning_rate - 0.00003 == pytest.approx(bc_module.current_lr, abs=0.01)
+
+
+# Test with RNN
+@pytest.mark.parametrize("is_sac", [True, False], ids=["sac", "ppo"])
+def test_bcmodule_rnn_update(is_sac):
+    mock_specs = mb.create_mock_3dball_behavior_specs()
+    bc_settings = BehavioralCloningSettings(
+        demo_path=os.path.dirname(os.path.abspath(__file__)) + "/" + "test.demo"
+    )
+    bc_module = create_bc_module(mock_specs, bc_settings, True, is_sac)
+    stats = bc_module.update()
+    for _, item in stats.items():
+        assert isinstance(item, np.float32)
+
+
+# Test with discrete control and visual observations
+@pytest.mark.parametrize("is_sac", [True, False], ids=["sac", "ppo"])
+def test_bcmodule_dc_visual_update(is_sac):
+    mock_specs = mb.create_mock_banana_behavior_specs()
+    bc_settings = BehavioralCloningSettings(
+        demo_path=os.path.dirname(os.path.abspath(__file__)) + "/" + "testdcvis.demo"
+    )
+    bc_module = create_bc_module(mock_specs, bc_settings, False, is_sac)
+    stats = bc_module.update()
+    for _, item in stats.items():
+        assert isinstance(item, np.float32)
+
+
+# Test with discrete control, visual observations and RNN
+@pytest.mark.parametrize("is_sac", [True, False], ids=["sac", "ppo"])
+def test_bcmodule_rnn_dc_update(is_sac):
+    mock_specs = mb.create_mock_banana_behavior_specs()
+    bc_settings = BehavioralCloningSettings(
+        demo_path=os.path.dirname(os.path.abspath(__file__)) + "/" + "testdcvis.demo"
+    )
+    bc_module = create_bc_module(mock_specs, bc_settings, True, is_sac)
+    stats = bc_module.update()
+    for _, item in stats.items():
+        assert isinstance(item, np.float32)
+
+
+if __name__ == "__main__":
+    pytest.main()

ml-agents/mlagents/trainers/tests/torch/testdcvis.demo

+

bcvis& 
-
****:VisualFoodCollectorLearning�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�X�ndE� �D,�������	D$|!19?EF�--�@0^��ڡ�oU�z�}�glK��v��8�8�}�ww��?�,�yg�ĉw#���o;�v~�h�~���)��o;��	�c��Mo^;���/p��/p��mߎ};��@;���v۷cߎ�'Ў�7����طc��	,�د.//G������<:`��U_on�T]��W#'�W���׉�ÎN޵rqq�r��S��:������ms��:9
���-o��y�䅑W̟�������g~�9��f�?>�������׷��?�?������ֱ.�/�S��v��\]_'K���W���z�������:�!�R�&�M~�>K�ۿ�<H�<www �$��wIډos��\o����3\��N �f�>��yr��ӈG	�0��;;���UX���,R8I�H$G����<�g?9(K�}V�T �K�ʘ�z
�˗_eL��n����@A�`���`����#������V5:)��͌EK������((%�I��7����NN
H���[��LB��,��f���@�0D��ّ��%�M�NU����G�~jP�������t�I��!#����Iç��ؗ>� )��T�F4�Y��1Ĥ!%���ن�
�0��9�	�� $�2�d"���<K�����11ϼc��yQR��Q�T��Z�@� �Vb�]"�u�FN�j7�#�P�(\���&q�D��h+���m%����%�ȪS���%�a����Q��L�EI�$�(�8�`�^�D[dE?�6�ʧ�6A��R)bI��1��{I�+��*�!��E����h��-=_xH]��G0\�3�V�f.�cI
��gp����3�]���
+F�|���d����	�x�����(�{
F��[
AL����|�hv��<���]V4�^���pAH�W�H"d��G����P�uż���A�A$��8�O��;��Ó��ý�UqpI^�(P�g�|���{��?G@[1��M�9h�E���샔�6���b��j�m-h�ڡ��"q"&
+b1�,)��p'�����s
+͗b�$P
����IԳN��O0s��J�)�Q����=�bs`��s)&4|H�@ujٸ��6�0�䃞�{\x�t8����6��I��Wꝏ�����c��k�k.���؈�Kh��������rD�^��fRݐ��.��BG�/<'�:C�Yt��S�&b�]%תz�Ǿ�R������}��� ��¿�&�>���t��uel#��w���T����j�����K�X��Ô���r��.#?�wqMD��D ��r�`(�N���'�;m��ko�����c���I�R�i�Э�/��6��d�`�9�j��D�̢O-�����7�S�Qv]U�iD� ���|5ÕI�IEND�B`�j
+
"
+�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�X�ndE� �D,�������	D$|!19?EF�--�@0^��ڡ�oU�z�}�glK��v��8�8�}�ww��?�,�yg�ĉw#���o;�v~�h�~���)��o;��	�c��Mo^;���/p��/p��mߎ};��@;���v۷cߎ�'Ў�7����طc��	,�د.//G������<:`��U_on�T]��W#'�W���׉�ÎN޵rqq�r��S��:������ms��:9
���-o��y�䅑W̟�������g~�9��f�?>�������׷��?�?������ֱ.�/�S��v��\]_'K���W���z�������:�!�R�&�M~�>K�ۿ�<H�<www �$��wIډos��\o����3\��N �f�>��yr��ӈG	�0��;;���UX���,R8I�H$G����<�g?9(K�}V�T �K�ʘ�z
�˗_eL��n����@A�`���`����#������V5:)��͌EK������((%�I��7����NN
H���[��LB��,��f���@�0D��ّ��%�M�NU����G�~jP�������t�I��!#����Iç��ؗ>� )��T�F4�Y��1Ĥ!%���ن�
�0��9�	�� $�2�d"���<K�����11ϼc��yQR��Q�T��Z�@� �Vb�]"�u�FN�j7�#�P�(\���&q�D��h+���m%����%�ȪS���%�a����Q��L�EI�$�(�8�`�^�D[dE?�6�ʧ�6A��R)bI��1��{I�+��*�!��E����h��-=_xH]��G0\�3�V�f.�cI
��gp����3�]���
+F�|���d����	�x�����(�{
F��[
AL����|�hv��<���]V4�^���pAH�W�H"d��G����P�uż���A�A$��8�O��;��Ó��ý�UqpI^�(P�g�|���{��?G@[1��M�9h�E���샔�6���b��j�m-h�ڡ��"q"&
+b1�,)��p'�����s
+͗b�$P
����IԳN��O0s��J�)�Q����=�bs`��s)&4|H�@ujٸ��6�0�䃞�{\x�t8����6��I��Wꝏ�����c��k�k.���؈�Kh��������rD�^��fRݐ��.��BG�/<'�:C�Yt��S�&b�]%תz�Ǿ�R������}��� ��¿�&�>���t��uel#��w���T����j�����K�X��Ô���r��.#?�wqMD��D ��r�`(�N���'�;m��ko�����c���I�R�i�Э�/��6��d�`�9�j��D�̢O-�����7�S�Qv]U�iD� ���|5ÕI�IEND�B`�j
+
"
+�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�X�ndE� �D,�������	D$|!19?EF�--�@0^��ڡ�oU�z�}�glK��v��8�8�}�ww��?�,�yg�ĉw#���o;�v~�h�~���)��o;��	�c��Mo^;���/p��/p��mߎ};��@;���v۷cߎ�'Ў�7����طc��	,�د.//G������<:`��U_on�T]��W#'�W���׉�ÎN޵rqq�r��S��:������ms��:9
���-o��y�䅑W̟�������g~�9��f�?>�������׷��?�?������ֱ.�/�S��v��\]_'K���W���z�������:�!�R�&�M~�>K�ۿ�<H�<www �$��wIډos��\o����3\��N �f�>��yr��ӈG	�0��;;���UX���,R8I�H$G����<�g?9(K�}V�T �K�ʘ�z
�˗_eL��n����@A�`���`����#������V5:)��͌EK������((%�I��7����NN
H���[��LB��,��f���@�0D��ّ��%�M�NU����G�~jP�������t�I��!#����Iç��ؗ>� )��T�F4�Y��1Ĥ!%���ن�
�0��9�	�� $�2�d"���<K�����11ϼc��yQR��Q�T��Z�@� �Vb�]"�u�FN�j7�#�P�(\���&q�D��h+���m%����%�ȪS���%�a����Q��L�EI�$�(�8�`�^�D[dE?�6�ʧ�6A��R)bI��1��{I�+��*�!��E����h��-=_xH]��G0\�3�V�f.�cI
��gp����3�]���
+F�|���d����	�x�����(�{
F��[
AL����|�hv��<���]V4�^���pAH�W�H"d��G����P�uż���A�A$��8�O��;��Ó��ý�UqpI^�(P�g�|���{��?G@[1��M�9h�E���샔�6���b��j�m-h�ڡ��"q"&
+b1�,)��p'�����s
+͗b�$P
����IԳN��O0s��J�)�Q����=�bs`��s)&4|H�@ujٸ��6�0�䃞�{\x�t8����6��I��Wꝏ�����c��k�k.���؈�Kh��������rD�^��fRݐ��.��BG�/<'�:C�Yt��S�&b�]%תz�Ǿ�R������}��� ��¿�&�>���t��uel#��w���T����j�����K�X��Ô���r��.#?�wqMD��D ��r�`(�N���'�;m��ko�����c���I�R�i�Э�/��6��d�`�9�j��D�̢O-�����7�S�Qv]U�iD� ���|5ÕI�IEND�B`�j
+
"
+�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�X�ndE� �D,�������	D$|!19?EF�--�@0^��ڡ�oU�z�}�glK��v��8�8�}�ww��?�,�yg�ĉw#���o;�v~�h�~���)��o;��	�c��Mo^;���/p��/p��mߎ};��@;���v۷cߎ�'Ў�7����طc��	,�د.//G������<:`��U_on�T]��W#'�W���׉�ÎN޵rqq�r��S��:������ms��:9
���-o��y�䅑W̟�������g~�9��f�?>�������׷��?�?������ֱ.�/�S��v��\]_'K���W���z�������:�!�R�&�M~�>K�ۿ�<H�<www �$��wIډos��\o����3\��N �f�>��yr��ӈG	�0��;;���UX���,R8I�H$G����<�g?9(K�}V�T �K�ʘ�z
�˗_eL��n����@A�`���`����#������V5:)��͌EK������((%�I��7����NN
H���[��LB��,��f���@�0D��ّ��%�M�NU����G�~jP�������t�I��!#����Iç��ؗ>� )��T�F4�Y��1Ĥ!%���ن�
�0��9�	�� $�2�d"���<K�����11ϼc��yQR��Q�T��Z�@� �Vb�]"�u�FN�j7�#�P�(\���&q�D��h+���m%����%�ȪS���%�a����Q��L�EI�$�(�8�`�^�D[dE?�6�ʧ�6A��R)bI��1��{I�+��*�!��E����h��-=_xH]��G0\�3�V�f.�cI
��gp����3�]���
+F�|���d����	�x�����(�{
F��[
AL����|�hv��<���]V4�^���pAH�W�H"d��G����P�uż���A�A$��8�O��;��Ó��ý�UqpI^�(P�g�|���{��?G@[1��M�9h�E���샔�6���b��j�m-h�ڡ��"q"&
+b1�,)��p'�����s
+͗b�$P
����IԳN��O0s��J�)�Q����=�bs`��s)&4|H�@ujٸ��6�0�䃞�{\x�t8����6��I��Wꝏ�����c��k�k.���؈�Kh��������rD�^��fRݐ��.��BG�/<'�:C�Yt��S�&b�]%תz�Ǿ�R������}��� ��¿�&�>���t��uel#��w���T����j�����K�X��Ô���r��.#?�wqMD��D ��r�`(�N���'�;m��ko�����c���I�R�i�Э�/��6��d�`�9�j��D�̢O-�����7�S�Qv]U�iD� ���|5ÕI�IEND�B`�j
+
"
+�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�X�ndE� �D,�������	D$|!19?EF�--�@0^��ڡ�oU�z�}�glK��v��8�8�}�ww��?�,�yg�ĉw#���o;�v~�h�~���)��o;��	�c��Mo^;���/p��/p��mߎ};��@;���v۷cߎ�'Ў�7����طc��	,�د.//G������<:`��U_on�T]��W#'�W���׉�ÎN޵rqq�r��S��:������ms��:9
���-o��y�䅑W̟�������g~�9��f�?>�������׷��?�?������ֱ.�/�S��v��\]_'K���W���z�������:�!�R�&�M~�>K�ۿ�<H�<www �$��wIډos��\o����3\��N �f�>��yr��ӈG	�0��;;���UX���,R8I�H$G����<�g?9(K�}V�T �K�ʘ�z
�˗_eL��n����@A�`���`����#������V5:)��͌EK������((%�I��7����NN
H���[��LB��,��f���@�0D��ّ��%�M�NU����G�~jP�������t�I��!#����Iç��ؗ>� )��T�F4�Y��1Ĥ!%���ن�
�0��9�	�� $�2�d"���<K�����11ϼc��yQR��Q�T��Z�@� �Vb�]"�u�FN�j7�#�P�(\���&q�D��h+���m%����%�ȪS���%�a����Q��L�EI�$�(�8�`�^�D[dE?�6�ʧ�6A��R)bI��1��{I�+��*�!��E����h��-=_xH]��G0\�3�V�f.�cI
��gp����3�]���
+F�|���d����	�x�����(�{
F��[
AL����|�hv��<���]V4�^���pAH�W�H"d��G����P�uż���A�A$��8�O��;��Ó��ý�UqpI^�(P�g�|���{��?G@[1��M�9h�E���샔�6���b��j�m-h�ڡ��"q"&
+b1�,)��p'�����s
+͗b�$P
����IԳN��O0s��J�)�Q����=�bs`��s)&4|H�@ujٸ��6�0�䃞�{\x�t8����6��I��Wꝏ�����c��k�k.���؈�Kh��������rD�^��fRݐ��.��BG�/<'�:C�Yt��S�&b�]%תz�Ǿ�R������}��� ��¿�&�>���t��uel#��w���T����j�����K�X��Ô���r��.#?�wqMD��D ��r�`(�N���'�;m��ko�����c���I�R�i�Э�/��6��d�`�9�j��D�̢O-�����7�S�Qv]U�iD� ���|5ÕI�IEND�B`�j
+
"
+@�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Xͮ�D����$W�@�-jT]!$6���'�Ux��a�W@����X���R����$�mn�1ߙ��s��6�l�u�����o��ط�7��4���J�x�䛺��ʷ+��h۾���(�+߮|g�m�.z{��m߶}g�m�.z{ڷm߶}g�m�.z{ڷm߶}g�m�.z{ڷm߶}g��m׾�������
��q�������F�&Z����>o@���9�n����|����|���������ۏ��8�����׵���򗿿�<K����x����P����`����1��݋4�����+��6��G7�E�H���Y2;N�w������OGx�&i��=��f~��ܣ8�:����{!�:Nn��?=��/���
���?>�%�EP���A�q7~v�?�͵�����X��ߣG��R*�*���KB�/��.]r|�x���ePA���"_fK�1X,�%�A�[�'y��7��y��]F��Mje"�%@���lET����X��#r���E�<��<���i��sh�<�=|��wG�9��G�b=<�آ��[j[�|*��ّY��^��B�"q¡�����q6���ϔ?ܘ�G�����8	�3 �BhORn���1�̕=���P0q���a	��Ģ#����/��Φ'��ω�"�γ�Lܳ""Po�`0@�q���+Ec���d�<�
�\�r����Y��^9�|�o
�������hoZd�g0�fx��NOSO��T��0����
"v��Y:�CSL]�Y({��ɘ3��W�V��^�����AX>�@Ap6:û
˚/��p4<�@0?Mg�Y��7,����!�A ֟ձb�%Jn�yi�_Q�5��Ѻ*IM���)�C
+/O^�8ϋe^�t���s���<z�p��0�H�	�l��dX�
+V	:W�\��y�Q��7��>9Ӱ
�=�����>�n��n�t"�ǿ
(���{�˃��Ӵ��FH]��Wް����N�a���liݸ��w
��?z~��"���Ǚ7?��$]�K��"�謿��Xvn���I����2׵�Q޹�%s�Hc�ge�J�7���#��h~���
S8��H�s�_��ܨ������VF�OQ��}n�?��<�L��ݧ��a���X�0�=@������bBi���˃��v}��Hd��W%dV�/F�+�1f9�§�gZ��$�����D�ҹJ�����̝���N�dȡVd!h5bo
�%K��H~�������A�o�s"�7���Hi��(���i�<
+�(	VHLS�m�)��5.�&��K�_��i�
�,�������,������U�d�_���=I&����'��R`u�+�Z�+�s�V�Uh��m��ל5��b�[����g��j�0ب����8�
+��\��F����B��2����a��c���	��5o��݄�3ʓI'e~�I�9VL���8��E�'\�=v�Q�$mﰁu�u�}����&ֵ��0��J��Y8&`�Y���z���<���NF0Yg����JՙJ�\��]��B
+������	������s���6q�D���Z�i>�[�D0<���E���럞N�z��H2b-�
3�?�J	�2s�5?v����z�=/���1�P;��D;�.UE��l��1#����#)�v{�M�ʫi@�yrH��^��d	����`�_ψF��[���R{6�~ɛT)E�J�~��\��(�^u����몔�J��ެ��
u=b/|y>����˛h�	s�D�Ԩـ���G	�Ul��Ni���I��@�Ƨ����~k�&Vx��5���G^��^�"ۃ����y�:a5E�dmF2K��M����Ug���-6�uI�z4K��t�F鉘��Ȗ��dv�^��-��-�>��b.�>��`V��������jR�r<s���ω�D`V��q�Y�ʨ���*�Ì�N#ov�Y
+���tT���i���{�~������ޔ�:x�P�a栭�����
���i��_IEND�B`�j
+
"
+@�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Xͮ�D����$W�@�-jT]!$6���'�Ux��a�W@����X���R����$�mn�1ߙ��s��6�l�u�����o��ط�7��4���J�x�䛺��ʷ+��h۾���(�+߮|g�m�.z{��m߶}g�m�.z{ڷm߶}g�m�.z{ڷm߶}g�m�.z{ڷm߶}g��m׾�������
��q�������F�&Z����>o@���9�n����|����|���������ۏ��8�����׵���򗿿�<K����x����P����`����1��݋4�����+��6��G7�E�H���Y2;N�w������OGx�&i��=��f~��ܣ8�:����{!�:Nn��?=��/���
���?>�%�EP���A�q7~v�?�͵�����X��ߣG��R*�*���KB�/��.]r|�x���ePA���"_fK�1X,�%�A�[�'y��7��y��]F��Mje"�%@���lET����X��#r���E�<��<���i��sh�<�=|��wG�9��G�b=<�آ��[j[�|*��ّY��^��B�"q¡�����q6���ϔ?ܘ�G�����8	�3 �BhORn���1�̕=���P0q���a	��Ģ#����/��Φ'��ω�"�γ�Lܳ""Po�`0@�q���+Ec���d�<�
�\�r����Y��^9�|�o
�������hoZd�g0�fx��NOSO��T��0����
"v��Y:�CSL]�Y({��ɘ3��W�V��^�����AX>�@Ap6:û
˚/��p4<�@0?Mg�Y��7,����!�A ֟ձb�%Jn�yi�_Q�5��Ѻ*IM���)�C
+/O^�8ϋe^�t���s���<z�p��0�H�	�l��dX�
+V	:W�\��y�Q��7��>9Ӱ
�=�����>�n��n�t"�ǿ
(���{�˃��Ӵ��FH]��Wް����N�a���liݸ��w
��?z~��"���Ǚ7?��$]�K��"�謿��Xvn���I����2׵�Q޹�%s�Hc�ge�J�7���#��h~���
S8��H�s�_��ܨ������VF�OQ��}n�?��<�L��ݧ��a���X�0�=@������bBi���˃��v}��Hd��W%dV�/F�+�1f9�§�gZ��$�����D�ҹJ�����̝���N�dȡVd!h5bo
�%K��H~�������A�o�s"�7���Hi��(���i�<
+�(	VHLS�m�)��5.�&��K�_��i�
�,�������,������U�d�_���=I&����'��R`u�+�Z�+�s�V�Uh��m��ל5��b�[����g��j�0ب����8�
+��\��F����B��2����a��c���	��5o��݄�3ʓI'e~�I�9VL���8��E�'\�=v�Q�$mﰁu�u�}����&ֵ��0��J��Y8&`�Y���z���<���NF0Yg����JՙJ�\��]��B
+������	������s���6q�D���Z�i>�[�D0<���E���럞N�z��H2b-�
3�?�J	�2s�5?v����z�=/���1�P;��D;�.UE��l��1#����#)�v{�M�ʫi@�yrH��^��d	����`�_ψF��[���R{6�~ɛT)E�J�~��\��(�^u����몔�J��ެ��
u=b/|y>����˛h�	s�D�Ԩـ���G	�Ul��Ni���I��@�Ƨ����~k�&Vx��5���G^��^�"ۃ����y�:a5E�dmF2K��M����Ug���-6�uI�z4K��t�F鉘��Ȗ��dv�^��-��-�>��b.�>��`V��������jR�r<s���ω�D`V��q�Y�ʨ���*�Ì�N#ov�Y
+���tT���i���{�~������ޔ�:x�P�a栭�����
���i��_IEND�B`�j
+
"
+@�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Xͮ�D����$W�@�-jT]!$6���'�Ux��a�W@����X���R����$�mn�1ߙ��s��6�l�u�����o��ط�7��4���J�x�䛺��ʷ+��h۾���(�+߮|g�m�.z{��m߶}g�m�.z{ڷm߶}g�m�.z{ڷm߶}g�m�.z{ڷm߶}g��m׾�������
��q�������F�&Z����>o@���9�n����|����|���������ۏ��8�����׵���򗿿�<K����x����P����`����1��݋4�����+��6��G7�E�H���Y2;N�w������OGx�&i��=��f~��ܣ8�:����{!�:Nn��?=��/���
���?>�%�EP���A�q7~v�?�͵�����X��ߣG��R*�*���KB�/��.]r|�x���ePA���"_fK�1X,�%�A�[�'y��7��y��]F��Mje"�%@���lET����X��#r���E�<��<���i��sh�<�=|��wG�9��G�b=<�آ��[j[�|*��ّY��^��B�"q¡�����q6���ϔ?ܘ�G�����8	�3 �BhORn���1�̕=���P0q���a	��Ģ#����/��Φ'��ω�"�γ�Lܳ""Po�`0@�q���+Ec���d�<�
�\�r����Y��^9�|�o
�������hoZd�g0�fx��NOSO��T��0����
"v��Y:�CSL]�Y({��ɘ3��W�V��^�����AX>�@Ap6:û
˚/��p4<�@0?Mg�Y��7,����!�A ֟ձb�%Jn�yi�_Q�5��Ѻ*IM���)�C
+/O^�8ϋe^�t���s���<z�p��0�H�	�l��dX�
+V	:W�\��y�Q��7��>9Ӱ
�=�����>�n��n�t"�ǿ
(���{�˃��Ӵ��FH]��Wް����N�a���liݸ��w
��?z~��"���Ǚ7?��$]�K��"�謿��Xvn���I����2׵�Q޹�%s�Hc�ge�J�7���#��h~���
S8��H�s�_��ܨ������VF�OQ��}n�?��<�L��ݧ��a���X�0�=@������bBi���˃��v}��Hd��W%dV�/F�+�1f9�§�gZ��$�����D�ҹJ�����̝���N�dȡVd!h5bo
�%K��H~�������A�o�s"�7���Hi��(���i�<
+�(	VHLS�m�)��5.�&��K�_��i�
�,�������,������U�d�_���=I&����'��R`u�+�Z�+�s�V�Uh��m��ל5��b�[����g��j�0ب����8�
+��\��F����B��2����a��c���	��5o��݄�3ʓI'e~�I�9VL���8��E�'\�=v�Q�$mﰁu�u�}����&ֵ��0��J��Y8&`�Y���z���<���NF0Yg����JՙJ�\��]��B
+������	������s���6q�D���Z�i>�[�D0<���E���럞N�z��H2b-�
3�?�J	�2s�5?v����z�=/���1�P;��D;�.UE��l��1#����#)�v{�M�ʫi@�yrH��^��d	����`�_ψF��[���R{6�~ɛT)E�J�~��\��(�^u����몔�J��ެ��
u=b/|y>����˛h�	s�D�Ԩـ���G	�Ul��Ni���I��@�Ƨ����~k�&Vx��5���G^��^�"ۃ����y�:a5E�dmF2K��M����Ug���-6�uI�z4K��t�F鉘��Ȗ��dv�^��-��-�>��b.�>��`V��������jR�r<s���ω�D`V��q�Y�ʨ���*�Ì�N#ov�Y
+���tT���i���{�~������ޔ�:x�P�a栭�����
���i��_IEND�B`�j
+
"
+@�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Xͮ�D����$W�@�-jT]!$6���'�Ux��a�W@����X���R����$�mn�1ߙ��s��6�l�u�����o��ط�7��4���J�x�䛺��ʷ+��h۾���(�+߮|g�m�.z{��m߶}g�m�.z{ڷm߶}g�m�.z{ڷm߶}g�m�.z{ڷm߶}g��m׾�������
��q�������F�&Z����>o@���9�n����|����|���������ۏ��8�����׵���򗿿�<K����x����P����`����1��݋4�����+��6��G7�E�H���Y2;N�w������OGx�&i��=��f~��ܣ8�:����{!�:Nn��?=��/���
���?>�%�EP���A�q7~v�?�͵�����X��ߣG��R*�*���KB�/��.]r|�x���ePA���"_fK�1X,�%�A�[�'y��7��y��]F��Mje"�%@���lET����X��#r���E�<��<���i��sh�<�=|��wG�9��G�b=<�آ��[j[�|*��ّY��^��B�"q¡�����q6���ϔ?ܘ�G�����8	�3 �BhORn���1�̕=���P0q���a	��Ģ#����/��Φ'��ω�"�γ�Lܳ""Po�`0@�q���+Ec���d�<�
�\�r����Y��^9�|�o
�������hoZd�g0�fx��NOSO��T��0����
"v��Y:�CSL]�Y({��ɘ3��W�V��^�����AX>�@Ap6:û
˚/��p4<�@0?Mg�Y��7,����!�A ֟ձb�%Jn�yi�_Q�5��Ѻ*IM���)�C
+/O^�8ϋe^�t���s���<z�p��0�H�	�l��dX�
+V	:W�\��y�Q��7��>9Ӱ
�=�����>�n��n�t"�ǿ
(���{�˃��Ӵ��FH]��Wް����N�a���liݸ��w
��?z~��"���Ǚ7?��$]�K��"�謿��Xvn���I����2׵�Q޹�%s�Hc�ge�J�7���#��h~���
S8��H�s�_��ܨ������VF�OQ��}n�?��<�L��ݧ��a���X�0�=@������bBi���˃��v}��Hd��W%dV�/F�+�1f9�§�gZ��$�����D�ҹJ�����̝���N�dȡVd!h5bo
�%K��H~�������A�o�s"�7���Hi��(���i�<
+�(	VHLS�m�)��5.�&��K�_��i�
�,�������,������U�d�_���=I&����'��R`u�+�Z�+�s�V�Uh��m��ל5��b�[����g��j�0ب����8�
+��\��F����B��2����a��c���	��5o��݄�3ʓI'e~�I�9VL���8��E�'\�=v�Q�$mﰁu�u�}����&ֵ��0��J��Y8&`�Y���z���<���NF0Yg����JՙJ�\��]��B
+������	������s���6q�D���Z�i>�[�D0<���E���럞N�z��H2b-�
3�?�J	�2s�5?v����z�=/���1�P;��D;�.UE��l��1#����#)�v{�M�ʫi@�yrH��^��d	����`�_ψF��[���R{6�~ɛT)E�J�~��\��(�^u����몔�J��ެ��
u=b/|y>����˛h�	s�D�Ԩـ���G	�Ul��Ni���I��@�Ƨ����~k�&Vx��5���G^��^�"ۃ����y�:a5E�dmF2K��M����Ug���-6�uI�z4K��t�F鉘��Ȗ��dv�^��-��-�>��b.�>��`V��������jR�r<s���ω�D`V��q�Y�ʨ���*�Ì�N#ov�Y
+���tT���i���{�~������ޔ�:x�P�a栭�����
���i��_IEND�B`�j
+
"
+@�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Xͮ�D����$W�@�-jT]!$6���'�Ux��a�W@����X���R����$�mn�1ߙ��s��6�l�u�����o��ط�7��4���J�x�䛺��ʷ+��h۾���(�+߮|g�m�.z{��m߶}g�m�.z{ڷm߶}g�m�.z{ڷm߶}g�m�.z{ڷm߶}g��m׾�������
��q�������F�&Z����>o@���9�n����|����|���������ۏ��8�����׵���򗿿�<K����x����P����`����1��݋4�����+��6��G7�E�H���Y2;N�w������OGx�&i��=��f~��ܣ8�:����{!�:Nn��?=��/���
���?>�%�EP���A�q7~v�?�͵�����X��ߣG��R*�*���KB�/��.]r|�x���ePA���"_fK�1X,�%�A�[�'y��7��y��]F��Mje"�%@���lET����X��#r���E�<��<���i��sh�<�=|��wG�9��G�b=<�آ��[j[�|*��ّY��^��B�"q¡�����q6���ϔ?ܘ�G�����8	�3 �BhORn���1�̕=���P0q���a	��Ģ#����/��Φ'��ω�"�γ�Lܳ""Po�`0@�q���+Ec���d�<�
�\�r����Y��^9�|�o
�������hoZd�g0�fx��NOSO��T��0����
"v��Y:�CSL]�Y({��ɘ3��W�V��^�����AX>�@Ap6:û
˚/��p4<�@0?Mg�Y��7,����!�A ֟ձb�%Jn�yi�_Q�5��Ѻ*IM���)�C
+/O^�8ϋe^�t���s���<z�p��0�H�	�l��dX�
+V	:W�\��y�Q��7��>9Ӱ
�=�����>�n��n�t"�ǿ
(���{�˃��Ӵ��FH]��Wް����N�a���liݸ��w
��?z~��"���Ǚ7?��$]�K��"�謿��Xvn���I����2׵�Q޹�%s�Hc�ge�J�7���#��h~���
S8��H�s�_��ܨ������VF�OQ��}n�?��<�L��ݧ��a���X�0�=@������bBi���˃��v}��Hd��W%dV�/F�+�1f9�§�gZ��$�����D�ҹJ�����̝���N�dȡVd!h5bo
�%K��H~�������A�o�s"�7���Hi��(���i�<
+�(	VHLS�m�)��5.�&��K�_��i�
�,�������,������U�d�_���=I&����'��R`u�+�Z�+�s�V�Uh��m��ל5��b�[����g��j�0ب����8�
+��\��F����B��2����a��c���	��5o��݄�3ʓI'e~�I�9VL���8��E�'\�=v�Q�$mﰁu�u�}����&ֵ��0��J��Y8&`�Y���z���<���NF0Yg����JՙJ�\��]��B
+������	������s���6q�D���Z�i>�[�D0<���E���럞N�z��H2b-�
3�?�J	�2s�5?v����z�=/���1�P;��D;�.UE��l��1#����#)�v{�M�ʫi@�yrH��^��d	����`�_ψF��[���R{6�~ɛT)E�J�~��\��(�^u����몔�J��ެ��
u=b/|y>����˛h�	s�D�Ԩـ���G	�Ul��Ni���I��@�Ƨ����~k�&Vx��5���G^��^�"ۃ����y�:a5E�dmF2K��M����Ug���-6�uI�z4K��t�F鉘��Ȗ��dv�^��-��-�>��b.�>��`V��������jR�r<s���ω�D`V��q�Y�ʨ���*�Ì�N#ov�Y
+���tT���i���{�~������ޔ�:x�P�a栭�����
���i��_IEND�B`�j
+
"
+�?�?�?�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�X�n\E�IF���X�����
k��
;~�=_6Y@/	$�x,#�P�����OUw�nj��-n�"�z�:U�n�뉗_|��b��Gsκ�����v�۝��ڱ��M�۝ow~�h�~�7���ڱo�~�h�~�7���۱o�~�h�~�7���۱o�~�h�~�7���۱o�~�x����__}������◿��k��]^^�-�����]�k!�^��ln�����.�Ƿ��/�G��C�ڷ��JL߄���z/�c��{8�i]G���/�������Ez�z�K_~���G�=��LMQ����?����O�럯7�l^���Tj-�Q~���o��o���v�	m��������ꇫ����Pp�.�S�_r��ψ�����b/4���H�?y��L�2fONO(uq~�*�7R?O�)����$�
+�@V���^$'�呍(ȥl���)�gg|?�S'���>�r׊l�l��
+	`T��[R%�8�t�"�ɐ��6ⶋ�͆��S��҇H9����tJ&�
q�!% i�,9�.����,��D\�Z��P�76�!&�ώ�m-	%bI�^8ڋ1qP&$"�%��l�X
@��S�q	IF���`I o	�_�H�vPsT��$����X
�}�	S~�9�*Q(a�Gia���*$����
)cGM�t5@DZOsXC)m̪Jq4�!<��;v�*�|P[x����Ql�����`�2��MJ�%y1�h�
�[^,A��ú1%O��ވe*5��;��M:��R�$�
�-���@����q�Pm���ݼ�9ҧ�H�au"c���� �,H�p��8j�V������_F�fHJ�)I��-
ϖ@Jd�`�>i�t�<:6�&��/�\R)*���"	��zsq��{j�}0O� ���0J��b��u��b�I
���u��#�H�dy"oh�yDj�pP�>l	8^�9�l�1��c��D"��T�#`�RT�K�	$�
Sci�I��.

iV�
�#T�PX�K#�nߧT)�0{��L�Qo)�F�B@Ks=7�H���3��l��EH�s?/~i�4'����g�m����2͡���1�B��J�dM�{>Ϳ����*�?mL(�-
��88!+$B����͢W�t�ʴ�`H�WGT��q�pP�
ˈF&)�pok�S>5�C
+U@�:P�*y�!�D�����a��)�����C~Vd�#Ym(���2��ӝGv�L��:�r��u��'ݍ^Җ-������y��^����kz�k�u'�z�į##t$�3
+�"�f�Ш7uo3����`V���V�oxY�v(���e�n;j�fq��)�<�C���>+�m�;���C�C춋��wT���7
��i��A�K삟y�ӹ�bw:���c9ԕ9��=��&\��(o{;e��͝�}�|��(�۪Ц���3�3��܅�v��m�v��T��!���Q��s�.�������7�˲K�ڱ���vQ����w{;[�iS�hgSU�YD�"��y����IEND�B`�j
+
"
+�?�?�?�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�X�n\E�IF���X�����
k��
;~�=_6Y@/	$�x,#�P�����OUw�nj��-n�"�z�:U�n�뉗_|��b��Gsκ�����v�۝��ڱ��M�۝ow~�h�~�7���ڱo�~�h�~�7���۱o�~�h�~�7���۱o�~�h�~�7���۱o�~�x����__}������◿��k��]^^�-�����]�k!�^��ln�����.�Ƿ��/�G��C�ڷ��JL߄���z/�c��{8�i]G���/�������Ez�z�K_~���G�=��LMQ����?����O�럯7�l^���Tj-�Q~���o��o���v�	m��������ꇫ����Pp�.�S�_r��ψ�����b/4���H�?y��L�2fONO(uq~�*�7R?O�)����$�
+�@V���^$'�呍(ȥl���)�gg|?�S'���>�r׊l�l��
+	`T��[R%�8�t�"�ɐ��6ⶋ�͆��S��҇H9����tJ&�
q�!% i�,9�.����,��D\�Z��P�76�!&�ώ�m-	%bI�^8ڋ1qP&$"�%��l�X
@��S�q	IF���`I o	�_�H�vPsT��$����X
�}�	S~�9�*Q(a�Gia���*$����
)cGM�t5@DZOsXC)m̪Jq4�!<��;v�*�|P[x����Ql�����`�2��MJ�%y1�h�
�[^,A��ú1%O��ވe*5��;��M:��R�$�
�-���@����q�Pm���ݼ�9ҧ�H�au"c���� �,H�p��8j�V������_F�fHJ�)I��-
ϖ@Jd�`�>i�t�<:6�&��/�\R)*���"	��zsq��{j�}0O� ���0J��b��u��b�I
���u��#�H�dy"oh�yDj�pP�>l	8^�9�l�1��c��D"��T�#`�RT�K�	$�
Sci�I��.

iV�
�#T�PX�K#�nߧT)�0{��L�Qo)�F�B@Ks=7�H���3��l��EH�s?/~i�4'����g�m����2͡���1�B��J�dM�{>Ϳ����*�?mL(�-
��88!+$B����͢W�t�ʴ�`H�WGT��q�pP�
ˈF&)�pok�S>5�C
+U@�:P�*y�!�D�����a��)�����C~Vd�#Ym(���2��ӝGv�L��:�r��u��'ݍ^Җ-������y��^����kz�k�u'�z�į##t$�3
+�"�f�Ш7uo3����`V���V�oxY�v(���e�n;j�fq��)�<�C���>+�m�;���C�C춋��wT���7
��i��A�K삟y�ӹ�bw:���c9ԕ9��=��&\��(o{;e��͝�}�|��(�۪Ц���3�3��܅�v��m�v��T��!���Q��s�.�������7�˲K�ڱ���vQ����w{;[�iS�hgSU�YD�"��y����IEND�B`�j
+
"
+�?�?�?�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�X�n\E�IF���X�����
k��
;~�=_6Y@/	$�x,#�P�����OUw�nj��-n�"�z�:U�n�뉗_|��b��Gsκ�����v�۝��ڱ��M�۝ow~�h�~�7���ڱo�~�h�~�7���۱o�~�h�~�7���۱o�~�h�~�7���۱o�~�x����__}������◿��k��]^^�-�����]�k!�^��ln�����.�Ƿ��/�G��C�ڷ��JL߄���z/�c��{8�i]G���/�������Ez�z�K_~���G�=��LMQ����?����O�럯7�l^���Tj-�Q~���o��o���v�	m��������ꇫ����Pp�.�S�_r��ψ�����b/4���H�?y��L�2fONO(uq~�*�7R?O�)����$�
+�@V���^$'�呍(ȥl���)�gg|?�S'���>�r׊l�l��
+	`T��[R%�8�t�"�ɐ��6ⶋ�͆��S��҇H9����tJ&�
q�!% i�,9�.����,��D\�Z��P�76�!&�ώ�m-	%bI�^8ڋ1qP&$"�%��l�X
@��S�q	IF���`I o	�_�H�vPsT��$����X
�}�	S~�9�*Q(a�Gia���*$����
)cGM�t5@DZOsXC)m̪Jq4�!<��;v�*�|P[x����Ql�����`�2��MJ�%y1�h�
�[^,A��ú1%O��ވe*5��;��M:��R�$�
�-���@����q�Pm���ݼ�9ҧ�H�au"c���� �,H�p��8j�V������_F�fHJ�)I��-
ϖ@Jd�`�>i�t�<:6�&��/�\R)*���"	��zsq��{j�}0O� ���0J��b��u��b�I
���u��#�H�dy"oh�yDj�pP�>l	8^�9�l�1��c��D"��T�#`�RT�K�	$�
Sci�I��.

iV�
�#T�PX�K#�nߧT)�0{��L�Qo)�F�B@Ks=7�H���3��l��EH�s?/~i�4'����g�m����2͡���1�B��J�dM�{>Ϳ����*�?mL(�-
��88!+$B����͢W�t�ʴ�`H�WGT��q�pP�
ˈF&)�pok�S>5�C
+U@�:P�*y�!�D�����a��)�����C~Vd�#Ym(���2��ӝGv�L��:�r��u��'ݍ^Җ-������y��^����kz�k�u'�z�į##t$�3
+�"�f�Ш7uo3����`V���V�oxY�v(���e�n;j�fq��)�<�C���>+�m�;���C�C춋��wT���7
��i��A�K삟y�ӹ�bw:���c9ԕ9��=��&\��(o{;e��͝�}�|��(�۪Ц���3�3��܅�v��m�v��T��!���Q��s�.�������7�˲K�ڱ���vQ����w{;[�iS�hgSU�YD�"��y����IEND�B`�j
+
"
+�?�?�?�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�X�n\E�IF���X�����
k��
;~�=_6Y@/	$�x,#�P�����OUw�nj��-n�"�z�:U�n�뉗_|��b��Gsκ�����v�۝��ڱ��M�۝ow~�h�~�7���ڱo�~�h�~�7���۱o�~�h�~�7���۱o�~�h�~�7���۱o�~�x����__}������◿��k��]^^�-�����]�k!�^��ln�����.�Ƿ��/�G��C�ڷ��JL߄���z/�c��{8�i]G���/�������Ez�z�K_~���G�=��LMQ����?����O�럯7�l^���Tj-�Q~���o��o���v�	m��������ꇫ����Pp�.�S�_r��ψ�����b/4���H�?y��L�2fONO(uq~�*�7R?O�)����$�
+�@V���^$'�呍(ȥl���)�gg|?�S'���>�r׊l�l��
+	`T��[R%�8�t�"�ɐ��6ⶋ�͆��S��҇H9����tJ&�
q�!% i�,9�.����,��D\�Z��P�76�!&�ώ�m-	%bI�^8ڋ1qP&$"�%��l�X
@��S�q	IF���`I o	�_�H�vPsT��$����X
�}�	S~�9�*Q(a�Gia���*$����
)cGM�t5@DZOsXC)m̪Jq4�!<��;v�*�|P[x����Ql�����`�2��MJ�%y1�h�
�[^,A��ú1%O��ވe*5��;��M:��R�$�
�-���@����q�Pm���ݼ�9ҧ�H�au"c���� �,H�p��8j�V������_F�fHJ�)I��-
ϖ@Jd�`�>i�t�<:6�&��/�\R)*���"	��zsq��{j�}0O� ���0J��b��u��b�I
���u��#�H�dy"oh�yDj�pP�>l	8^�9�l�1��c��D"��T�#`�RT�K�	$�
Sci�I��.

iV�
�#T�PX�K#�nߧT)�0{��L�Qo)�F�B@Ks=7�H���3��l��EH�s?/~i�4'����g�m����2͡���1�B��J�dM�{>Ϳ����*�?mL(�-
��88!+$B����͢W�t�ʴ�`H�WGT��q�pP�
ˈF&)�pok�S>5�C
+U@�:P�*y�!�D�����a��)�����C~Vd�#Ym(���2��ӝGv�L��:�r��u��'ݍ^Җ-������y��^����kz�k�u'�z�į##t$�3
+�"�f�Ш7uo3����`V���V�oxY�v(���e�n;j�fq��)�<�C���>+�m�;���C�C춋��wT���7
��i��A�K삟y�ӹ�bw:���c9ԕ9��=��&\��(o{;e��͝�}�|��(�۪Ц���3�3��܅�v��m�v��T��!���Q��s�.�������7�˲K�ڱ���vQ����w{;[�iS�hgSU�YD�"��y����IEND�B`�j
+
"
+�?�?�?�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�X�n\E�IF���X�����
k��
;~�=_6Y@/	$�x,#�P�����OUw�nj��-n�"�z�:U�n�뉗_|��b��Gsκ�����v�۝��ڱ��M�۝ow~�h�~�7���ڱo�~�h�~�7���۱o�~�h�~�7���۱o�~�h�~�7���۱o�~�x����__}������◿��k��]^^�-�����]�k!�^��ln�����.�Ƿ��/�G��C�ڷ��JL߄���z/�c��{8�i]G���/�������Ez�z�K_~���G�=��LMQ����?����O�럯7�l^���Tj-�Q~���o��o���v�	m��������ꇫ����Pp�.�S�_r��ψ�����b/4���H�?y��L�2fONO(uq~�*�7R?O�)����$�
+�@V���^$'�呍(ȥl���)�gg|?�S'���>�r׊l�l��
+	`T��[R%�8�t�"�ɐ��6ⶋ�͆��S��҇H9����tJ&�
q�!% i�,9�.����,��D\�Z��P�76�!&�ώ�m-	%bI�^8ڋ1qP&$"�%��l�X
@��S�q	IF���`I o	�_�H�vPsT��$����X
�}�	S~�9�*Q(a�Gia���*$����
)cGM�t5@DZOsXC)m̪Jq4�!<��;v�*�|P[x����Ql�����`�2��MJ�%y1�h�
�[^,A��ú1%O��ވe*5��;��M:��R�$�
�-���@����q�Pm���ݼ�9ҧ�H�au"c���� �,H�p��8j�V������_F�fHJ�)I��-
ϖ@Jd�`�>i�t�<:6�&��/�\R)*���"	��zsq��{j�}0O� ���0J��b��u��b�I
���u��#�H�dy"oh�yDj�pP�>l	8^�9�l�1��c��D"��T�#`�RT�K�	$�
Sci�I��.

iV�
�#T�PX�K#�nߧT)�0{��L�Qo)�F�B@Ks=7�H���3��l��EH�s?/~i�4'����g�m����2͡���1�B��J�dM�{>Ϳ����*�?mL(�-
��88!+$B����͢W�t�ʴ�`H�WGT��q�pP�
ˈF&)�pok�S>5�C
+U@�:P�*y�!�D�����a��)�����C~Vd�#Ym(���2��ӝGv�L��:�r��u��'ݍ^Җ-������y��^����kz�k�u'�z�į##t$�3
+�"�f�Ш7uo3����`V���V�oxY�v(���e�n;j�fq��)�<�C���>+�m�;���C�C춋��wT���7
��i��A�K삟y�ӹ�bw:���c9ԕ9��=��&\��(o{;e��͝�}�|��(�۪Ц���3�3��܅�v��m�v��T��!���Q��s�.�������7�˲K�ڱ���vQ����w{;[�iS�hgSU�YD�"��y����IEND�B`�j
+
"
+@@�?�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Y1r$7Ժ�"m��WN�������p��p��Y�8�*�A��93������V��M��pV��/������7{m��=����ϝ�;����~���-ϝ�;����~��>/�y�����
+�c��M���<����p�������~�y�w����p��m?��<�;\��������O?���d?��{��.�kV����U�����J===	t���n�Թ��^��;�Q�񼛷������;�Wd�����ߝ�緿���]������D^T����;��^�b�&��}yyq:���ϟ�m�����1r���*
��"�5c���h�g�%��P!A���S׿)���Q���[�/s��m
���ӧ���U.Hg1�B4i@�e��s�1���n�S�u?Pq�[��z�
+�h��8eX$/��t�����$��or���x���Z
+Y��e0������S<#��Mb�������������8(��r���@$z2ul 
��}}Խ.([$��I��('/|��l�
F�!��3�{�J�>����l
+
+}�M�LU�BK:�R��b�f����59M���?<<ʣn�_�F0?T�b���b���d*P�z�S�o��*�O�*9F�A��G��o���̈�J��Ki43�L�5�(�e��Ҹ
u�c��\0\�Ǟ����.��h�7�P)ʎ'��:J�0��2���˻���O�K�z�@C�`��5��8��k�#��؇G]:HO�϶����*�)!�z�+�d�r�w��u����n)`�i�Z;(�jv��a���b�Pj��C��=��,��01g��A&s�Lٶ7D.�E[r4&�p��:���j��;�N���}T���g�=\�C,$�@KC��u�s��<�%��ݙg0�nģ7���(ŷ�A���K�"Ȁe)�؇|^�6)��	��-���Y*�{�/89ϋ��[�=O�m{(����<Y�nu�29���I��$��	R�x���}l�d-��Hb��h���j����Ԑk�:��X�� &ŀ��5`+��;ۓCٍ5�zB��?�TWA$�ů��[L/a>&�<J1s��EQҘ���N�O�t4<=|��`'�T�T1W�H��kr�e`�L���V����U�	%/#�vGg8�50�l:�h-�yNr��T�B8����P s��3v��Z�T�����W�,�s���p�F��o�F����k�Rv�%�v�1!z^����,��/�b!ʅ��'Qә�bE$z�i��a�F07���;]�2]�Ȩ�<�#>f\�Up�q{p8ZLpX�l���%/��E5/#D���]�cD�"����n�N����pԺ��X%'���򖍅�P�������7vjB��XKW���ܔ�`5����(�ml18��nG<z������4���1����u)-�,�QS�=ܔj�b7�'�������OW���0�
+�0o#8�����1���4?	�I�[8�w���cs:!��B���Hf_�*�������<|17�[N��IEND�B`�j
+
"
+@@�?�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Y1r$7Ժ�"m��WN�������p��p��Y�8�*�A��93������V��M��pV��/������7{m��=����ϝ�;����~���-ϝ�;����~��>/�y�����
+�c��M���<����p�������~�y�w����p��m?��<�;\��������O?���d?��{��.�kV����U�����J===	t���n�Թ��^��;�Q�񼛷������;�Wd�����ߝ�緿���]������D^T����;��^�b�&��}yyq:���ϟ�m�����1r���*
��"�5c���h�g�%��P!A���S׿)���Q���[�/s��m
���ӧ���U.Hg1�B4i@�e��s�1���n�S�u?Pq�[��z�
+�h��8eX$/��t�����$��or���x���Z
+Y��e0������S<#��Mb�������������8(��r���@$z2ul 
��}}Խ.([$��I��('/|��l�
F�!��3�{�J�>����l
+
+}�M�LU�BK:�R��b�f����59M���?<<ʣn�_�F0?T�b���b���d*P�z�S�o��*�O�*9F�A��G��o���̈�J��Ki43�L�5�(�e��Ҹ
u�c��\0\�Ǟ����.��h�7�P)ʎ'��:J�0��2���˻���O�K�z�@C�`��5��8��k�#��؇G]:HO�϶����*�)!�z�+�d�r�w��u����n)`�i�Z;(�jv��a���b�Pj��C��=��,��01g��A&s�Lٶ7D.�E[r4&�p��:���j��;�N���}T���g�=\�C,$�@KC��u�s��<�%��ݙg0�nģ7���(ŷ�A���K�"Ȁe)�؇|^�6)��	��-���Y*�{�/89ϋ��[�=O�m{(����<Y�nu�29���I��$��	R�x���}l�d-��Hb��h���j����Ԑk�:��X�� &ŀ��5`+��;ۓCٍ5�zB��?�TWA$�ů��[L/a>&�<J1s��EQҘ���N�O�t4<=|��`'�T�T1W�H��kr�e`�L���V����U�	%/#�vGg8�50�l:�h-�yNr��T�B8����P s��3v��Z�T�����W�,�s���p�F��o�F����k�Rv�%�v�1!z^����,��/�b!ʅ��'Qә�bE$z�i��a�F07���;]�2]�Ȩ�<�#>f\�Up�q{p8ZLpX�l���%/��E5/#D���]�cD�"����n�N����pԺ��X%'���򖍅�P�������7vjB��XKW���ܔ�`5����(�ml18��nG<z������4���1����u)-�,�QS�=ܔj�b7�'�������OW���0�
+�0o#8�����1���4?	�I�[8�w���cs:!��B���Hf_�*�������<|17�[N��IEND�B`�j
+
"
+@@�?�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Y1r$7Ժ�"m��WN�������p��p��Y�8�*�A��93������V��M��pV��/������7{m��=����ϝ�;����~���-ϝ�;����~��>/�y�����
+�c��M���<����p�������~�y�w����p��m?��<�;\��������O?���d?��{��.�kV����U�����J===	t���n�Թ��^��;�Q�񼛷������;�Wd�����ߝ�緿���]������D^T����;��^�b�&��}yyq:���ϟ�m�����1r���*
��"�5c���h�g�%��P!A���S׿)���Q���[�/s��m
���ӧ���U.Hg1�B4i@�e��s�1���n�S�u?Pq�[��z�
+�h��8eX$/��t�����$��or���x���Z
+Y��e0������S<#��Mb�������������8(��r���@$z2ul 
��}}Խ.([$��I��('/|��l�
F�!��3�{�J�>����l
+
+}�M�LU�BK:�R��b�f����59M���?<<ʣn�_�F0?T�b���b���d*P�z�S�o��*�O�*9F�A��G��o���̈�J��Ki43�L�5�(�e��Ҹ
u�c��\0\�Ǟ����.��h�7�P)ʎ'��:J�0��2���˻���O�K�z�@C�`��5��8��k�#��؇G]:HO�϶����*�)!�z�+�d�r�w��u����n)`�i�Z;(�jv��a���b�Pj��C��=��,��01g��A&s�Lٶ7D.�E[r4&�p��:���j��;�N���}T���g�=\�C,$�@KC��u�s��<�%��ݙg0�nģ7���(ŷ�A���K�"Ȁe)�؇|^�6)��	��-���Y*�{�/89ϋ��[�=O�m{(����<Y�nu�29���I��$��	R�x���}l�d-��Hb��h���j����Ԑk�:��X�� &ŀ��5`+��;ۓCٍ5�zB��?�TWA$�ů��[L/a>&�<J1s��EQҘ���N�O�t4<=|��`'�T�T1W�H��kr�e`�L���V����U�	%/#�vGg8�50�l:�h-�yNr��T�B8����P s��3v��Z�T�����W�,�s���p�F��o�F����k�Rv�%�v�1!z^����,��/�b!ʅ��'Qә�bE$z�i��a�F07���;]�2]�Ȩ�<�#>f\�Up�q{p8ZLpX�l���%/��E5/#D���]�cD�"����n�N����pԺ��X%'���򖍅�P�������7vjB��XKW���ܔ�`5����(�ml18��nG<z������4���1����u)-�,�QS�=ܔj�b7�'�������OW���0�
+�0o#8�����1���4?	�I�[8�w���cs:!��B���Hf_�*�������<|17�[N��IEND�B`�j
+
"
+@@�?�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Y1r$7Ժ�"m��WN�������p��p��Y�8�*�A��93������V��M��pV��/������7{m��=����ϝ�;����~���-ϝ�;����~��>/�y�����
+�c��M���<����p�������~�y�w����p��m?��<�;\��������O?���d?��{��.�kV����U�����J===	t���n�Թ��^��;�Q�񼛷������;�Wd�����ߝ�緿���]������D^T����;��^�b�&��}yyq:���ϟ�m�����1r���*
��"�5c���h�g�%��P!A���S׿)���Q���[�/s��m
���ӧ���U.Hg1�B4i@�e��s�1���n�S�u?Pq�[��z�
+�h��8eX$/��t�����$��or���x���Z
+Y��e0������S<#��Mb�������������8(��r���@$z2ul 
��}}Խ.([$��I��('/|��l�
F�!��3�{�J�>����l
+
+}�M�LU�BK:�R��b�f����59M���?<<ʣn�_�F0?T�b���b���d*P�z�S�o��*�O�*9F�A��G��o���̈�J��Ki43�L�5�(�e��Ҹ
u�c��\0\�Ǟ����.��h�7�P)ʎ'��:J�0��2���˻���O�K�z�@C�`��5��8��k�#��؇G]:HO�϶����*�)!�z�+�d�r�w��u����n)`�i�Z;(�jv��a���b�Pj��C��=��,��01g��A&s�Lٶ7D.�E[r4&�p��:���j��;�N���}T���g�=\�C,$�@KC��u�s��<�%��ݙg0�nģ7���(ŷ�A���K�"Ȁe)�؇|^�6)��	��-���Y*�{�/89ϋ��[�=O�m{(����<Y�nu�29���I��$��	R�x���}l�d-��Hb��h���j����Ԑk�:��X�� &ŀ��5`+��;ۓCٍ5�zB��?�TWA$�ů��[L/a>&�<J1s��EQҘ���N�O�t4<=|��`'�T�T1W�H��kr�e`�L���V����U�	%/#�vGg8�50�l:�h-�yNr��T�B8����P s��3v��Z�T�����W�,�s���p�F��o�F����k�Rv�%�v�1!z^����,��/�b!ʅ��'Qә�bE$z�i��a�F07���;]�2]�Ȩ�<�#>f\�Up�q{p8ZLpX�l���%/��E5/#D���]�cD�"����n�N����pԺ��X%'���򖍅�P�������7vjB��XKW���ܔ�`5����(�ml18��nG<z������4���1����u)-�,�QS�=ܔj�b7�'�������OW���0�
+�0o#8�����1���4?	�I�[8�w���cs:!��B���Hf_�*�������<|17�[N��IEND�B`�j
+
"
+@@�?�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Y1r$7Ժ�"m��WN�������p��p��Y�8�*�A��93������V��M��pV��/������7{m��=����ϝ�;����~���-ϝ�;����~��>/�y�����
+�c��M���<����p�������~�y�w����p��m?��<�;\��������O?���d?��{��.�kV����U�����J===	t���n�Թ��^��;�Q�񼛷������;�Wd�����ߝ�緿���]������D^T����;��^�b�&��}yyq:���ϟ�m�����1r���*
��"�5c���h�g�%��P!A���S׿)���Q���[�/s��m
���ӧ���U.Hg1�B4i@�e��s�1���n�S�u?Pq�[��z�
+�h��8eX$/��t�����$��or���x���Z
+Y��e0������S<#��Mb�������������8(��r���@$z2ul 
��}}Խ.([$��I��('/|��l�
F�!��3�{�J�>����l
+
+}�M�LU�BK:�R��b�f����59M���?<<ʣn�_�F0?T�b���b���d*P�z�S�o��*�O�*9F�A��G��o���̈�J��Ki43�L�5�(�e��Ҹ
u�c��\0\�Ǟ����.��h�7�P)ʎ'��:J�0��2���˻���O�K�z�@C�`��5��8��k�#��؇G]:HO�϶����*�)!�z�+�d�r�w��u����n)`�i�Z;(�jv��a���b�Pj��C��=��,��01g��A&s�Lٶ7D.�E[r4&�p��:���j��;�N���}T���g�=\�C,$�@KC��u�s��<�%��ݙg0�nģ7���(ŷ�A���K�"Ȁe)�؇|^�6)��	��-���Y*�{�/89ϋ��[�=O�m{(����<Y�nu�29���I��$��	R�x���}l�d-��Hb��h���j����Ԑk�:��X�� &ŀ��5`+��;ۓCٍ5�zB��?�TWA$�ů��[L/a>&�<J1s��EQҘ���N�O�t4<=|��`'�T�T1W�H��kr�e`�L���V����U�	%/#�vGg8�50�l:�h-�yNr��T�B8����P s��3v��Z�T�����W�,�s���p�F��o�F����k�Rv�%�v�1!z^����,��/�b!ʅ��'Qә�bE$z�i��a�F07���;]�2]�Ȩ�<�#>f\�Up�q{p8ZLpX�l���%/��E5/#D���]�cD�"����n�N����pԺ��X%'���򖍅�P�������7vjB��XKW���ܔ�`5����(�ml18��nG<z������4���1����u)-�,�QS�=ܔj�b7�'�������OW���0�
+�0o#8�����1���4?	�I�[8�w���cs:!��B���Hf_�*�������<|17�[N��IEND�B`�j
+
"
+�?�?@�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Y��U݁�T��d���GP�|!99F������]�
+�}������i%���*�t�9}�1wF��}�ÏWs�|0��y�6�\�~�����p���𦇑۝ow~�h�~�7���ڱo�~�h�~�7���۱o�~�h�~�7���۱o�~�h�~�7���۱o�~n��|����M�˯���<'�fH�=l�H��c��D->}YO�V}J�����ooo�`�c�y����LN�>���3�����ɯ������~�ݫ�����ۿ��i]�	�>���۩>���w��7������食-tB���~��j�v�~�:ᨉ�����~u�Z�Y����I��r��T��D�^�xA�������hwu��ޖ�^ݼ߾�#_l?����lr*1�G�r����v��ۋc*Z(�!�ai�U'��܀.>^����eL.)&^��L��������
+1o1F��b�2wd
�]��Mn m6$��<����&䣫$נ�ł�ZJ�z5�<�5O�emp��,�
��NEPBe��$��P�eh�/��.�����
+�PM����t���ti�)��`$�#E�n�6�e
+���	��
+}��� 
��������r�F���!@Ƞ���ѱ���S���lJ�����0H���E�<�!�Q�rԋ(��
�A��ßժJ�4��Gjfi��e�iPYM>4<�b񊡒.��!<|��d�:��"�a�,���F@e��`#�cO����r,JH���z�ɰ����X�ҋ��9��~��L��*����:���Ԛj��1p�%|�^<(�Fj7��kđ:5�¡uz2�"���"��,L�0`h�R�9�;��z��)&��D�ǖ�gK(9s�v�cg@wB?�6���gq\����Dg
��%���hG?�T2��� �`n0�J{1t��Dd��F
�����	��J���IR -h���H�� u�r��s�I�S\�G��@I�U�G
+����8.Q4&�n�_x���
�\�L{u��q��d��R�~�כ�҆8cfL��S���nT
+`)��k�`fᙇ����^V#��R?M~ )�X��
j��3��#O^���2�1��1�l<Ǖ,��:��|�m�Y�w�P�[�98
��|�}����t�r�G��+3
+E���T8���e�G:���`k\Sh>KCJU��:��,y�!�B��db%BX�,T
+!LAZ�!?I2��$7$_x�Ą��I��-;8�`��%~���ƒ��Fw���yɱu8�$�3�|^�"	_9���_��k8J��y#~��I�1#�������m���0}0�Nwa3�/�$O+����<u��m�t
+��8z_)i�<���U#݌�#�O�k�1�c췋��wR�4~�ƙR6)�d�삟yǟRy8c9���c���1��5%��D�/�mo���s}Gxh#V���lx��d�X�|F��t#�]�l��0�EKghlJ�?���4iNۅm�^x�����&U�
�Ixn:������ږ�Is�e��?��σ���IEND�B`�j
+
"
+�?�?@�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Y��U݁�T��d���GP�|!99F������]�
+�}������i%���*�t�9}�1wF��}�ÏWs�|0��y�6�\�~�����p���𦇑۝ow~�h�~�7���ڱo�~�h�~�7���۱o�~�h�~�7���۱o�~�h�~�7���۱o�~n��|����M�˯���<'�fH�=l�H��c��D->}YO�V}J�����ooo�`�c�y����LN�>���3�����ɯ������~�ݫ�����ۿ��i]�	�>���۩>���w��7������食-tB���~��j�v�~�:ᨉ�����~u�Z�Y����I��r��T��D�^�xA�������hwu��ޖ�^ݼ߾�#_l?����lr*1�G�r����v��ۋc*Z(�!�ai�U'��܀.>^����eL.)&^��L��������
+1o1F��b�2wd
�]��Mn m6$��<����&䣫$נ�ł�ZJ�z5�<�5O�emp��,�
��NEPBe��$��P�eh�/��.�����
+�PM����t���ti�)��`$�#E�n�6�e
+���	��
+}��� 
��������r�F���!@Ƞ���ѱ���S���lJ�����0H���E�<�!�Q�rԋ(��
�A��ßժJ�4��Gjfi��e�iPYM>4<�b񊡒.��!<|��d�:��"�a�,���F@e��`#�cO����r,JH���z�ɰ����X�ҋ��9��~��L��*����:���Ԛj��1p�%|�^<(�Fj7��kđ:5�¡uz2�"���"��,L�0`h�R�9�;��z��)&��D�ǖ�gK(9s�v�cg@wB?�6���gq\����Dg
��%���hG?�T2��� �`n0�J{1t��Dd��F
�����	��J���IR -h���H�� u�r��s�I�S\�G��@I�U�G
+����8.Q4&�n�_x���
�\�L{u��q��d��R�~�כ�҆8cfL��S���nT
+`)��k�`fᙇ����^V#��R?M~ )�X��
j��3��#O^���2�1��1�l<Ǖ,��:��|�m�Y�w�P�[�98
��|�}����t�r�G��+3
+E���T8���e�G:���`k\Sh>KCJU��:��,y�!�B��db%BX�,T
+!LAZ�!?I2��$7$_x�Ą��I��-;8�`��%~���ƒ��Fw���yɱu8�$�3�|^�"	_9���_��k8J��y#~��I�1#�������m���0}0�Nwa3�/�$O+����<u��m�t
+��8z_)i�<���U#݌�#�O�k�1�c췋��wR�4~�ƙR6)�d�삟yǟRy8c9���c���1��5%��D�/�mo���s}Gxh#V���lx��d�X�|F��t#�]�l��0�EKghlJ�?���4iNۅm�^x�����&U�
�Ixn:������ږ�Is�e��?��σ���IEND�B`�j
+
"
+�?�?@�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Y��U݁�T��d���GP�|!99F������]�
+�}������i%���*�t�9}�1wF��}�ÏWs�|0��y�6�\�~�����p���𦇑۝ow~�h�~�7���ڱo�~�h�~�7���۱o�~�h�~�7���۱o�~�h�~�7���۱o�~n��|����M�˯���<'�fH�=l�H��c��D->}YO�V}J�����ooo�`�c�y����LN�>���3�����ɯ������~�ݫ�����ۿ��i]�	�>���۩>���w��7������食-tB���~��j�v�~�:ᨉ�����~u�Z�Y����I��r��T��D�^�xA�������hwu��ޖ�^ݼ߾�#_l?����lr*1�G�r����v��ۋc*Z(�!�ai�U'��܀.>^����eL.)&^��L��������
+1o1F��b�2wd
�]��Mn m6$��<����&䣫$נ�ł�ZJ�z5�<�5O�emp��,�
��NEPBe��$��P�eh�/��.�����
+�PM����t���ti�)��`$�#E�n�6�e
+���	��
+}��� 
��������r�F���!@Ƞ���ѱ���S���lJ�����0H���E�<�!�Q�rԋ(��
�A��ßժJ�4��Gjfi��e�iPYM>4<�b񊡒.��!<|��d�:��"�a�,���F@e��`#�cO����r,JH���z�ɰ����X�ҋ��9��~��L��*����:���Ԛj��1p�%|�^<(�Fj7��kđ:5�¡uz2�"���"��,L�0`h�R�9�;��z��)&��D�ǖ�gK(9s�v�cg@wB?�6���gq\����Dg
��%���hG?�T2��� �`n0�J{1t��Dd��F
�����	��J���IR -h���H�� u�r��s�I�S\�G��@I�U�G
+����8.Q4&�n�_x���
�\�L{u��q��d��R�~�כ�҆8cfL��S���nT
+`)��k�`fᙇ����^V#��R?M~ )�X��
j��3��#O^���2�1��1�l<Ǖ,��:��|�m�Y�w�P�[�98
��|�}����t�r�G��+3
+E���T8���e�G:���`k\Sh>KCJU��:��,y�!�B��db%BX�,T
+!LAZ�!?I2��$7$_x�Ą��I��-;8�`��%~���ƒ��Fw���yɱu8�$�3�|^�"	_9���_��k8J��y#~��I�1#�������m���0}0�Nwa3�/�$O+����<u��m�t
+��8z_)i�<���U#݌�#�O�k�1�c췋��wR�4~�ƙR6)�d�삟yǟRy8c9���c���1��5%��D�/�mo���s}Gxh#V���lx��d�X�|F��t#�]�l��0�EKghlJ�?���4iNۅm�^x�����&U�
�Ixn:������ږ�Is�e��?��σ���IEND�B`�j
+
"
+�?�?@�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Y��U݁�T��d���GP�|!99F������]�
+�}������i%���*�t�9}�1wF��}�ÏWs�|0��y�6�\�~�����p���𦇑۝ow~�h�~�7���ڱo�~�h�~�7���۱o�~�h�~�7���۱o�~�h�~�7���۱o�~n��|����M�˯���<'�fH�=l�H��c��D->}YO�V}J�����ooo�`�c�y����LN�>���3�����ɯ������~�ݫ�����ۿ��i]�	�>���۩>���w��7������食-tB���~��j�v�~�:ᨉ�����~u�Z�Y����I��r��T��D�^�xA�������hwu��ޖ�^ݼ߾�#_l?����lr*1�G�r����v��ۋc*Z(�!�ai�U'��܀.>^����eL.)&^��L��������
+1o1F��b�2wd
�]��Mn m6$��<����&䣫$נ�ł�ZJ�z5�<�5O�emp��,�
��NEPBe��$��P�eh�/��.�����
+�PM����t���ti�)��`$�#E�n�6�e
+���	��
+}��� 
��������r�F���!@Ƞ���ѱ���S���lJ�����0H���E�<�!�Q�rԋ(��
�A��ßժJ�4��Gjfi��e�iPYM>4<�b񊡒.��!<|��d�:��"�a�,���F@e��`#�cO����r,JH���z�ɰ����X�ҋ��9��~��L��*����:���Ԛj��1p�%|�^<(�Fj7��kđ:5�¡uz2�"���"��,L�0`h�R�9�;��z��)&��D�ǖ�gK(9s�v�cg@wB?�6���gq\����Dg
��%���hG?�T2��� �`n0�J{1t��Dd��F
�����	��J���IR -h���H�� u�r��s�I�S\�G��@I�U�G
+����8.Q4&�n�_x���
�\�L{u��q��d��R�~�כ�҆8cfL��S���nT
+`)��k�`fᙇ����^V#��R?M~ )�X��
j��3��#O^���2�1��1�l<Ǖ,��:��|�m�Y�w�P�[�98
��|�}����t�r�G��+3
+E���T8���e�G:���`k\Sh>KCJU��:��,y�!�B��db%BX�,T
+!LAZ�!?I2��$7$_x�Ą��I��-;8�`��%~���ƒ��Fw���yɱu8�$�3�|^�"	_9���_��k8J��y#~��I�1#�������m���0}0�Nwa3�/�$O+����<u��m�t
+��8z_)i�<���U#݌�#�O�k�1�c췋��wR�4~�ƙR6)�d�삟yǟRy8c9���c���1��5%��D�/�mo���s}Gxh#V���lx��d�X�|F��t#�]�l��0�EKghlJ�?���4iNۅm�^x�����&U�
�Ixn:������ږ�Is�e��?��σ���IEND�B`�j
+
"
+�?�?@�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Y��U݁�T��d���GP�|!99F������]�
+�}������i%���*�t�9}�1wF��}�ÏWs�|0��y�6�\�~�����p���𦇑۝ow~�h�~�7���ڱo�~�h�~�7���۱o�~�h�~�7���۱o�~�h�~�7���۱o�~n��|����M�˯���<'�fH�=l�H��c��D->}YO�V}J�����ooo�`�c�y����LN�>���3�����ɯ������~�ݫ�����ۿ��i]�	�>���۩>���w��7������食-tB���~��j�v�~�:ᨉ�����~u�Z�Y����I��r��T��D�^�xA�������hwu��ޖ�^ݼ߾�#_l?����lr*1�G�r����v��ۋc*Z(�!�ai�U'��܀.>^����eL.)&^��L��������
+1o1F��b�2wd
�]��Mn m6$��<����&䣫$נ�ł�ZJ�z5�<�5O�emp��,�
��NEPBe��$��P�eh�/��.�����
+�PM����t���ti�)��`$�#E�n�6�e
+���	��
+}��� 
��������r�F���!@Ƞ���ѱ���S���lJ�����0H���E�<�!�Q�rԋ(��
�A��ßժJ�4��Gjfi��e�iPYM>4<�b񊡒.��!<|��d�:��"�a�,���F@e��`#�cO����r,JH���z�ɰ����X�ҋ��9��~��L��*����:���Ԛj��1p�%|�^<(�Fj7��kđ:5�¡uz2�"���"��,L�0`h�R�9�;��z��)&��D�ǖ�gK(9s�v�cg@wB?�6���gq\����Dg
��%���hG?�T2��� �`n0�J{1t��Dd��F
�����	��J���IR -h���H�� u�r��s�I�S\�G��@I�U�G
+����8.Q4&�n�_x���
�\�L{u��q��d��R�~�כ�҆8cfL��S���nT
+`)��k�`fᙇ����^V#��R?M~ )�X��
j��3��#O^���2�1��1�l<Ǖ,��:��|�m�Y�w�P�[�98
��|�}����t�r�G��+3
+E���T8���e�G:���`k\Sh>KCJU��:��,y�!�B��db%BX�,T
+!LAZ�!?I2��$7$_x�Ą��I��-;8�`��%~���ƒ��Fw���yɱu8�$�3�|^�"	_9���_��k8J��y#~��I�1#�������m���0}0�Nwa3�/�$O+����<u��m�t
+��8z_)i�<���U#݌�#�O�k�1�c췋��wR�4~�ƙR6)�d�삟yǟRy8c9���c���1��5%��D�/�mo���s}Gxh#V���lx��d�X�|F��t#�]�l��0�EKghlJ�?���4iNۅm�^x�����&U�
�Ixn:������ږ�Is�e��?��σ���IEND�B`�j
+
"
+�?�?�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�XˎUuG��h2�`�0�	�D>!{����
�Ċ�%!�Nƚl� ��v�=Uuo����3�o/��q�T��붓����^-�z�Tᬻ�_��oO�=�n��>�$�=�����v������v���v����۾�v���v����۾�v���v����۾�����ꊴ������ߟi��m�!=��ɦF4�O;�=D�]5�d��P����)��.//��.[wj�����(���Z��Q~j��J��;���/�x��g~��������|%Ms�c��|���_����//������7�n�o~�l~����i�;��#������y������ʩ�q���<� ����B7�����E��߾�U�=;;�V���R~��?~|N雛W4��م�JK!+Su��P[�>zċ;?g	9�~�pmq����Y; ����Q�ܑ5�vW�rS�L��n��y����z�,&�S}��Z[)X�:��tG�K@��Ũ*Ra㪨��	(�'��hN������O?%�i���h�;G�K(X�FV�paX5����<�\o�M���E��n�$,X��%@�qtcΈ�2A"1�,���v�5T��zԈ�H2z.����
���M �d=`5G�QO"I)��hV���ǯԲJ����Gia���d��|
e6`��xt�����J}�9�:G�N1�|R[x����Qld�����&�PjD�Up� �b	Jo6�)y�{�,�J����:�I�g���Pǀ����M z�I�8i�v���@G�Uk:͠N�X䳃���
Rn�00�Z��9�;�B�@�f��/QԱ%@��H����g
�NB�G�����A�K
+!E��Y$�<[Bonv�cO-ÅyB�\�s�(Uۋ�����-'5yf���\����c�P���G��
�A	��%������Q����Z��J�$"�;�:�G����,� H7�^t����+*��NG�N����F��~x,msp��4�„Ȕq��B@Ks�}h��|�A�'8�W����S�cA�H3���9�A��̣���+t��=�Lsh��s����R<Y��=���c`��J�M�'��G�	�����)6����z��뺕i��>���4)7�J��H��LR0���~NE`��,����u�U��C@���2f�S
+Aq9��Pd�#�6V_x�b��DŽ�t��]:H��[��R~���LǓ�F/i˖���Cyg>�e/��bš>��x����-����$�gFAU�,�hԛ���vwb�`�����^���}2�,����Na؈���¨�cO?rB��m_�(8j`�;T:ļ]�ϼ�:D���3�mh�b�س��;�����鸧;��PW�(R���p-R����픥>7wN�m��
.�o�R�"���Ψ�72�[����l��S����g�'i"�i����/�66:�d�(˜d�ce#
�vQ���{�U7�����Y��"��]�>���IEND�B`�j
+
"
+�?�?�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�XˎUuG��h2�`�0�	�D>!{����
�Ċ�%!�Nƚl� ��v�=Uuo����3�o/��q�T��붓����^-�z�Tᬻ�_��oO�=�n��>�$�=�����v������v���v����۾�v���v����۾�v���v����۾�����ꊴ������ߟi��m�!=��ɦF4�O;�=D�]5�d��P����)��.//��.[wj�����(���Z��Q~j��J��;���/�x��g~��������|%Ms�c��|���_����//������7�n�o~�l~����i�;��#������y������ʩ�q���<� ����B7�����E��߾�U�=;;�V���R~��?~|N雛W4��م�JK!+Su��P[�>zċ;?g	9�~�pmq����Y; ����Q�ܑ5�vW�rS�L��n��y����z�,&�S}��Z[)X�:��tG�K@��Ũ*Ra㪨��	(�'��hN������O?%�i���h�;G�K(X�FV�paX5����<�\o�M���E��n�$,X��%@�qtcΈ�2A"1�,���v�5T��zԈ�H2z.����
���M �d=`5G�QO"I)��hV���ǯԲJ����Gia���d��|
e6`��xt�����J}�9�:G�N1�|R[x����Qld�����&�PjD�Up� �b	Jo6�)y�{�,�J����:�I�g���Pǀ����M z�I�8i�v���@G�Uk:͠N�X䳃���
Rn�00�Z��9�;�B�@�f��/QԱ%@��H����g
�NB�G�����A�K
+!E��Y$�<[Bonv�cO-ÅyB�\�s�(Uۋ�����-'5yf���\����c�P���G��
�A	��%������Q����Z��J�$"�;�:�G����,� H7�^t����+*��NG�N����F��~x,msp��4�„Ȕq��B@Ks�}h��|�A�'8�W����S�cA�H3���9�A��̣���+t��=�Lsh��s����R<Y��=���c`��J�M�'��G�	�����)6����z��뺕i��>���4)7�J��H��LR0���~NE`��,����u�U��C@���2f�S
+Aq9��Pd�#�6V_x�b��DŽ�t��]:H��[��R~���LǓ�F/i˖���Cyg>�e/��bš>��x����-����$�gFAU�,�hԛ���vwb�`�����^���}2�,����Na؈���¨�cO?rB��m_�(8j`�;T:ļ]�ϼ�:D���3�mh�b�س��;�����鸧;��PW�(R���p-R����픥>7wN�m��
.�o�R�"���Ψ�72�[����l��S����g�'i"�i����/�66:�d�(˜d�ce#
�vQ���{�U7�����Y��"��]�>���IEND�B`�j
+
"
+�?�?�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�XˎUuG��h2�`�0�	�D>!{����
�Ċ�%!�Nƚl� ��v�=Uuo����3�o/��q�T��붓����^-�z�Tᬻ�_��oO�=�n��>�$�=�����v������v���v����۾�v���v����۾�v���v����۾�����ꊴ������ߟi��m�!=��ɦF4�O;�=D�]5�d��P����)��.//��.[wj�����(���Z��Q~j��J��;���/�x��g~��������|%Ms�c��|���_����//������7�n�o~�l~����i�;��#������y������ʩ�q���<� ����B7�����E��߾�U�=;;�V���R~��?~|N雛W4��م�JK!+Su��P[�>zċ;?g	9�~�pmq����Y; ����Q�ܑ5�vW�rS�L��n��y����z�,&�S}��Z[)X�:��tG�K@��Ũ*Ra㪨��	(�'��hN������O?%�i���h�;G�K(X�FV�paX5����<�\o�M���E��n�$,X��%@�qtcΈ�2A"1�,���v�5T��zԈ�H2z.����
���M �d=`5G�QO"I)��hV���ǯԲJ����Gia���d��|
e6`��xt�����J}�9�:G�N1�|R[x����Qld�����&�PjD�Up� �b	Jo6�)y�{�,�J����:�I�g���Pǀ����M z�I�8i�v���@G�Uk:͠N�X䳃���
Rn�00�Z��9�;�B�@�f��/QԱ%@��H����g
�NB�G�����A�K
+!E��Y$�<[Bonv�cO-ÅyB�\�s�(Uۋ�����-'5yf���\����c�P���G��
�A	��%������Q����Z��J�$"�;�:�G����,� H7�^t����+*��NG�N����F��~x,msp��4�„Ȕq��B@Ks�}h��|�A�'8�W����S�cA�H3���9�A��̣���+t��=�Lsh��s����R<Y��=���c`��J�M�'��G�	�����)6����z��뺕i��>���4)7�J��H��LR0���~NE`��,����u�U��C@���2f�S
+Aq9��Pd�#�6V_x�b��DŽ�t��]:H��[��R~���LǓ�F/i˖���Cyg>�e/��bš>��x����-����$�gFAU�,�hԛ���vwb�`�����^���}2�,����Na؈���¨�cO?rB��m_�(8j`�;T:ļ]�ϼ�:D���3�mh�b�س��;�����鸧;��PW�(R���p-R����픥>7wN�m��
.�o�R�"���Ψ�72�[����l��S����g�'i"�i����/�66:�d�(˜d�ce#
�vQ���{�U7�����Y��"��]�>���IEND�B`�j
+
"
+�?�?�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�XˎUuG��h2�`�0�	�D>!{����
�Ċ�%!�Nƚl� ��v�=Uuo����3�o/��q�T��붓����^-�z�Tᬻ�_��oO�=�n��>�$�=�����v������v���v����۾�v���v����۾�v���v����۾�����ꊴ������ߟi��m�!=��ɦF4�O;�=D�]5�d��P����)��.//��.[wj�����(���Z��Q~j��J��;���/�x��g~��������|%Ms�c��|���_����//������7�n�o~�l~����i�;��#������y������ʩ�q���<� ����B7�����E��߾�U�=;;�V���R~��?~|N雛W4��م�JK!+Su��P[�>zċ;?g	9�~�pmq����Y; ����Q�ܑ5�vW�rS�L��n��y����z�,&�S}��Z[)X�:��tG�K@��Ũ*Ra㪨��	(�'��hN������O?%�i���h�;G�K(X�FV�paX5����<�\o�M���E��n�$,X��%@�qtcΈ�2A"1�,���v�5T��zԈ�H2z.����
���M �d=`5G�QO"I)��hV���ǯԲJ����Gia���d��|
e6`��xt�����J}�9�:G�N1�|R[x����Qld�����&�PjD�Up� �b	Jo6�)y�{�,�J����:�I�g���Pǀ����M z�I�8i�v���@G�Uk:͠N�X䳃���
Rn�00�Z��9�;�B�@�f��/QԱ%@��H����g
�NB�G�����A�K
+!E��Y$�<[Bonv�cO-ÅyB�\�s�(Uۋ�����-'5yf���\����c�P���G��
�A	��%������Q����Z��J�$"�;�:�G����,� H7�^t����+*��NG�N����F��~x,msp��4�„Ȕq��B@Ks�}h��|�A�'8�W����S�cA�H3���9�A��̣���+t��=�Lsh��s����R<Y��=���c`��J�M�'��G�	�����)6����z��뺕i��>���4)7�J��H��LR0���~NE`��,����u�U��C@���2f�S
+Aq9��Pd�#�6V_x�b��DŽ�t��]:H��[��R~���LǓ�F/i˖���Cyg>�e/��bš>��x����-����$�gFAU�,�hԛ���vwb�`�����^���}2�,����Na؈���¨�cO?rB��m_�(8j`�;T:ļ]�ϼ�:D���3�mh�b�س��;�����鸧;��PW�(R���p-R����픥>7wN�m��
.�o�R�"���Ψ�72�[����l��S����g�'i"�i����/�66:�d�(˜d�ce#
�vQ���{�U7�����Y��"��]�>���IEND�B`�j
+
"
+�?�?�?�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�XˎUuG��h2�`�0�	�D>!{����
�Ċ�%!�Nƚl� ��v�=Uuo����3�o/��q�T��붓����^-�z�Tᬻ�_��oO�=�n��>�$�=�����v������v���v����۾�v���v����۾�v���v����۾�����ꊴ������ߟi��m�!=��ɦF4�O;�=D�]5�d��P����)��.//��.[wj�����(���Z��Q~j��J��;���/�x��g~��������|%Ms�c��|���_����//������7�n�o~�l~����i�;��#������y������ʩ�q���<� ����B7�����E��߾�U�=;;�V���R~��?~|N雛W4��م�JK!+Su��P[�>zċ;?g	9�~�pmq����Y; ����Q�ܑ5�vW�rS�L��n��y����z�,&�S}��Z[)X�:��tG�K@��Ũ*Ra㪨��	(�'��hN������O?%�i���h�;G�K(X�FV�paX5����<�\o�M���E��n�$,X��%@�qtcΈ�2A"1�,���v�5T��zԈ�H2z.����
���M �d=`5G�QO"I)��hV���ǯԲJ����Gia���d��|
e6`��xt�����J}�9�:G�N1�|R[x����Qld�����&�PjD�Up� �b	Jo6�)y�{�,�J����:�I�g���Pǀ����M z�I�8i�v���@G�Uk:͠N�X䳃���
Rn�00�Z��9�;�B�@�f��/QԱ%@��H����g
�NB�G�����A�K
+!E��Y$�<[Bonv�cO-ÅyB�\�s�(Uۋ�����-'5yf���\����c�P���G��
�A	��%������Q����Z��J�$"�;�:�G����,� H7�^t����+*��NG�N����F��~x,msp��4�„Ȕq��B@Ks�}h��|�A�'8�W����S�cA�H3���9�A��̣���+t��=�Lsh��s����R<Y��=���c`��J�M�'��G�	�����)6����z��뺕i��>���4)7�J��H��LR0���~NE`��,����u�U��C@���2f�S
+Aq9��Pd�#�6V_x�b��DŽ�t��]:H��[��R~���LǓ�F/i˖���Cyg>�e/��bš>��x����-����$�gFAU�,�hԛ���vwb�`�����^���}2�,����Na؈���¨�cO?rB��m_�(8j`�;T:ļ]�ϼ�:D���3�mh�b�س��;�����鸧;��PW�(R���p-R����픥>7wN�m��
.�o�R�"���Ψ�72�[����l��S����g�'i"�i����/�66:�d�(˜d�ce#
�vQ���{�U7�����Y��"��]�>���IEND�B`�j
+
"
+@�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Z�nE�I��bG�+BA
)�&���y�#$x�9'`��H�8ή����P�S�U��3;��ǑfF#w�|����q���Ά�\j�X�X�Pw��q�؁q�����q�ǝ`Ʊএ�8������p���~�q�؁q����m?��8���8������~�c�?�'��I\��������*	қ����M�Z$k�TPI�2Z�kQ����^���#�_�n�^q��F?�}�������u6W�����/fEq
+�l�f�w�WU�{ˬ���'��~�+Z]�ݟ�}����W�z.����p���b1_,J~���>r������_7��Ԛk���̳��/���y{*�W���LJ�/�O�z��"+��r^��eyo��9���̿����/_L������ӽϋ��`������.<����ѝ;�c�s�`?�r�������G�TQΦg���WR�9V�.�e���������8J���3��g2��
+�k6�Nn���Z���6�Fq�j{{�;�v�������
+�{���'����Ln>ŝ����}��CQQ�9�-1���̀�G�� �R�����g�I��^�wߜ����jQ-���m���,��/N��www�ъ�����Z�
+ 8������<���-,�G��������P[��!ѐC��m�(�s�l�A�Z�ٙ'����s^;�lD{rt
+�l>�y<���+���~�n��h����ZPB	���h���X��i'�ݸ�^}�a��7n�_�Ώ�7�a��ڽFP�hd$�HLX1!������f�@сs����ӗ7�߄�O�9�x�ލ=���f��Rl���PFFL-E@���X���|2X'(=+��ك��W=�hi����(F�4K�ӈ�����)�L�*
#�����l�%~��*~إՈ//E�����Q|Д$LwR�������M���:4!�i)ђ�W��XF��,E@�E(^,��Ym�}]1��%���*��N6��(��o��b�ġR����w&���d�j�35X�Q[���W`B҂
+C�v����~$^�Y��J��)�;�(x�#[`k{]��C���iV�4�:���+�[�H!z��[��KaLI>,�U�La��ڄQp_n�����|�p�$��2qBeW���0��M_Ď(Lu��	\/K��A�KYe,9��ҁ$Rk��2ӵ#1�b���M[�����C��UX�� J-5HG�Q�w��l=��C�Kv�E:$�["�b��v 2q�Ł�j�tFJ�c�|h�xb�X�:j���j?�b���#����*��Z�(PT������������I+F�ߴ����	°ځ�tB9����y����&Z��I�ʅK#A��Sa�V:�(U�`(�H ^�Z�����ʀ�v��Xх'v԰v6�wS��HN$U�/��{$�u��/p�!�uBZ�c�F�j�����z�O<��t�z�6���a`��3�I�gL���^�_��w��Ɖ�v���3"W���r�D:���8�)��a/�����8B�vƦ�It��):b�u� Uu��}��}`�Z�o�n�.U�j�ϼqu��/�@��6#�F��x�
�K�v���aL3�qR]�I��„p�w���Y�������#�_c#�ⵕ	|���daG{��M#\�Z���eUZ�e���j��z�S�ϱZ֚\g`��U�Ɉ��8ӋR����:���V]!d�(N#��?��˙$�%JIEND�B`�j
+
"
+@�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Z�nE�I��bG�+BA
)�&���y�#$x�9'`��H�8ή����P�S�U��3;��ǑfF#w�|����q���Ά�\j�X�X�Pw��q�؁q�����q�ǝ`Ʊএ�8������p���~�q�؁q����m?��8���8������~�c�?�'��I\��������*	қ����M�Z$k�TPI�2Z�kQ����^���#�_�n�^q��F?�}�������u6W�����/fEq
+�l�f�w�WU�{ˬ���'��~�+Z]�ݟ�}����W�z.����p���b1_,J~���>r������_7��Ԛk���̳��/���y{*�W���LJ�/�O�z��"+��r^��eyo��9���̿����/_L������ӽϋ��`������.<����ѝ;�c�s�`?�r�������G�TQΦg���WR�9V�.�e���������8J���3��g2��
+�k6�Nn���Z���6�Fq�j{{�;�v�������
+�{���'����Ln>ŝ����}��CQQ�9�-1���̀�G�� �R�����g�I��^�wߜ����jQ-���m���,��/N��www�ъ�����Z�
+ 8������<���-,�G��������P[��!ѐC��m�(�s�l�A�Z�ٙ'����s^;�lD{rt
+�l>�y<���+���~�n��h����ZPB	���h���X��i'�ݸ�^}�a��7n�_�Ώ�7�a��ڽFP�hd$�HLX1!������f�@сs����ӗ7�߄�O�9�x�ލ=���f��Rl���PFFL-E@���X���|2X'(=+��ك��W=�hi����(F�4K�ӈ�����)�L�*
#�����l�%~��*~إՈ//E�����Q|Д$LwR�������M���:4!�i)ђ�W��XF��,E@�E(^,��Ym�}]1��%���*��N6��(��o��b�ġR����w&���d�j�35X�Q[���W`B҂
+C�v����~$^�Y��J��)�;�(x�#[`k{]��C���iV�4�:���+�[�H!z��[��KaLI>,�U�La��ڄQp_n�����|�p�$��2qBeW���0��M_Ď(Lu��	\/K��A�KYe,9��ҁ$Rk��2ӵ#1�b���M[�����C��UX�� J-5HG�Q�w��l=��C�Kv�E:$�["�b��v 2q�Ł�j�tFJ�c�|h�xb�X�:j���j?�b���#����*��Z�(PT������������I+F�ߴ����	°ځ�tB9����y����&Z��I�ʅK#A��Sa�V:�(U�`(�H ^�Z�����ʀ�v��Xх'v԰v6�wS��HN$U�/��{$�u��/p�!�uBZ�c�F�j�����z�O<��t�z�6���a`��3�I�gL���^�_��w��Ɖ�v���3"W���r�D:���8�)��a/�����8B�vƦ�It��):b�u� Uu��}��}`�Z�o�n�.U�j�ϼqu��/�@��6#�F��x�
�K�v���aL3�qR]�I��„p�w���Y�������#�_c#�ⵕ	|���daG{��M#\�Z���eUZ�e���j��z�S�ϱZ֚\g`��U�Ɉ��8ӋR����:���V]!d�(N#��?��˙$�%JIEND�B`�j
+
"
+@�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Z�nE�I��bG�+BA
)�&���y�#$x�9'`��H�8ή����P�S�U��3;��ǑfF#w�|����q���Ά�\j�X�X�Pw��q�؁q�����q�ǝ`Ʊএ�8������p���~�q�؁q����m?��8���8������~�c�?�'��I\��������*	қ����M�Z$k�TPI�2Z�kQ����^���#�_�n�^q��F?�}�������u6W�����/fEq
+�l�f�w�WU�{ˬ���'��~�+Z]�ݟ�}����W�z.����p���b1_,J~���>r������_7��Ԛk���̳��/���y{*�W���LJ�/�O�z��"+��r^��eyo��9���̿����/_L������ӽϋ��`������.<����ѝ;�c�s�`?�r�������G�TQΦg���WR�9V�.�e���������8J���3��g2��
+�k6�Nn���Z���6�Fq�j{{�;�v�������
+�{���'����Ln>ŝ����}��CQQ�9�-1���̀�G�� �R�����g�I��^�wߜ����jQ-���m���,��/N��www�ъ�����Z�
+ 8������<���-,�G��������P[��!ѐC��m�(�s�l�A�Z�ٙ'����s^;�lD{rt
+�l>�y<���+���~�n��h����ZPB	���h���X��i'�ݸ�^}�a��7n�_�Ώ�7�a��ڽFP�hd$�HLX1!������f�@сs����ӗ7�߄�O�9�x�ލ=���f��Rl���PFFL-E@���X���|2X'(=+��ك��W=�hi����(F�4K�ӈ�����)�L�*
#�����l�%~��*~إՈ//E�����Q|Д$LwR�������M���:4!�i)ђ�W��XF��,E@�E(^,��Ym�}]1��%���*��N6��(��o��b�ġR����w&���d�j�35X�Q[���W`B҂
+C�v����~$^�Y��J��)�;�(x�#[`k{]��C���iV�4�:���+�[�H!z��[��KaLI>,�U�La��ڄQp_n�����|�p�$��2qBeW���0��M_Ď(Lu��	\/K��A�KYe,9��ҁ$Rk��2ӵ#1�b���M[�����C��UX�� J-5HG�Q�w��l=��C�Kv�E:$�["�b��v 2q�Ł�j�tFJ�c�|h�xb�X�:j���j?�b���#����*��Z�(PT������������I+F�ߴ����	°ځ�tB9����y����&Z��I�ʅK#A��Sa�V:�(U�`(�H ^�Z�����ʀ�v��Xх'v԰v6�wS��HN$U�/��{$�u��/p�!�uBZ�c�F�j�����z�O<��t�z�6���a`��3�I�gL���^�_��w��Ɖ�v���3"W���r�D:���8�)��a/�����8B�vƦ�It��):b�u� Uu��}��}`�Z�o�n�.U�j�ϼqu��/�@��6#�F��x�
�K�v���aL3�qR]�I��„p�w���Y�������#�_c#�ⵕ	|���daG{��M#\�Z���eUZ�e���j��z�S�ϱZ֚\g`��U�Ɉ��8ӋR����:���V]!d�(N#��?��˙$�%JIEND�B`�j
+
"
+@�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Z�nE�I��bG�+BA
)�&���y�#$x�9'`��H�8ή����P�S�U��3;��ǑfF#w�|����q���Ά�\j�X�X�Pw��q�؁q�����q�ǝ`Ʊএ�8������p���~�q�؁q����m?��8���8������~�c�?�'��I\��������*	қ����M�Z$k�TPI�2Z�kQ����^���#�_�n�^q��F?�}�������u6W�����/fEq
+�l�f�w�WU�{ˬ���'��~�+Z]�ݟ�}����W�z.����p���b1_,J~���>r������_7��Ԛk���̳��/���y{*�W���LJ�/�O�z��"+��r^��eyo��9���̿����/_L������ӽϋ��`������.<����ѝ;�c�s�`?�r�������G�TQΦg���WR�9V�.�e���������8J���3��g2��
+�k6�Nn���Z���6�Fq�j{{�;�v�������
+�{���'����Ln>ŝ����}��CQQ�9�-1���̀�G�� �R�����g�I��^�wߜ����jQ-���m���,��/N��www�ъ�����Z�
+ 8������<���-,�G��������P[��!ѐC��m�(�s�l�A�Z�ٙ'����s^;�lD{rt
+�l>�y<���+���~�n��h����ZPB	���h���X��i'�ݸ�^}�a��7n�_�Ώ�7�a��ڽFP�hd$�HLX1!������f�@сs����ӗ7�߄�O�9�x�ލ=���f��Rl���PFFL-E@���X���|2X'(=+��ك��W=�hi����(F�4K�ӈ�����)�L�*
#�����l�%~��*~إՈ//E�����Q|Д$LwR�������M���:4!�i)ђ�W��XF��,E@�E(^,��Ym�}]1��%���*��N6��(��o��b�ġR����w&���d�j�35X�Q[���W`B҂
+C�v����~$^�Y��J��)�;�(x�#[`k{]��C���iV�4�:���+�[�H!z��[��KaLI>,�U�La��ڄQp_n�����|�p�$��2qBeW���0��M_Ď(Lu��	\/K��A�KYe,9��ҁ$Rk��2ӵ#1�b���M[�����C��UX�� J-5HG�Q�w��l=��C�Kv�E:$�["�b��v 2q�Ł�j�tFJ�c�|h�xb�X�:j���j?�b���#����*��Z�(PT������������I+F�ߴ����	°ځ�tB9����y����&Z��I�ʅK#A��Sa�V:�(U�`(�H ^�Z�����ʀ�v��Xх'v԰v6�wS��HN$U�/��{$�u��/p�!�uBZ�c�F�j�����z�O<��t�z�6���a`��3�I�gL���^�_��w��Ɖ�v���3"W���r�D:���8�)��a/�����8B�vƦ�It��):b�u� Uu��}��}`�Z�o�n�.U�j�ϼqu��/�@��6#�F��x�
�K�v���aL3�qR]�I��„p�w���Y�������#�_c#�ⵕ	|���daG{��M#\�Z���eUZ�e���j��z�S�ϱZ֚\g`��U�Ɉ��8ӋR����:���V]!d�(N#��?��˙$�%JIEND�B`�j
+
"
+@�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�Z�nE�I��bG�+BA
)�&���y�#$x�9'`��H�8ή����P�S�U��3;��ǑfF#w�|����q���Ά�\j�X�X�Pw��q�؁q�����q�ǝ`Ʊএ�8������p���~�q�؁q����m?��8���8������~�c�?�'��I\��������*	қ����M�Z$k�TPI�2Z�kQ����^���#�_�n�^q��F?�}�������u6W�����/fEq
+�l�f�w�WU�{ˬ���'��~�+Z]�ݟ�}����W�z.����p���b1_,J~���>r������_7��Ԛk���̳��/���y{*�W���LJ�/�O�z��"+��r^��eyo��9���̿����/_L������ӽϋ��`������.<����ѝ;�c�s�`?�r�������G�TQΦg���WR�9V�.�e���������8J���3��g2��
+�k6�Nn���Z���6�Fq�j{{�;�v�������
+�{���'����Ln>ŝ����}��CQQ�9�-1���̀�G�� �R�����g�I��^�wߜ����jQ-���m���,��/N��www�ъ�����Z�
+ 8������<���-,�G��������P[��!ѐC��m�(�s�l�A�Z�ٙ'����s^;�lD{rt
+�l>�y<���+���~�n��h����ZPB	���h���X��i'�ݸ�^}�a��7n�_�Ώ�7�a��ڽFP�hd$�HLX1!������f�@сs����ӗ7�߄�O�9�x�ލ=���f��Rl���PFFL-E@���X���|2X'(=+��ك��W=�hi����(F�4K�ӈ�����)�L�*
#�����l�%~��*~إՈ//E�����Q|Д$LwR�������M���:4!�i)ђ�W��XF��,E@�E(^,��Ym�}]1��%���*��N6��(��o��b�ġR����w&���d�j�35X�Q[���W`B҂
+C�v����~$^�Y��J��)�;�(x�#[`k{]��C���iV�4�:���+�[�H!z��[��KaLI>,�U�La��ڄQp_n�����|�p�$��2qBeW���0��M_Ď(Lu��	\/K��A�KYe,9��ҁ$Rk��2ӵ#1�b���M[�����C��UX�� J-5HG�Q�w��l=��C�Kv�E:$�["�b��v 2q�Ł�j�tFJ�c�|h�xb�X�:j���j?�b���#����*��Z�(PT������������I+F�ߴ����	°ځ�tB9����y����&Z��I�ʅK#A��Sa�V:�(U�`(�H ^�Z�����ʀ�v��Xх'v԰v6�wS��HN$U�/��{$�u��/p�!�uBZ�c�F�j�����z�O<��t�z�6���a`��3�I�gL���^�_��w��Ɖ�v���3"W���r�D:���8�)��a/�����8B�vƦ�It��):b�u� Uu��}��}`�Z�o�n�.U�j�ϼqu��/�@��6#�F��x�
�K�v���aL3�qR]�I��„p�w���Y�������#�_c#�ⵕ	|���daG{��M#\�Z���eUZ�e���j��z�S�ϱZ֚\g`��U�Ɉ��8ӋR����:���V]!d�(N#��?��˙$�%JIEND�B`�j
+
"
+@�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�ZK�E�%��ټ �\,ʕK�q��=?q������ʕ�P�8�1+���;�;TuWuU�􌽎c��L��z|_���=�d�ٗ_e]���j��w�|WW�_�~�;8�~�wp�]����+��	�۾���x����}'�o�.z��۾���@��;���i�o�~�wp�������}����3����{<zx���7��$|�ɭ&���-͜����_?�c���eY?���-7�T����v
��m��/�8yx2=<�}���y�߻}�ȊrR��tyZU���j2�\�.o��p�?�e?�}����lA�b�,������<-�C��>}u�Q�T�f�yV��:���r>E
?�f�,����&��t���W>}t����QSA�*ۍo�~����q	{������q����?<��Gw���tAYg�}&A(g���Y�Aͬ�W͎<�91�3:�S^36j6&�
+�ް�a��ɬ�Y<*�m8���<F�R�lG͇
+�}�0?��������ɿ�E�\�~;ߵ]�GEYu���_�ze<�T�"��h�H��w�8����X���ȥ�ޗ�_�x|�/]���O���s\�G|��-	���7o�Mz��E�`��'
h����
�ܺ�:��Ǹ���E�������� \y<ܡ����4�C�(ժ��B�
+���Λv9.����%;���s��Ň�G�^/����09ע(*p�޻)3a����L����\��"��+�� ��۲|
��z����m�	1�CY��ɵ���	B%�xBG�&QH��u������G�]�
O�����i�V����=�n��l��q��M
+��N�����`�I�����$_p������ݶctqxk��ntT�&�	+��7��b�D� 
+ ����	�q�x�%4���M�,4���P�D@{##�
+#7��P`�h������=`�E7��8���f�B�I\?�3�D�8�	��A�P]
+а�CZ��ڦ�
+vI�M�h�n Bpa�D@�GCI��D$6��82��m��%��}!�rq�О��0��J<I�⛅5c�qO{��Z"���yE
]�C�L�CB�`�Dx{#߬�4t�P��w{S�j��4
��j`Ń�,�Q��(x&$�С`h
����=~���.��G�A��)�
+%
+%�ԑu��>��^U�ʋ�R=��F�z꩔/ԏ�E%�n�"\��Ho ��Uy���-[��Q�RV�a}�1pQŞF�����(,��D���7zG�W���+8�O)�΍}qB���'��Jȓ�������򫎣ի_âȡ�)�jnm2PV�=C�@�r�����o����uO�����]��� �[
�M�Q֠(w��ds���!:�I�A�3ρB��Bu&h:��2�!��qHwA��y�#�(t�Ӟc�j��٪PK#��<���@�� sT�EVohm�	�'��{Ȥ�KW�5�Qo�·C�ȇ74�n�x<��3��/�F
G�H]x�(U>�	���s��L�`i}�}\N����U~Ðz�Y)iD"I}"y�3������0�j�)���!x6V�d����&�:@ӂ��pnc���t��0!YBX���i����L���p����1~�`�
�<�y�,|���؇qW����W�N�g1��?=��܏q�I"o?e������}&�m�a"�I��b���H覚��l������f����3$s��7p&H����j�P��ImWCw��"��$��Y��Y�D��'
TԔ�60E>��{p��-�ݒ��K� �
o^���.O�h���8;ou�<��}��Ρj0~��o�7`+N��5G_�ƴ�����A6:��Ԧ�V�Aj�.����m���N+��V�袩7�8��5[
�
+qF�% t^i(Dn�e'�N��՚���f�w!��^/#��wEj��k^{-v��,=�J�]nq�,�s;��hp�9U����
L��6r�IB�l4X���Y4
�i��������
+�)�/��ED$(DLIEND�B`�j
+
"
+@�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�ZK�E�%��ټ �\,ʕK�q��=?q������ʕ�P�8�1+���;�;TuWuU�􌽎c��L��z|_���=�d�ٗ_e]���j��w�|WW�_�~�;8�~�wp�]����+��	�۾���x����}'�o�.z��۾���@��;���i�o�~�wp�������}����3����{<zx���7��$|�ɭ&���-͜����_?�c���eY?���-7�T����v
��m��/�8yx2=<�}���y�߻}�ȊrR��tyZU���j2�\�.o��p�?�e?�}����lA�b�,������<-�C��>}u�Q�T�f�yV��:���r>E
?�f�,����&��t���W>}t����QSA�*ۍo�~����q	{������q����?<��Gw���tAYg�}&A(g���Y�Aͬ�W͎<�91�3:�S^36j6&�
+�ް�a��ɬ�Y<*�m8���<F�R�lG͇
+�}�0?��������ɿ�E�\�~;ߵ]�GEYu���_�ze<�T�"��h�H��w�8����X���ȥ�ޗ�_�x|�/]���O���s\�G|��-	���7o�Mz��E�`��'
h����
�ܺ�:��Ǹ���E�������� \y<ܡ����4�C�(ժ��B�
+���Λv9.����%;���s��Ň�G�^/����09ע(*p�޻)3a����L����\��"��+�� ��۲|
��z����m�	1�CY��ɵ���	B%�xBG�&QH��u������G�]�
O�����i�V����=�n��l��q��M
+��N�����`�I�����$_p������ݶctqxk��ntT�&�	+��7��b�D� 
+ ����	�q�x�%4���M�,4���P�D@{##�
+#7��P`�h������=`�E7��8���f�B�I\?�3�D�8�	��A�P]
+а�CZ��ڦ�
+vI�M�h�n Bpa�D@�GCI��D$6��82��m��%��}!�rq�О��0��J<I�⛅5c�qO{��Z"���yE
]�C�L�CB�`�Dx{#߬�4t�P��w{S�j��4
��j`Ń�,�Q��(x&$�С`h
����=~���.��G�A��)�
+%
+%�ԑu��>��^U�ʋ�R=��F�z꩔/ԏ�E%�n�"\��Ho ��Uy���-[��Q�RV�a}�1pQŞF�����(,��D���7zG�W���+8�O)�΍}qB���'��Jȓ�������򫎣ի_âȡ�)�jnm2PV�=C�@�r�����o����uO�����]��� �[
�M�Q֠(w��ds���!:�I�A�3ρB��Bu&h:��2�!��qHwA��y�#�(t�Ӟc�j��٪PK#��<���@�� sT�EVohm�	�'��{Ȥ�KW�5�Qo�·C�ȇ74�n�x<��3��/�F
G�H]x�(U>�	���s��L�`i}�}\N����U~Ðz�Y)iD"I}"y�3������0�j�)���!x6V�d����&�:@ӂ��pnc���t��0!YBX���i����L���p����1~�`�
�<�y�,|���؇qW����W�N�g1��?=��܏q�I"o?e������}&�m�a"�I��b���H覚��l������f����3$s��7p&H����j�P��ImWCw��"��$��Y��Y�D��'
TԔ�60E>��{p��-�ݒ��K� �
o^���.O�h���8;ou�<��}��Ρj0~��o�7`+N��5G_�ƴ�����A6:��Ԧ�V�Aj�.����m���N+��V�袩7�8��5[
�
+qF�% t^i(Dn�e'�N��՚���f�w!��^/#��wEj��k^{-v��,=�J�]nq�,�s;��hp�9U����
L��6r�IB�l4X���Y4
�i��������
+�)�/��ED$(DLIEND�B`�j
+
"
+@�
+�P‡�������
j�
+TT
��PNG
+
+
IHDRTT�	���IDATx
�ZK�E�%��ټ �\,ʕK�q��=?q������ʕ�P�8�1+���;�;TuWuU�􌽎c��L��z|_���=�d�ٗ_e]���j��w�|WW�_�~�;8�~�wp�]����+��	�۾���x����}'�o�.z��۾���@��;���i�o�~�wp�������}����3����{<zx���7��$|�ɭ&���-͜����_?�c���eY?���-7�T����v
��m��/�8yx2=<�}���y�߻}�ȊrR��tyZU���j2�\�.o��p�?�e?�}����lA�b�,������<-�C��>}u�Q�T�f�yV��:���r>E
?�f�,����&��t���W>}t����QSA�*ۍo�~����q	{������q����?<��Gw���tAYg�}&A(g���Y�Aͬ�W͎<�91�3:�S^36j6&�
+�ް�a��ɬ�Y<*�m8���<F�R�lG͇
+�}�0?��������ɿ�E�\�~;ߵ]�GEYu���_�ze<�T�"��h�H��w�8����X���ȥ�ޗ�_�x|�/]���O���s\�G|��-	���7o�Mz��E�`��'
h����
�ܺ�:��Ǹ���E�������� \y<ܡ����4�C�(ժ��B�
+���Λv9.����%;���s��Ň�G�^/����09ע(*p�޻)3a����L����\��"��+�� ��۲|
��z����m�	1�CY��ɵ���	B%�xBG�&QH��u������G�]�
O�����i�V����=�n��l��q��M
+��N�����`�I�����$_p������ݶctqxk��ntT�&�	+��7��b�D� 
+ ����	�q�x�%4���M�,4���P�D@{##�
+#7��P`�h������=`�E7��8���f�B�I\?�3�D�8�	��A�P]
+а�CZ��ڦ�
+vI�M�h�n Bpa�D@�GCI��D$6��82��m��%��}!�rq�О��0��J<I�⛅5c�qO{��Z"���yE
]�C�L�CB�`�Dx{#߬�4t�P��w{S�j��4
��j`Ń�,�Q��(x&$�С`h
����=~���.��G�A��)�
+%
+%�ԑu��>��^U�ʋ�R=��F�z꩔/ԏ�E%�n�"\��Ho ��Uy���-[��Q�RV�a}�1pQŞF�����(,��D���7zG�W���+8�O)�΍}qB���'��Jȓ�������򫎣ի_âȡ�)�jnm2PV�=C�@�r�����o����uO�����]��� �[
�M�Q֠(w��ds���!:�I�A�3ρB��Bu&h:��2�!��qHwA��y�#�(t�Ӟc�j��٪PK#��<���@�� sT�EVohm�	�'��{Ȥ�KW�5�Qo�·C�ȇ74�n�x<��3��/�F
G�H]x�(U>�	���s��L�`i}�}\N����U~Ðz�Y)iD"I}"y�3������0�j�)���!x6V�d����&�:@ӂ��pnc���t��0!YBX���i����L���p����1~�`�
�<�y�,|���؇qW����W�N�g1��?=��܏q�I"o?e������}&�m�a"�I��b���H覚��l������f����3$s��7p&H����j�P��ImWCw��"��$��Y��Y�D��'
TԔ�60E>��{p��-�ݒ��K� �
o^���.O�h���8;ou�<��}��Ρj0~��o�7`+N��5G_�ƴ�����A6:��Ԧ�V�Aj�.����m���N+��V�袩7�8��5[
�
+qF�% t^i(Dn�e'�N��՚���f�w!��^/#��wEj��k^{-v��,=�J�]nq�,�s;��hp�9U����
L��6r�IB�l4X���Y4
�i��������
+�)�/��ED$(DLIEND�B`�j
+
"

ml-agents/mlagents/trainers/tests/torch/test_policy.py

+import pytest
+
+import torch
+from mlagents.trainers.policy.torch_policy import TorchPolicy
+from mlagents.trainers.tests import mock_brain as mb
+from mlagents.trainers.settings import TrainerSettings, NetworkSettings
+from mlagents.trainers.torch.utils import ModelUtils
+
+VECTOR_ACTION_SPACE = 2
+VECTOR_OBS_SPACE = 8
+DISCRETE_ACTION_SPACE = [3, 3, 3, 2]
+BUFFER_INIT_SAMPLES = 32
+NUM_AGENTS = 12
+EPSILON = 1e-7
+
+
+def create_policy_mock(
+    dummy_config: TrainerSettings,
+    use_rnn: bool = False,
+    use_discrete: bool = True,
+    use_visual: bool = False,
+    seed: int = 0,
+) -> TorchPolicy:
+    mock_spec = mb.setup_test_behavior_specs(
+        use_discrete,
+        use_visual,
+        vector_action_space=DISCRETE_ACTION_SPACE
+        if use_discrete
+        else VECTOR_ACTION_SPACE,
+        vector_obs_space=VECTOR_OBS_SPACE,
+    )
+
+    trainer_settings = dummy_config
+    trainer_settings.keep_checkpoints = 3
+    trainer_settings.network_settings.memory = (
+        NetworkSettings.MemorySettings() if use_rnn else None
+    )
+    policy = TorchPolicy(seed, mock_spec, trainer_settings)
+    return policy
+
+
+@pytest.mark.parametrize("discrete", [True, False], ids=["discrete", "continuous"])
+@pytest.mark.parametrize("visual", [True, False], ids=["visual", "vector"])
+@pytest.mark.parametrize("rnn", [True, False], ids=["rnn", "no_rnn"])
+def test_policy_evaluate(rnn, visual, discrete):
+    # Test evaluate
+    policy = create_policy_mock(
+        TrainerSettings(), use_rnn=rnn, use_discrete=discrete, use_visual=visual
+    )
+    decision_step, terminal_step = mb.create_steps_from_behavior_spec(
+        policy.behavior_spec, num_agents=NUM_AGENTS
+    )
+
+    run_out = policy.evaluate(decision_step, list(decision_step.agent_id))
+    if discrete:
+        run_out["action"].shape == (NUM_AGENTS, len(DISCRETE_ACTION_SPACE))
+    else:
+        assert run_out["action"].shape == (NUM_AGENTS, VECTOR_ACTION_SPACE)
+
+
+@pytest.mark.parametrize("discrete", [True, False], ids=["discrete", "continuous"])
+@pytest.mark.parametrize("visual", [True, False], ids=["visual", "vector"])
+@pytest.mark.parametrize("rnn", [True, False], ids=["rnn", "no_rnn"])
+def test_evaluate_actions(rnn, visual, discrete):
+    policy = create_policy_mock(
+        TrainerSettings(), use_rnn=rnn, use_discrete=discrete, use_visual=visual
+    )
+    buffer = mb.simulate_rollout(64, policy.behavior_spec, memory_size=policy.m_size)
+    vec_obs = [ModelUtils.list_to_tensor(buffer["vector_obs"])]
+    act_masks = ModelUtils.list_to_tensor(buffer["action_mask"])
+    if policy.use_continuous_act:
+        actions = ModelUtils.list_to_tensor(buffer["actions"]).unsqueeze(-1)
+    else:
+        actions = ModelUtils.list_to_tensor(buffer["actions"], dtype=torch.long)
+    vis_obs = []
+    for idx, _ in enumerate(policy.actor_critic.network_body.visual_encoders):
+        vis_ob = ModelUtils.list_to_tensor(buffer["visual_obs%d" % idx])
+        vis_obs.append(vis_ob)
+
+    memories = [
+        ModelUtils.list_to_tensor(buffer["memory"][i])
+        for i in range(0, len(buffer["memory"]), policy.sequence_length)
+    ]
+    if len(memories) > 0:
+        memories = torch.stack(memories).unsqueeze(0)
+
+    log_probs, entropy, values = policy.evaluate_actions(
+        vec_obs,
+        vis_obs,
+        masks=act_masks,
+        actions=actions,
+        memories=memories,
+        seq_len=policy.sequence_length,
+    )
+    assert log_probs.shape == (64, policy.behavior_spec.action_size)
+    assert entropy.shape == (64, policy.behavior_spec.action_size)
+    for val in values.values():
+        assert val.shape == (64,)
+
+
+@pytest.mark.parametrize("discrete", [True, False], ids=["discrete", "continuous"])
+@pytest.mark.parametrize("visual", [True, False], ids=["visual", "vector"])
+@pytest.mark.parametrize("rnn", [True, False], ids=["rnn", "no_rnn"])
+def test_sample_actions(rnn, visual, discrete):
+    policy = create_policy_mock(
+        TrainerSettings(), use_rnn=rnn, use_discrete=discrete, use_visual=visual
+    )
+    buffer = mb.simulate_rollout(64, policy.behavior_spec, memory_size=policy.m_size)
+    vec_obs = [ModelUtils.list_to_tensor(buffer["vector_obs"])]
+    act_masks = ModelUtils.list_to_tensor(buffer["action_mask"])
+
+    vis_obs = []
+    for idx, _ in enumerate(policy.actor_critic.network_body.visual_encoders):
+        vis_ob = ModelUtils.list_to_tensor(buffer["visual_obs%d" % idx])
+        vis_obs.append(vis_ob)
+
+    memories = [
+        ModelUtils.list_to_tensor(buffer["memory"][i])
+        for i in range(0, len(buffer["memory"]), policy.sequence_length)
+    ]
+    if len(memories) > 0:
+        memories = torch.stack(memories).unsqueeze(0)
+
+    (
+        sampled_actions,
+        log_probs,
+        entropies,
+        sampled_values,
+        memories,
+    ) = policy.sample_actions(
+        vec_obs,
+        vis_obs,
+        masks=act_masks,
+        memories=memories,
+        seq_len=policy.sequence_length,
+        all_log_probs=not policy.use_continuous_act,
+    )
+    if discrete:
+        assert log_probs.shape == (
+            64,
+            sum(policy.behavior_spec.discrete_action_branches),
+        )
+    else:
+        assert log_probs.shape == (64, policy.behavior_spec.action_shape)
+    assert entropies.shape == (64, policy.behavior_spec.action_size)
+    for val in sampled_values.values():
+        assert val.shape == (64,)
+
+    if rnn:
+        assert memories.shape == (1, 1, policy.m_size)

ml-agents/mlagents/trainers/saver/tf_saver.py

+import os
+import shutil
+from typing import Optional, Union, cast
+from mlagents_envs.exception import UnityPolicyException
+from mlagents_envs.logging_util import get_logger
+from mlagents.tf_utils import tf
+from mlagents.trainers.saver.saver import BaseSaver
+from mlagents.trainers.tf.model_serialization import export_policy_model
+from mlagents.trainers.settings import TrainerSettings, SerializationSettings
+from mlagents.trainers.policy.tf_policy import TFPolicy
+from mlagents.trainers.optimizer.tf_optimizer import TFOptimizer
+from mlagents.trainers import __version__
+
+
+logger = get_logger(__name__)
+
+
+class TFSaver(BaseSaver):
+    """
+    Saver class for TensorFlow
+    """
+
+    def __init__(
+        self, trainer_settings: TrainerSettings, model_path: str, load: bool = False
+    ):
+        super().__init__()
+        self.model_path = model_path
+        self.initialize_path = trainer_settings.init_path
+        self._keep_checkpoints = trainer_settings.keep_checkpoints
+        self.load = load
+
+        # Currently only support saving one policy. This is the one to be saved.
+        self.policy: Optional[TFPolicy] = None
+        self.graph = None
+        self.sess = None
+        self.tf_saver = None
+
+    def register(self, module: Union[TFPolicy, TFOptimizer]) -> None:
+        if isinstance(module, TFPolicy):
+            self._register_policy(module)
+        elif isinstance(module, TFOptimizer):
+            self._register_optimizer(module)
+        else:
+            raise UnityPolicyException(
+                "Registering Object of unsupported type {} to Saver ".format(
+                    type(module)
+                )
+            )
+
+    def _register_policy(self, policy: TFPolicy) -> None:
+        if self.policy is None:
+            self.policy = policy
+            self.graph = self.policy.graph
+            self.sess = self.policy.sess
+            with self.policy.graph.as_default():
+                self.tf_saver = tf.train.Saver(max_to_keep=self._keep_checkpoints)
+
+    def save_checkpoint(self, behavior_name: str, step: int) -> str:
+        checkpoint_path = os.path.join(self.model_path, f"{behavior_name}-{step}")
+        # Save the TF checkpoint and graph definition
+        if self.graph:
+            with self.graph.as_default():
+                if self.tf_saver:
+                    self.tf_saver.save(self.sess, f"{checkpoint_path}.ckpt")
+                tf.train.write_graph(
+                    self.graph, self.model_path, "raw_graph_def.pb", as_text=False
+                )
+        # also save the policy so we have optimized model files for each checkpoint
+        self.export(checkpoint_path, behavior_name)
+        return checkpoint_path
+
+    def export(self, output_filepath: str, behavior_name: str) -> None:
+        # save model if there is only one worker or
+        # only on worker-0 if there are multiple workers
+        if self.policy and self.policy.rank is not None and self.policy.rank != 0:
+            return
+        export_policy_model(
+            self.model_path, output_filepath, behavior_name, self.graph, self.sess
+        )
+
+    def initialize_or_load(self, policy: Optional[TFPolicy] = None) -> None:
+        # If there is an initialize path, load from that. Else, load from the set model path.
+        # If load is set to True, don't reset steps to 0. Else, do. This allows a user to,
+        # e.g., resume from an initialize path.
+        if policy is None:
+            policy = self.policy
+        policy = cast(TFPolicy, policy)
+        reset_steps = not self.load
+        if self.initialize_path is not None:
+            self._load_graph(
+                policy, self.initialize_path, reset_global_steps=reset_steps
+            )
+        elif self.load:
+            self._load_graph(policy, self.model_path, reset_global_steps=reset_steps)
+        else:
+            policy.initialize()
+
+        TFPolicy.broadcast_global_variables(0)
+
+    def _load_graph(
+        self, policy: TFPolicy, model_path: str, reset_global_steps: bool = False
+    ) -> None:
+        with policy.graph.as_default():
+            logger.info(f"Loading model from {model_path}.")
+            ckpt = tf.train.get_checkpoint_state(model_path)
+            if ckpt is None:
+                raise UnityPolicyException(
+                    "The model {} could not be loaded. Make "
+                    "sure you specified the right "
+                    "--run-id and that the previous run you are loading from had the same "
+                    "behavior names.".format(model_path)
+                )
+            if self.tf_saver:
+                try:
+                    self.tf_saver.restore(policy.sess, ckpt.model_checkpoint_path)
+                except tf.errors.NotFoundError:
+                    raise UnityPolicyException(
+                        "The model {} was found but could not be loaded. Make "
+                        "sure the model is from the same version of ML-Agents, has the same behavior parameters, "
+                        "and is using the same trainer configuration as the current run.".format(
+                            model_path
+                        )
+                    )
+            self._check_model_version(__version__)
+            if reset_global_steps:
+                policy.set_step(0)
+                logger.info(
+                    "Starting training from step 0 and saving to {}.".format(
+                        self.model_path
+                    )
+                )
+            else:
+                logger.info(f"Resuming training from step {policy.get_current_step()}.")
+
+    def _check_model_version(self, version: str) -> None:
+        """
+        Checks whether the model being loaded was created with the same version of
+        ML-Agents, and throw a warning if not so.
+        """
+        if self.policy is not None and self.policy.version_tensors is not None:
+            loaded_ver = tuple(
+                num.eval(session=self.sess) for num in self.policy.version_tensors
+            )
+            if loaded_ver != TFPolicy._convert_version_string(version):
+                logger.warning(
+                    f"The model checkpoint you are loading from was saved with ML-Agents version "
+                    f"{loaded_ver[0]}.{loaded_ver[1]}.{loaded_ver[2]} but your current ML-Agents"
+                    f"version is {version}. Model may not behave properly."
+                )
+
+    def copy_final_model(self, source_nn_path: str) -> None:
+        """
+        Copy the .nn file at the given source to the destination.
+        Also copies the corresponding .onnx file if it exists.
+        """
+        final_model_name = os.path.splitext(source_nn_path)[0]
+
+        if SerializationSettings.convert_to_barracuda:
+            source_path = f"{final_model_name}.nn"
+            destination_path = f"{self.model_path}.nn"
+            shutil.copyfile(source_path, destination_path)
+            logger.info(f"Copied {source_path} to {destination_path}.")
+
+        if SerializationSettings.convert_to_onnx:
+            try:
+                source_path = f"{final_model_name}.onnx"
+                destination_path = f"{self.model_path}.onnx"
+                shutil.copyfile(source_path, destination_path)
+                logger.info(f"Copied {source_path} to {destination_path}.")
+            except OSError:
+                pass

ml-agents/mlagents/trainers/saver/torch_saver.py

+import os
+import shutil
+import torch
+from typing import Dict, Union, Optional, cast
+from mlagents_envs.exception import UnityPolicyException
+from mlagents_envs.logging_util import get_logger
+from mlagents.trainers.saver.saver import BaseSaver
+from mlagents.trainers.settings import TrainerSettings, SerializationSettings
+from mlagents.trainers.policy.torch_policy import TorchPolicy
+from mlagents.trainers.optimizer.torch_optimizer import TorchOptimizer
+from mlagents.trainers.torch.model_serialization import ModelSerializer
+
+
+logger = get_logger(__name__)
+
+
+class TorchSaver(BaseSaver):
+    """
+    Saver class for PyTorch
+    """
+
+    def __init__(
+        self, trainer_settings: TrainerSettings, model_path: str, load: bool = False
+    ):
+        super().__init__()
+        self.model_path = model_path
+        self.initialize_path = trainer_settings.init_path
+        self._keep_checkpoints = trainer_settings.keep_checkpoints
+        self.load = load
+
+        self.policy: Optional[TorchPolicy] = None
+        self.exporter: Optional[ModelSerializer] = None
+        self.modules: Dict[str, torch.nn.Modules] = {}
+
+    def register(self, module: Union[TorchPolicy, TorchOptimizer]) -> None:
+        if isinstance(module, TorchPolicy) or isinstance(module, TorchOptimizer):
+            self.modules.update(module.get_modules())  # type: ignore
+        else:
+            raise UnityPolicyException(
+                "Registering Object of unsupported type {} to Saver ".format(
+                    type(module)
+                )
+            )
+        if self.policy is None and isinstance(module, TorchPolicy):
+            self.policy = module
+            self.exporter = ModelSerializer(self.policy)
+
+    def save_checkpoint(self, behavior_name: str, step: int) -> str:
+        if not os.path.exists(self.model_path):
+            os.makedirs(self.model_path)
+        checkpoint_path = os.path.join(self.model_path, f"{behavior_name}-{step}")
+        state_dict = {
+            name: module.state_dict() for name, module in self.modules.items()
+        }
+        torch.save(state_dict, f"{checkpoint_path}.pt")
+        torch.save(state_dict, os.path.join(self.model_path, "checkpoint.pt"))
+        self.export(checkpoint_path, behavior_name)
+        return checkpoint_path
+
+    def export(self, output_filepath: str, behavior_name: str) -> None:
+        if self.exporter is not None:
+            self.exporter.export_policy_model(output_filepath)
+
+    def initialize_or_load(self, policy: Optional[TorchPolicy] = None) -> None:
+        # Initialize/Load registered self.policy by default.
+        # If given input argument policy, use the input policy instead.
+        # This argument is mainly for initialization of the ghost trainer's fixed policy.
+        reset_steps = not self.load
+        if self.initialize_path is not None:
+            self._load_model(
+                self.initialize_path, policy, reset_global_steps=reset_steps
+            )
+        elif self.load:
+            self._load_model(self.model_path, policy, reset_global_steps=reset_steps)
+
+    def _load_model(
+        self,
+        load_path: str,
+        policy: Optional[TorchPolicy] = None,
+        reset_global_steps: bool = False,
+    ) -> None:
+        model_path = os.path.join(load_path, "checkpoint.pt")
+        saved_state_dict = torch.load(model_path)
+        if policy is None:
+            modules = self.modules
+            policy = self.policy
+        else:
+            modules = policy.get_modules()
+        policy = cast(TorchPolicy, policy)
+
+        for name, mod in modules.items():
+            mod.load_state_dict(saved_state_dict[name])
+
+        if reset_global_steps:
+            policy.set_step(0)
+            logger.info(
+                "Starting training from step 0 and saving to {}.".format(
+                    self.model_path
+                )
+            )
+        else:
+            logger.info(f"Resuming training from step {policy.get_current_step()}.")
+
+    def copy_final_model(self, source_nn_path: str) -> None:
+        """
+        Copy the .nn file at the given source to the destination.
+        Also copies the corresponding .onnx file if it exists.
+        """
+        final_model_name = os.path.splitext(source_nn_path)[0]
+
+        if SerializationSettings.convert_to_onnx:
+            try:
+                source_path = f"{final_model_name}.onnx"
+                destination_path = f"{self.model_path}.onnx"
+                shutil.copyfile(source_path, destination_path)
+                logger.info(f"Copied {source_path} to {destination_path}.")
+            except OSError:
+                pass

ml-agents/mlagents/trainers/torch/components/reward_providers/__init__.py

+from mlagents.trainers.torch.components.reward_providers.base_reward_provider import (  # noqa F401
+    BaseRewardProvider,
+)
+from mlagents.trainers.torch.components.reward_providers.extrinsic_reward_provider import (  # noqa F401
+    ExtrinsicRewardProvider,
+)
+from mlagents.trainers.torch.components.reward_providers.curiosity_reward_provider import (  # noqa F401
+    CuriosityRewardProvider,
+)
+from mlagents.trainers.torch.components.reward_providers.gail_reward_provider import (  # noqa F401
+    GAILRewardProvider,
+)
+from mlagents.trainers.torch.components.reward_providers.reward_provider_factory import (  # noqa F401
+    create_reward_provider,
+)

ml-agents/mlagents/trainers/torch/components/reward_providers/base_reward_provider.py

+import numpy as np
+from abc import ABC, abstractmethod
+from typing import Dict
+
+from mlagents.trainers.buffer import AgentBuffer
+from mlagents.trainers.settings import RewardSignalSettings
+from mlagents_envs.base_env import BehaviorSpec
+
+
+class BaseRewardProvider(ABC):
+    def __init__(self, specs: BehaviorSpec, settings: RewardSignalSettings) -> None:
+        self._policy_specs = specs
+        self._gamma = settings.gamma
+        self._strength = settings.strength
+        self._ignore_done = False
+
+    @property
+    def gamma(self) -> float:
+        """
+        The discount factor for the reward signal
+        """
+        return self._gamma
+
+    @property
+    def strength(self) -> float:
+        """
+        The strength multiplier of the reward provider
+        """
+        return self._strength
+
+    @property
+    def name(self) -> str:
+        """
+        The name of the reward provider. Is used for reporting and identification
+        """
+        class_name = self.__class__.__name__
+        return class_name.replace("RewardProvider", "")
+
+    @property
+    def ignore_done(self) -> bool:
+        """
+        If true, when the agent is done, the rewards of the next episode must be
+        used to calculate the return of the current episode.
+        Is used to mitigate the positive bias in rewards with no natural end.
+        """
+        return self._ignore_done
+
+    @abstractmethod
+    def evaluate(self, mini_batch: AgentBuffer) -> np.ndarray:
+        """
+        Evaluates the reward for the data present in the Dict mini_batch. Use this when evaluating a reward
+        function drawn straight from a Buffer.
+        :param mini_batch: A Dict of numpy arrays (the format used by our Buffer)
+            when drawing from the update buffer.
+        :return: a np.ndarray of rewards generated by the reward provider
+        """
+        raise NotImplementedError(
+            "The reward provider's evaluate method has not been implemented "
+        )
+
+    @abstractmethod
+    def update(self, mini_batch: AgentBuffer) -> Dict[str, np.ndarray]:
+        """
+        Update the reward for the data present in the Dict mini_batch. Use this when updating a reward
+        function drawn straight from a Buffer.
+        :param mini_batch: A Dict of numpy arrays (the format used by our Buffer)
+            when drawing from the update buffer.
+        :return: A dictionary from string to stats values
+        """
+        raise NotImplementedError(
+            "The reward provider's update method has not been implemented "
+        )

ml-agents/mlagents/trainers/torch/components/reward_providers/extrinsic_reward_provider.py

+import numpy as np
+from typing import Dict
+
+from mlagents.trainers.buffer import AgentBuffer
+from mlagents.trainers.torch.components.reward_providers.base_reward_provider import (
+    BaseRewardProvider,
+)
+
+
+class ExtrinsicRewardProvider(BaseRewardProvider):
+    def evaluate(self, mini_batch: AgentBuffer) -> np.ndarray:
+        return np.array(mini_batch["environment_rewards"], dtype=np.float32)
+
+    def update(self, mini_batch: AgentBuffer) -> Dict[str, np.ndarray]:
+        return {}

ml-agents/mlagents/trainers/torch/components/reward_providers/reward_provider_factory.py

+from typing import Dict, Type
+from mlagents.trainers.exception import UnityTrainerException
+
+from mlagents.trainers.settings import RewardSignalSettings, RewardSignalType
+
+from mlagents.trainers.torch.components.reward_providers.base_reward_provider import (
+    BaseRewardProvider,
+)
+from mlagents.trainers.torch.components.reward_providers.extrinsic_reward_provider import (
+    ExtrinsicRewardProvider,
+)
+from mlagents.trainers.torch.components.reward_providers.curiosity_reward_provider import (
+    CuriosityRewardProvider,
+)
+from mlagents.trainers.torch.components.reward_providers.gail_reward_provider import (
+    GAILRewardProvider,
+)
+
+from mlagents_envs.base_env import BehaviorSpec
+
+NAME_TO_CLASS: Dict[RewardSignalType, Type[BaseRewardProvider]] = {
+    RewardSignalType.EXTRINSIC: ExtrinsicRewardProvider,
+    RewardSignalType.CURIOSITY: CuriosityRewardProvider,
+    RewardSignalType.GAIL: GAILRewardProvider,
+}
+
+
+def create_reward_provider(
+    name: RewardSignalType, specs: BehaviorSpec, settings: RewardSignalSettings
+) -> BaseRewardProvider:
+    """
+    Creates a reward provider class based on the name and config entry provided as a dict.
+    :param name: The name of the reward signal
+    :param specs: The BehaviorSpecs of the policy
+    :param settings: The RewardSignalSettings for that reward signal
+    :return: The reward signal class instantiated
+    """
+    rcls = NAME_TO_CLASS.get(name)
+    if not rcls:
+        raise UnityTrainerException(f"Unknown reward signal type {name}")
+
+    class_inst = rcls(specs, settings)
+    return class_inst

ml-agents/mlagents/trainers/torch/components/reward_providers/curiosity_reward_provider.py

+import numpy as np
+from typing import Dict
+import torch
+
+from mlagents.trainers.buffer import AgentBuffer
+from mlagents.trainers.torch.components.reward_providers.base_reward_provider import (
+    BaseRewardProvider,
+)
+from mlagents.trainers.settings import CuriositySettings
+
+from mlagents_envs.base_env import BehaviorSpec
+from mlagents.trainers.torch.utils import ModelUtils
+from mlagents.trainers.torch.networks import NetworkBody
+from mlagents.trainers.torch.layers import linear_layer, Swish
+from mlagents.trainers.settings import NetworkSettings, EncoderType
+
+
+class CuriosityRewardProvider(BaseRewardProvider):
+    beta = 0.2  # Forward vs Inverse loss weight
+    loss_multiplier = 10.0  # Loss multiplier
+
+    def __init__(self, specs: BehaviorSpec, settings: CuriositySettings) -> None:
+        super().__init__(specs, settings)
+        self._ignore_done = True
+        self._network = CuriosityNetwork(specs, settings)
+        self.optimizer = torch.optim.Adam(
+            self._network.parameters(), lr=settings.learning_rate
+        )
+        self._has_updated_once = False
+
+    def evaluate(self, mini_batch: AgentBuffer) -> np.ndarray:
+        with torch.no_grad():
+            rewards = self._network.compute_reward(mini_batch).detach().cpu().numpy()
+        rewards = np.minimum(rewards, 1.0 / self.strength)
+        return rewards * self._has_updated_once
+
+    def update(self, mini_batch: AgentBuffer) -> Dict[str, np.ndarray]:
+        self._has_updated_once = True
+        forward_loss = self._network.compute_forward_loss(mini_batch)
+        inverse_loss = self._network.compute_inverse_loss(mini_batch)
+
+        loss = self.loss_multiplier * (
+            self.beta * forward_loss + (1.0 - self.beta) * inverse_loss
+        )
+        self.optimizer.zero_grad()
+        loss.backward()
+        self.optimizer.step()
+        return {
+            "Losses/Curiosity Forward Loss": forward_loss.detach().cpu().numpy(),
+            "Losses/Curiosity Inverse Loss": inverse_loss.detach().cpu().numpy(),
+        }
+
+
+class CuriosityNetwork(torch.nn.Module):
+    EPSILON = 1e-10
+
+    def __init__(self, specs: BehaviorSpec, settings: CuriositySettings) -> None:
+        super().__init__()
+        self._policy_specs = specs
+        state_encoder_settings = NetworkSettings(
+            normalize=False,
+            hidden_units=settings.encoding_size,
+            num_layers=2,
+            vis_encode_type=EncoderType.SIMPLE,
+            memory=None,
+        )
+        self._state_encoder = NetworkBody(
+            specs.observation_shapes, state_encoder_settings
+        )
+
+        self._action_flattener = ModelUtils.ActionFlattener(specs)
+
+        self.inverse_model_action_predition = torch.nn.Sequential(
+            linear_layer(2 * settings.encoding_size, 256),
+            Swish(),
+            linear_layer(256, self._action_flattener.flattened_size),
+        )
+
+        self.forward_model_next_state_prediction = torch.nn.Sequential(
+            linear_layer(
+                settings.encoding_size + self._action_flattener.flattened_size, 256
+            ),
+            Swish(),
+            linear_layer(256, settings.encoding_size),
+        )
+
+    def get_current_state(self, mini_batch: AgentBuffer) -> torch.Tensor:
+        """
+        Extracts the current state embedding from a mini_batch.
+        """
+        n_vis = len(self._state_encoder.visual_encoders)
+        hidden, _ = self._state_encoder.forward(
+            vec_inputs=[
+                ModelUtils.list_to_tensor(mini_batch["vector_obs"], dtype=torch.float)
+            ],
+            vis_inputs=[
+                ModelUtils.list_to_tensor(
+                    mini_batch["visual_obs%d" % i], dtype=torch.float
+                )
+                for i in range(n_vis)
+            ],
+        )
+        return hidden
+
+    def get_next_state(self, mini_batch: AgentBuffer) -> torch.Tensor:
+        """
+        Extracts the next state embedding from a mini_batch.
+        """
+        n_vis = len(self._state_encoder.visual_encoders)
+        hidden, _ = self._state_encoder.forward(
+            vec_inputs=[
+                ModelUtils.list_to_tensor(
+                    mini_batch["next_vector_in"], dtype=torch.float
+                )
+            ],
+            vis_inputs=[
+                ModelUtils.list_to_tensor(
+                    mini_batch["next_visual_obs%d" % i], dtype=torch.float
+                )
+                for i in range(n_vis)
+            ],
+        )
+        return hidden
+
+    def predict_action(self, mini_batch: AgentBuffer) -> torch.Tensor:
+        """
+        In the continuous case, returns the predicted action.
+        In the discrete case, returns the logits.
+        """
+        inverse_model_input = torch.cat(
+            (self.get_current_state(mini_batch), self.get_next_state(mini_batch)), dim=1
+        )
+        hidden = self.inverse_model_action_predition(inverse_model_input)
+        if self._policy_specs.is_action_continuous():
+            return hidden
+        else:
+            branches = ModelUtils.break_into_branches(
+                hidden, self._policy_specs.discrete_action_branches
+            )
+            branches = [torch.softmax(b, dim=1) for b in branches]
+            return torch.cat(branches, dim=1)
+
+    def predict_next_state(self, mini_batch: AgentBuffer) -> torch.Tensor:
+        """
+        Uses the current state embedding and the action of the mini_batch to predict
+        the next state embedding.
+        """
+        if self._policy_specs.is_action_continuous():
+            action = ModelUtils.list_to_tensor(mini_batch["actions"], dtype=torch.float)
+        else:
+            action = torch.cat(
+                ModelUtils.actions_to_onehot(
+                    ModelUtils.list_to_tensor(mini_batch["actions"], dtype=torch.long),
+                    self._policy_specs.discrete_action_branches,
+                ),
+                dim=1,
+            )
+        forward_model_input = torch.cat(
+            (self.get_current_state(mini_batch), action), dim=1
+        )
+
+        return self.forward_model_next_state_prediction(forward_model_input)
+
+    def compute_inverse_loss(self, mini_batch: AgentBuffer) -> torch.Tensor:
+        """
+        Computes the inverse loss for a mini_batch. Corresponds to the error on the
+        action prediction (given the current and next state).
+        """
+        predicted_action = self.predict_action(mini_batch)
+        if self._policy_specs.is_action_continuous():
+            sq_difference = (
+                ModelUtils.list_to_tensor(mini_batch["actions"], dtype=torch.float)
+                - predicted_action
+            ) ** 2
+            sq_difference = torch.sum(sq_difference, dim=1)
+            return torch.mean(
+                ModelUtils.dynamic_partition(
+                    sq_difference,
+                    ModelUtils.list_to_tensor(mini_batch["masks"], dtype=torch.float),
+                    2,
+                )[1]
+            )
+        else:
+            true_action = torch.cat(
+                ModelUtils.actions_to_onehot(
+                    ModelUtils.list_to_tensor(mini_batch["actions"], dtype=torch.long),
+                    self._policy_specs.discrete_action_branches,
+                ),
+                dim=1,
+            )
+            cross_entropy = torch.sum(
+                -torch.log(predicted_action + self.EPSILON) * true_action, dim=1
+            )
+            return torch.mean(
+                ModelUtils.dynamic_partition(
+                    cross_entropy,
+                    ModelUtils.list_to_tensor(
+                        mini_batch["masks"], dtype=torch.float
+                    ),  # use masks not action_masks
+                    2,
+                )[1]
+            )
+
+    def compute_reward(self, mini_batch: AgentBuffer) -> torch.Tensor:
+        """
+        Calculates the curiosity reward for the mini_batch. Corresponds to the error
+        between the predicted and actual next state.
+        """
+        predicted_next_state = self.predict_next_state(mini_batch)
+        target = self.get_next_state(mini_batch)
+        sq_difference = 0.5 * (target - predicted_next_state) ** 2
+        sq_difference = torch.sum(sq_difference, dim=1)
+        return sq_difference
+
+    def compute_forward_loss(self, mini_batch: AgentBuffer) -> torch.Tensor:
+        """
+        Computes the loss for the next state prediction
+        """
+        return torch.mean(
+            ModelUtils.dynamic_partition(
+                self.compute_reward(mini_batch),
+                ModelUtils.list_to_tensor(mini_batch["masks"], dtype=torch.float),
+                2,
+            )[1]
+        )

ml-agents/mlagents/trainers/torch/components/reward_providers/gail_reward_provider.py

+from typing import Optional, Dict
+import numpy as np
+import torch
+
+from mlagents.trainers.buffer import AgentBuffer
+from mlagents.trainers.torch.components.reward_providers.base_reward_provider import (
+    BaseRewardProvider,
+)
+from mlagents.trainers.settings import GAILSettings
+from mlagents_envs.base_env import BehaviorSpec
+from mlagents.trainers.torch.utils import ModelUtils
+from mlagents.trainers.torch.networks import NetworkBody
+from mlagents.trainers.torch.layers import linear_layer, Swish, Initialization
+from mlagents.trainers.settings import NetworkSettings, EncoderType
+from mlagents.trainers.demo_loader import demo_to_buffer
+
+
+class GAILRewardProvider(BaseRewardProvider):
+    def __init__(self, specs: BehaviorSpec, settings: GAILSettings) -> None:
+        super().__init__(specs, settings)
+        self._ignore_done = True
+        self._discriminator_network = DiscriminatorNetwork(specs, settings)
+        _, self._demo_buffer = demo_to_buffer(
+            settings.demo_path, 1, specs
+        )  # This is supposed to be the sequence length but we do not have access here
+        params = list(self._discriminator_network.parameters())
+        self.optimizer = torch.optim.Adam(params, lr=settings.learning_rate)
+
+    def evaluate(self, mini_batch: AgentBuffer) -> np.ndarray:
+        with torch.no_grad():
+            estimates, _ = self._discriminator_network.compute_estimate(
+                mini_batch, use_vail_noise=False
+            )
+            return (
+                -torch.log(
+                    1.0
+                    - estimates.squeeze(dim=1)
+                    * (1.0 - self._discriminator_network.EPSILON)
+                )
+                .detach()
+                .cpu()
+                .numpy()
+            )
+
+    def update(self, mini_batch: AgentBuffer) -> Dict[str, np.ndarray]:
+        expert_batch = self._demo_buffer.sample_mini_batch(
+            mini_batch.num_experiences, 1
+        )
+        loss, stats_dict = self._discriminator_network.compute_loss(
+            mini_batch, expert_batch
+        )
+        self.optimizer.zero_grad()
+        loss.backward()
+        self.optimizer.step()
+        return stats_dict
+
+
+class DiscriminatorNetwork(torch.nn.Module):
+    gradient_penalty_weight = 10.0
+    z_size = 128
+    alpha = 0.0005
+    mutual_information = 0.5
+    EPSILON = 1e-7
+    initial_beta = 0.0
+
+    def __init__(self, specs: BehaviorSpec, settings: GAILSettings) -> None:
+        super().__init__()
+        self._policy_specs = specs
+        self._use_vail = settings.use_vail
+        self._settings = settings
+
+        state_encoder_settings = NetworkSettings(
+            normalize=False,
+            hidden_units=settings.encoding_size,
+            num_layers=2,
+            vis_encode_type=EncoderType.SIMPLE,
+            memory=None,
+        )
+        self._state_encoder = NetworkBody(
+            specs.observation_shapes, state_encoder_settings
+        )
+
+        self._action_flattener = ModelUtils.ActionFlattener(specs)
+
+        encoder_input_size = settings.encoding_size
+        if settings.use_actions:
+            encoder_input_size += (
+                self._action_flattener.flattened_size + 1
+            )  # + 1 is for done
+
+        self.encoder = torch.nn.Sequential(
+            linear_layer(encoder_input_size, settings.encoding_size),
+            Swish(),
+            linear_layer(settings.encoding_size, settings.encoding_size),
+            Swish(),
+        )
+
+        estimator_input_size = settings.encoding_size
+        if settings.use_vail:
+            estimator_input_size = self.z_size
+            self._z_sigma = torch.nn.Parameter(
+                torch.ones((self.z_size), dtype=torch.float), requires_grad=True
+            )
+            self._z_mu_layer = linear_layer(
+                settings.encoding_size,
+                self.z_size,
+                kernel_init=Initialization.KaimingHeNormal,
+                kernel_gain=0.1,
+            )
+            self._beta = torch.nn.Parameter(
+                torch.tensor(self.initial_beta, dtype=torch.float), requires_grad=False
+            )
+
+        self._estimator = torch.nn.Sequential(
+            linear_layer(estimator_input_size, 1), torch.nn.Sigmoid()
+        )
+
+    def get_action_input(self, mini_batch: AgentBuffer) -> torch.Tensor:
+        """
+        Creates the action Tensor. In continuous case, corresponds to the action. In
+        the discrete case, corresponds to the concatenation of one hot action Tensors.
+        """
+        return self._action_flattener.forward(
+            torch.as_tensor(mini_batch["actions"], dtype=torch.float)
+        )
+
+    def get_state_encoding(self, mini_batch: AgentBuffer) -> torch.Tensor:
+        """
+        Creates the observation input.
+        """
+        n_vis = len(self._state_encoder.visual_encoders)
+        hidden, _ = self._state_encoder.forward(
+            vec_inputs=[torch.as_tensor(mini_batch["vector_obs"], dtype=torch.float)],
+            vis_inputs=[
+                torch.as_tensor(mini_batch["visual_obs%d" % i], dtype=torch.float)
+                for i in range(n_vis)
+            ],
+        )
+        return hidden
+
+    def compute_estimate(
+        self, mini_batch: AgentBuffer, use_vail_noise: bool = False
+    ) -> torch.Tensor:
+        """
+        Given a mini_batch, computes the estimate (How much the discriminator believes
+        the data was sampled from the demonstration data).
+        :param mini_batch: The AgentBuffer of data
+        :param use_vail_noise: Only when using VAIL : If true, will sample the code, if
+        false, will return the mean of the code.
+        """
+        encoder_input = self.get_state_encoding(mini_batch)
+        if self._settings.use_actions:
+            actions = self.get_action_input(mini_batch)
+            dones = torch.as_tensor(mini_batch["done"], dtype=torch.float)
+            encoder_input = torch.cat([encoder_input, actions, dones], dim=1)
+        hidden = self.encoder(encoder_input)
+        z_mu: Optional[torch.Tensor] = None
+        if self._settings.use_vail:
+            z_mu = self._z_mu_layer(hidden)
+            hidden = torch.normal(z_mu, self._z_sigma * use_vail_noise)
+        estimate = self._estimator(hidden)
+        return estimate, z_mu
+
+    def compute_loss(
+        self, policy_batch: AgentBuffer, expert_batch: AgentBuffer
+    ) -> torch.Tensor:
+        """
+        Given a policy mini_batch and an expert mini_batch, computes the loss of the discriminator.
+        """
+        total_loss = torch.zeros(1)
+        stats_dict: Dict[str, np.ndarray] = {}
+        policy_estimate, policy_mu = self.compute_estimate(
+            policy_batch, use_vail_noise=True
+        )
+        expert_estimate, expert_mu = self.compute_estimate(
+            expert_batch, use_vail_noise=True
+        )
+        stats_dict["Policy/GAIL Policy Estimate"] = (
+            policy_estimate.mean().detach().cpu().numpy()
+        )
+        stats_dict["Policy/GAIL Expert Estimate"] = (
+            expert_estimate.mean().detach().cpu().numpy()
+        )
+        discriminator_loss = -(
+            torch.log(expert_estimate + self.EPSILON)
+            + torch.log(1.0 - policy_estimate + self.EPSILON)
+        ).mean()
+        stats_dict["Losses/GAIL Loss"] = discriminator_loss.detach().cpu().numpy()
+        total_loss += discriminator_loss
+        if self._settings.use_vail:
+            # KL divergence loss (encourage latent representation to be normal)
+            kl_loss = torch.mean(
+                -torch.sum(
+                    1
+                    + (self._z_sigma ** 2).log()
+                    - 0.5 * expert_mu ** 2
+                    - 0.5 * policy_mu ** 2
+                    - (self._z_sigma ** 2),
+                    dim=1,
+                )
+            )
+            vail_loss = self._beta * (kl_loss - self.mutual_information)
+            with torch.no_grad():
+                self._beta.data = torch.max(
+                    self._beta + self.alpha * (kl_loss - self.mutual_information),
+                    torch.tensor(0.0),
+                )
+            total_loss += vail_loss
+            stats_dict["Policy/GAIL Beta"] = self._beta.detach().cpu().numpy()
+            stats_dict["Losses/GAIL KL Loss"] = kl_loss.detach().cpu().numpy()
+        if self.gradient_penalty_weight > 0.0:
+            total_loss += (
+                self.gradient_penalty_weight
+                * self.compute_gradient_magnitude(policy_batch, expert_batch)
+            )
+        return total_loss, stats_dict
+
+    def compute_gradient_magnitude(
+        self, policy_batch: AgentBuffer, expert_batch: AgentBuffer
+    ) -> torch.Tensor:
+        """
+        Gradient penalty from https://arxiv.org/pdf/1704.00028. Adds stability esp.
+        for off-policy. Compute gradients w.r.t randomly interpolated input.
+        """
+        policy_obs = self.get_state_encoding(policy_batch)
+        expert_obs = self.get_state_encoding(expert_batch)
+        obs_epsilon = torch.rand(policy_obs.shape)
+        encoder_input = obs_epsilon * policy_obs + (1 - obs_epsilon) * expert_obs
+        if self._settings.use_actions:
+            policy_action = self.get_action_input(policy_batch)
+            expert_action = self.get_action_input(policy_batch)
+            action_epsilon = torch.rand(policy_action.shape)
+            policy_dones = torch.as_tensor(policy_batch["done"], dtype=torch.float)
+            expert_dones = torch.as_tensor(expert_batch["done"], dtype=torch.float)
+            dones_epsilon = torch.rand(policy_dones.shape)
+            encoder_input = torch.cat(
+                [
+                    encoder_input,
+                    action_epsilon * policy_action
+                    + (1 - action_epsilon) * expert_action,
+                    dones_epsilon * policy_dones + (1 - dones_epsilon) * expert_dones,
+                ],
+                dim=1,
+            )
+        hidden = self.encoder(encoder_input)
+        if self._settings.use_vail:
+            use_vail_noise = True
+            z_mu = self._z_mu_layer(hidden)
+            hidden = torch.normal(z_mu, self._z_sigma * use_vail_noise)
+        hidden = self._estimator(hidden)
+        estimate = torch.mean(torch.sum(hidden, dim=1))
+        gradient = torch.autograd.grad(estimate, encoder_input)[0]
+        # Norm's gradient could be NaN at 0. Use our own safe_norm
+        safe_norm = (torch.sum(gradient ** 2, dim=1) + self.EPSILON).sqrt()
+        gradient_mag = torch.mean((safe_norm - 1) ** 2)
+        return gradient_mag

ml-agents/mlagents/trainers/torch/components/bc/module.py

+from typing import Dict
+import numpy as np
+import torch
+
+from mlagents.trainers.policy.torch_policy import TorchPolicy
+from mlagents.trainers.demo_loader import demo_to_buffer
+from mlagents.trainers.settings import BehavioralCloningSettings, ScheduleType
+from mlagents.trainers.torch.utils import ModelUtils
+
+
+class BCModule:
+    def __init__(
+        self,
+        policy: TorchPolicy,
+        settings: BehavioralCloningSettings,
+        policy_learning_rate: float,
+        default_batch_size: int,
+        default_num_epoch: int,
+    ):
+        """
+        A BC trainer that can be used inline with RL.
+        :param policy: The policy of the learning model
+        :param settings: The settings for BehavioralCloning including LR strength, batch_size,
+        num_epochs, samples_per_update and LR annealing steps.
+        :param policy_learning_rate: The initial Learning Rate of the policy. Used to set an appropriate learning rate
+            for the pretrainer.
+        """
+        self.policy = policy
+        self._anneal_steps = settings.steps
+        self.current_lr = policy_learning_rate * settings.strength
+
+        learning_rate_schedule: ScheduleType = ScheduleType.LINEAR if self._anneal_steps > 0 else ScheduleType.CONSTANT
+        self.decay_learning_rate = ModelUtils.DecayedValue(
+            learning_rate_schedule, self.current_lr, 1e-10, self._anneal_steps
+        )
+        params = self.policy.actor_critic.parameters()
+        self.optimizer = torch.optim.Adam(params, lr=self.current_lr)
+        _, self.demonstration_buffer = demo_to_buffer(
+            settings.demo_path, policy.sequence_length, policy.behavior_spec
+        )
+
+        self.batch_size = (
+            settings.batch_size if settings.batch_size else default_batch_size
+        )
+        self.num_epoch = settings.num_epoch if settings.num_epoch else default_num_epoch
+        self.n_sequences = max(
+            min(self.batch_size, self.demonstration_buffer.num_experiences)
+            // policy.sequence_length,
+            1,
+        )
+
+        self.has_updated = False
+        self.use_recurrent = self.policy.use_recurrent
+        self.samples_per_update = settings.samples_per_update
+
+    def update(self) -> Dict[str, np.ndarray]:
+        """
+        Updates model using buffer.
+        :param max_batches: The maximum number of batches to use per update.
+        :return: The loss of the update.
+        """
+        # Don't continue training if the learning rate has reached 0, to reduce training time.
+
+        decay_lr = self.decay_learning_rate.get_value(self.policy.get_current_step())
+        if self.current_lr <= 0:
+            return {"Losses/Pretraining Loss": 0}
+
+        batch_losses = []
+        possible_demo_batches = (
+            self.demonstration_buffer.num_experiences // self.n_sequences
+        )
+        possible_batches = possible_demo_batches
+
+        max_batches = self.samples_per_update // self.n_sequences
+
+        n_epoch = self.num_epoch
+        for _ in range(n_epoch):
+            self.demonstration_buffer.shuffle(
+                sequence_length=self.policy.sequence_length
+            )
+            if max_batches == 0:
+                num_batches = possible_batches
+            else:
+                num_batches = min(possible_batches, max_batches)
+            for i in range(num_batches // self.policy.sequence_length):
+                demo_update_buffer = self.demonstration_buffer
+                start = i * self.n_sequences * self.policy.sequence_length
+                end = (i + 1) * self.n_sequences * self.policy.sequence_length
+                mini_batch_demo = demo_update_buffer.make_mini_batch(start, end)
+                run_out = self._update_batch(mini_batch_demo, self.n_sequences)
+                loss = run_out["loss"]
+                batch_losses.append(loss)
+
+        ModelUtils.update_learning_rate(self.optimizer, decay_lr)
+        self.current_lr = decay_lr
+
+        self.has_updated = True
+        update_stats = {"Losses/Pretraining Loss": np.mean(batch_losses)}
+        return update_stats
+
+    def _behavioral_cloning_loss(self, selected_actions, log_probs, expert_actions):
+        if self.policy.use_continuous_act:
+            bc_loss = torch.nn.functional.mse_loss(selected_actions, expert_actions)
+        else:
+            log_prob_branches = ModelUtils.break_into_branches(
+                log_probs, self.policy.act_size
+            )
+            bc_loss = torch.mean(
+                torch.stack(
+                    [
+                        torch.sum(
+                            -torch.nn.functional.log_softmax(log_prob_branch, dim=1)
+                            * expert_actions_branch,
+                            dim=1,
+                        )
+                        for log_prob_branch, expert_actions_branch in zip(
+                            log_prob_branches, expert_actions
+                        )
+                    ]
+                )
+            )
+        return bc_loss
+
+    def _update_batch(
+        self, mini_batch_demo: Dict[str, np.ndarray], n_sequences: int
+    ) -> Dict[str, float]:
+        """
+        Helper function for update_batch.
+        """
+        vec_obs = [ModelUtils.list_to_tensor(mini_batch_demo["vector_obs"])]
+        act_masks = None
+        if self.policy.use_continuous_act:
+            expert_actions = ModelUtils.list_to_tensor(mini_batch_demo["actions"])
+        else:
+            raw_expert_actions = ModelUtils.list_to_tensor(
+                mini_batch_demo["actions"], dtype=torch.long
+            )
+            expert_actions = ModelUtils.actions_to_onehot(
+                raw_expert_actions, self.policy.act_size
+            )
+            act_masks = ModelUtils.list_to_tensor(
+                np.ones(
+                    (
+                        self.n_sequences * self.policy.sequence_length,
+                        sum(self.policy.behavior_spec.discrete_action_branches),
+                    ),
+                    dtype=np.float32,
+                )
+            )
+
+        memories = []
+        if self.policy.use_recurrent:
+            memories = torch.zeros(1, self.n_sequences, self.policy.m_size)
+
+        if self.policy.use_vis_obs:
+            vis_obs = []
+            for idx, _ in enumerate(
+                self.policy.actor_critic.network_body.visual_encoders
+            ):
+                vis_ob = ModelUtils.list_to_tensor(
+                    mini_batch_demo["visual_obs%d" % idx]
+                )
+                vis_obs.append(vis_ob)
+        else:
+            vis_obs = []
+
+        selected_actions, all_log_probs, _, _, _ = self.policy.sample_actions(
+            vec_obs,
+            vis_obs,
+            masks=act_masks,
+            memories=memories,
+            seq_len=self.policy.sequence_length,
+            all_log_probs=True,
+        )
+        bc_loss = self._behavioral_cloning_loss(
+            selected_actions, all_log_probs, expert_actions
+        )
+        self.optimizer.zero_grad()
+        bc_loss.backward()
+
+        self.optimizer.step()
+        run_out = {"loss": bc_loss.detach().cpu().numpy()}
+        return run_out

ml-agents/mlagents/trainers/torch/model_serialization.py

+import os
+import torch
+
+from mlagents_envs.logging_util import get_logger
+from mlagents.trainers.settings import SerializationSettings
+
+
+logger = get_logger(__name__)
+
+
+class ModelSerializer:
+    def __init__(self, policy):
+        self.policy = policy
+        batch_dim = [1]
+        dummy_vec_obs = [torch.zeros(batch_dim + [self.policy.vec_obs_size])]
+        dummy_vis_obs = (
+            [torch.zeros(batch_dim + list(self.policy.vis_obs_shape))]
+            if self.policy.vis_obs_size > 0
+            else []
+        )
+        dummy_masks = torch.ones(batch_dim + [sum(self.policy.actor_critic.act_size)])
+        dummy_memories = torch.zeros(batch_dim + [1] + [self.policy.export_memory_size])
+
+        # Need to pass all posslible inputs since currently keyword arguments is not
+        # supported by torch.nn.export()
+        self.dummy_input = (dummy_vec_obs, dummy_vis_obs, dummy_masks, dummy_memories)
+
+        # Input names can only contain actual input used since in torch.nn.export
+        # it maps input_names only to input nodes that exist in the graph
+        self.input_names = []
+        self.dynamic_axes = {"action": {0: "batch"}, "action_probs": {0: "batch"}}
+        if self.policy.use_vec_obs:
+            self.input_names.append("vector_observation")
+            self.dynamic_axes.update({"vector_observation": {0: "batch"}})
+        if self.policy.use_vis_obs:
+            self.input_names.append("visual_observation")
+            self.dynamic_axes.update({"visual_observation": {0: "batch"}})
+        if not self.policy.use_continuous_act:
+            self.input_names.append("action_masks")
+            self.dynamic_axes.update({"action_masks": {0: "batch"}})
+        if self.policy.use_recurrent:
+            self.input_names.append("memories")
+            self.dynamic_axes.update({"memories": {0: "batch"}})
+
+        self.output_names = [
+            "action",
+            "action_probs",
+            "version_number",
+            "memory_size",
+            "is_continuous_control",
+            "action_output_shape",
+        ]
+
+    def export_policy_model(self, output_filepath: str) -> None:
+        """
+        Exports a Torch model for a Policy to .onnx format for Unity embedding.
+
+        :param output_filepath: file path to output the model (without file suffix)
+        :param brain_name: Brain name of brain to be trained
+        """
+        if not os.path.exists(output_filepath):
+            os.makedirs(output_filepath)
+
+        onnx_output_path = f"{output_filepath}.onnx"
+        logger.info(f"Converting to {onnx_output_path}")
+
+        torch.onnx.export(
+            self.policy.actor_critic,
+            self.dummy_input,
+            onnx_output_path,
+            verbose=False,
+            opset_version=SerializationSettings.onnx_opset,
+            input_names=self.input_names,
+            output_names=self.output_names,
+            dynamic_axes=self.dynamic_axes,
+        )
+        logger.info(f"Exported {onnx_output_path}")

Project/Assets/ML-Agents/Examples/SharedAssets/Scripts/OrientationCubeController.cs.meta

-fileFormatVersion: 2
-guid: 771e78c5e980e440e8cd19716b55075f
-MonoImporter:
-  externalObjects: {}
-  serializedVersion: 2
-  defaultReferences: []
-  executionOrder: 0
-  icon: {instanceID: 0}
-  userData: 
-  assetBundleName: 
-  assetBundleVariant: 

Project/Assets/csc.rsp

--warnaserror+
--warnaserror-:618

com.unity.ml-agents/Runtime/SensorHelper.cs.meta

-fileFormatVersion: 2
-guid: 7c1189c0af42c46f7b533350d49ad3e7
-MonoImporter:
-  externalObjects: {}
-  serializedVersion: 2
-  defaultReferences: []
-  executionOrder: 0
-  icon: {instanceID: 0}
-  userData: 
-  assetBundleName: 
-  assetBundleVariant: 

utils/validate_release_links.py

-#!/usr/bin/env python3
-
-import ast
-import sys
-import os
-import re
-import subprocess
-from typing import List, Optional, Pattern
-
-RELEASE_PATTERN = re.compile(r"release_[0-9]+(_docs)*")
-TRAINER_INIT_FILE = "ml-agents/mlagents/trainers/__init__.py"
-
-# Filename -> regex list to allow specific lines.
-# To allow everything in the file, use None for the value
-ALLOW_LIST = {
-    # Previous release table
-    "README.md": re.compile(r"\*\*Release [0-9]+\*\*"),
-    "docs/Versioning.md": None,
-    "com.unity.ml-agents/CHANGELOG.md": None,
-    "utils/make_readme_table.py": None,
-    "utils/validate_release_links.py": None,
-}
-
-
-def test_pattern():
-    # Just some sanity check that the regex works as expected.
-    assert RELEASE_PATTERN.search(
-        "https://github.com/Unity-Technologies/ml-agents/blob/release_4_docs/Food.md"
-    )
-    assert RELEASE_PATTERN.search(
-        "https://github.com/Unity-Technologies/ml-agents/blob/release_4/Foo.md"
-    )
-    assert RELEASE_PATTERN.search(
-        "git clone --branch release_4 https://github.com/Unity-Technologies/ml-agents.git"
-    )
-    assert RELEASE_PATTERN.search(
-        "https://github.com/Unity-Technologies/ml-agents/blob/release_123_docs/Foo.md"
-    )
-    assert RELEASE_PATTERN.search(
-        "https://github.com/Unity-Technologies/ml-agents/blob/release_123/Foo.md"
-    )
-    assert not RELEASE_PATTERN.search(
-        "https://github.com/Unity-Technologies/ml-agents/blob/latest_release/docs/Foo.md"
-    )
-    print("tests OK!")
-
-
-def git_ls_files() -> List[str]:
-    """
-    Run "git ls-files" and return a list with one entry per line.
-    This returns the list of all files tracked by git.
-    """
-    return subprocess.check_output(["git", "ls-files"], universal_newlines=True).split(
-        "\n"
-    )
-
-
-def get_release_tag() -> Optional[str]:
-    """
-    Returns the release tag for the mlagents python package.
-    This will be None on the master branch.
-    :return:
-    """
-    with open(TRAINER_INIT_FILE) as f:
-        for line in f:
-            if "__release_tag__" in line:
-                lhs, equals_string, rhs = line.strip().partition(" = ")
-                # Evaluate the right hand side of the expression
-                return ast.literal_eval(rhs)
-    # If we couldn't find the release tag, raise an exception
-    # (since we can't return None here)
-    raise RuntimeError("Can't determine release tag")
-
-
-def check_file(filename: str, global_allow_pattern: Pattern) -> List[str]:
-    """
-    Validate a single file and return any offending lines.
-    """
-    bad_lines = []
-    with open(filename) as f:
-        for line in f:
-            if not RELEASE_PATTERN.search(line):
-                continue
-
-            if global_allow_pattern.search(line):
-                continue
-
-            if filename in ALLOW_LIST:
-                if ALLOW_LIST[filename] is None or ALLOW_LIST[filename].search(line):
-                    continue
-
-            bad_lines.append(f"{filename}: {line.strip()}")
-    return bad_lines
-
-
-def check_all_files(allow_pattern: Pattern) -> List[str]:
-    """
-    Validate all files tracked by git.
-    :param allow_pattern:
-    """
-    bad_lines = []
-    file_types = {".py", ".md", ".cs"}
-    for file_name in git_ls_files():
-        if "localized" in file_name or os.path.splitext(file_name)[1] not in file_types:
-            continue
-        bad_lines += check_file(file_name, allow_pattern)
-    return bad_lines
-
-
-def main():
-    release_tag = get_release_tag()
-    if not release_tag:
-        print("Release tag is None, exiting")
-        sys.exit(0)
-
-    print(f"Release tag: {release_tag}")
-    allow_pattern = re.compile(f"{release_tag}(_docs)*")
-    bad_lines = check_all_files(allow_pattern)
-    if bad_lines:
-        print(
-            f"Found lines referring to previous release. Either update the files, or add an exclusion to {__file__}"
-        )
-        for line in bad_lines:
-            print(line)
-
-    sys.exit(1 if bad_lines else 0)
-
-
-if __name__ == "__main__":
-    if "--test" in sys.argv:
-        test_pattern()
-    main()

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

-import pytest
-
-from mlagents.trainers.tf.models import ModelUtils
-from mlagents.tf_utils import tf
-from mlagents_envs.base_env import BehaviorSpec, ActionType
-
-
-def create_behavior_spec(num_visual, num_vector, vector_size):
-    behavior_spec = BehaviorSpec(
-        [(84, 84, 3)] * int(num_visual) + [(vector_size,)] * int(num_vector),
-        ActionType.DISCRETE,
-        (1,),
-    )
-    return behavior_spec
-
-
-@pytest.mark.parametrize("num_visual", [1, 2, 4])
-@pytest.mark.parametrize("num_vector", [1, 2, 4])
-def test_create_input_placeholders(num_vector, num_visual):
-    vec_size = 8
-    name_prefix = "test123"
-    bspec = create_behavior_spec(num_visual, num_vector, vec_size)
-    vec_in, vis_in = ModelUtils.create_input_placeholders(
-        bspec.observation_shapes, name_prefix=name_prefix
-    )
-
-    assert isinstance(vis_in, list)
-    assert len(vis_in) == num_visual
-    assert isinstance(vec_in, tf.Tensor)
-    assert vec_in.get_shape().as_list()[1] == num_vector * 8
-
-    # Check names contain prefix and vis shapes are correct
-    for _vis in vis_in:
-        assert _vis.get_shape().as_list() == [None, 84, 84, 3]
-        assert _vis.name.startswith(name_prefix)
-    assert vec_in.name.startswith(name_prefix)

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

-import pytest
-from unittest import mock
-import os
-import unittest
-import tempfile
-
-import numpy as np
-from mlagents.tf_utils import tf
-from mlagents.trainers.saver.tf_saver import TFSaver
-from mlagents.trainers import __version__
-from mlagents.trainers.settings import TrainerSettings
-from mlagents.trainers.policy.tf_policy import TFPolicy
-from mlagents.trainers.tests import mock_brain as mb
-from mlagents.trainers.tests.test_nn_policy import create_policy_mock
-from mlagents.trainers.ppo.optimizer import PPOOptimizer
-
-
-def test_register(tmp_path):
-    trainer_params = TrainerSettings()
-    saver = TFSaver(trainer_params, tmp_path)
-
-    opt = mock.Mock(spec=PPOOptimizer)
-    saver.register(opt)
-    assert saver.policy is None
-
-    trainer_params = TrainerSettings()
-    policy = create_policy_mock(trainer_params)
-    saver.register(policy)
-    assert saver.policy is not None
-
-
-class ModelVersionTest(unittest.TestCase):
-    def test_version_compare(self):
-        # Test write_stats
-        with self.assertLogs("mlagents.trainers", level="WARNING") as cm:
-            trainer_params = TrainerSettings()
-            mock_path = tempfile.mkdtemp()
-            policy = create_policy_mock(trainer_params)
-            saver = TFSaver(trainer_params, mock_path)
-            saver.register(policy)
-
-            saver._check_model_version(
-                "0.0.0"
-            )  # This is not the right version for sure
-            # Assert that 1 warning has been thrown with incorrect version
-            assert len(cm.output) == 1
-            saver._check_model_version(__version__)  # This should be the right version
-            # Assert that no additional warnings have been thrown wth correct ver
-            assert len(cm.output) == 1
-
-
-def test_load_save(tmp_path):
-    path1 = os.path.join(tmp_path, "runid1")
-    path2 = os.path.join(tmp_path, "runid2")
-    trainer_params = TrainerSettings()
-    policy = create_policy_mock(trainer_params)
-    saver = TFSaver(trainer_params, path1)
-    saver.register(policy)
-    saver.initialize_or_load(policy)
-    policy.set_step(2000)
-
-    mock_brain_name = "MockBrain"
-    saver.save_checkpoint(mock_brain_name, 2000)
-    assert len(os.listdir(tmp_path)) > 0
-
-    # Try load from this path
-    saver = TFSaver(trainer_params, path1, load=True)
-    policy2 = create_policy_mock(trainer_params)
-    saver.register(policy2)
-    saver.initialize_or_load(policy2)
-    _compare_two_policies(policy, policy2)
-    assert policy2.get_current_step() == 2000
-
-    # Try initialize from path 1
-    trainer_params.init_path = path1
-    saver = TFSaver(trainer_params, path2)
-    policy3 = create_policy_mock(trainer_params)
-    saver.register(policy3)
-    saver.initialize_or_load(policy3)
-
-    _compare_two_policies(policy2, policy3)
-    # Assert that the steps are 0.
-    assert policy3.get_current_step() == 0
-
-
-def _compare_two_policies(policy1: TFPolicy, policy2: TFPolicy) -> None:
-    """
-    Make sure two policies have the same output for the same input.
-    """
-    decision_step, _ = mb.create_steps_from_behavior_spec(
-        policy1.behavior_spec, num_agents=1
-    )
-    run_out1 = policy1.evaluate(decision_step, list(decision_step.agent_id))
-    run_out2 = policy2.evaluate(decision_step, list(decision_step.agent_id))
-
-    np.testing.assert_array_equal(run_out2["log_probs"], run_out1["log_probs"])
-
-
-@pytest.mark.parametrize("discrete", [True, False], ids=["discrete", "continuous"])
-@pytest.mark.parametrize("visual", [True, False], ids=["visual", "vector"])
-@pytest.mark.parametrize("rnn", [True, False], ids=["rnn", "no_rnn"])
-def test_checkpoint_conversion(tmpdir, rnn, visual, discrete):
-    tf.reset_default_graph()
-    dummy_config = TrainerSettings()
-    model_path = os.path.join(tmpdir, "Mock_Brain")
-    policy = create_policy_mock(
-        dummy_config, use_rnn=rnn, use_discrete=discrete, use_visual=visual
-    )
-    trainer_params = TrainerSettings()
-    saver = TFSaver(trainer_params, model_path)
-    saver.register(policy)
-    saver.save_checkpoint("Mock_Brain", 100)
-    assert os.path.isfile(model_path + "/Mock_Brain-100.nn")