Move test_env_perf file

config/ppo/3DBall.yaml

     summary_freq: 12000
     use_recurrent: false
     vis_encode_type: simple
+    threaded: false
     reward_signals:
       extrinsic:
         strength: 1.0

Project/ProjectSettings/EditorBuildSettings.asset

 EditorBuildSettings:
   m_ObjectHideFlags: 0
   serializedVersion: 2
-  m_Scenes: []
+  m_Scenes:
+  - enabled: 1
+    path: Assets/ML-Agents/Examples/3DBall/Scenes/3DBall.unity
+    guid: b9ac0cbf961bf4dacbfa0aa9c0d60aaa
   m_configObjects: {}

Project/ProjectSettings/ProjectVersion.txt

-m_EditorVersion: 2018.4.17f1
+m_EditorVersion: 2018.4.20f1

Project/ProjectSettings/UnityConnectSettings.asset

 UnityConnectSettings:
   m_ObjectHideFlags: 0
   serializedVersion: 1
-  m_Enabled: 1
+  m_Enabled: 0
   m_TestMode: 0
   m_EventOldUrl: https://api.uca.cloud.unity3d.com/v1/events
   m_EventUrl: https://cdp.cloud.unity3d.com/v1/events

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

 import abc
 import time
 
-from mlagents.trainers.optimizer.tf_optimizer import TFOptimizer
+from mlagents.trainers.optimizer.optimizer import Optimizer
 from mlagents.trainers.buffer import AgentBuffer
 from mlagents.trainers.trainer import Trainer
 from mlagents.trainers.exception import UnityTrainerException
             for agent_id in rewards:
                 rewards[agent_id] = 0
 
-    def _update_end_episode_stats(self, agent_id: str, optimizer: TFOptimizer) -> None:
+    def _update_end_episode_stats(self, agent_id: str, optimizer: Optimizer) -> None:
         for name, rewards in self.collected_rewards.items():
             if name == "environment":
                 self.stats_reporter.add_stat(

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

 from collections import deque
 
 from mlagents_envs.logging_util import get_logger
-from mlagents.model_serialization import export_policy_model, SerializationSettings
-from mlagents.trainers.policy.tf_policy import TFPolicy
+
-from mlagents.trainers.policy import Policy
+from mlagents.trainers.policy.policy import Policy
 from mlagents.trainers.exception import UnityTrainerException
 from mlagents.trainers.behavior_id_utils import BehaviorIdentifiers
 
         Exports the model
         """
         policy = self.get_policy(name_behavior_id)
-        settings = SerializationSettings(policy.model_path, policy.brain.brain_name)
-        export_policy_model(settings, policy.graph, policy.sess)
+        policy.export_model()
 
     @abc.abstractmethod
     def end_episode(self):
     @abc.abstractmethod
     def create_policy(
         self, parsed_behavior_id: BehaviorIdentifiers, brain_parameters: BrainParameters
-    ) -> TFPolicy:
+    ) -> Policy:
         """
         Creates policy
         """
     def add_policy(
-        self, parsed_behavior_id: BehaviorIdentifiers, policy: TFPolicy
+        self, parsed_behavior_id: BehaviorIdentifiers, policy: Policy
     ) -> None:
         """
         Adds policy to trainer.
     @abc.abstractmethod
-    def get_policy(self, name_behavior_id: str) -> TFPolicy:
+    def get_policy(self, name_behavior_id: str) -> Policy:
         """
         Gets policy from trainer.
         """

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

 import yaml
 
 from mlagents.trainers.ppo.trainer import PPOTrainer, discount_rewards
-from mlagents.trainers.ppo.optimizer import PPOOptimizer
+from mlagents.trainers.ppo.optimizer_tf import TFPPOOptimizer
 from mlagents.trainers.policy.nn_policy import NNPolicy
 from mlagents.trainers.brain import BrainParameters
 from mlagents.trainers.agent_processor import AgentManagerQueue
     policy = NNPolicy(
         0, mock_brain, trainer_parameters, False, False, create_tf_graph=False
     )
-    optimizer = PPOOptimizer(policy, trainer_parameters)
+    optimizer = TFPPOOptimizer(policy, trainer_parameters)
     return optimizer
 
 

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

 import mlagents.trainers.tests.mock_brain as mb
 from mlagents.trainers.policy.nn_policy import NNPolicy
 from mlagents.trainers.sac.optimizer import SACOptimizer
-from mlagents.trainers.ppo.optimizer import PPOOptimizer
+from mlagents.trainers.ppo.optimizer_tf import TFPPOOptimizer
 
 CONTINUOUS_PATH = os.path.dirname(os.path.abspath(__file__)) + "/test.demo"
 DISCRETE_PATH = os.path.dirname(os.path.abspath(__file__)) + "/testdcvis.demo"
     if trainer_parameters["trainer"] == "sac":
         optimizer = SACOptimizer(policy, trainer_parameters)
     else:
-        optimizer = PPOOptimizer(policy, trainer_parameters)
+        optimizer = TFPPOOptimizer(policy, trainer_parameters)
     return optimizer
 
 

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

-import abc
-class Optimizer(abc.ABC):
+class Optimizer(object):
-    @abc.abstractmethod
+    def __init__(self):
+        self.reward_signals = {}
+
     def update(self, batch: AgentBuffer, num_sequences: int) -> Dict[str, float]:
         """
         Update the Policy based on the batch that was passed in.

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

 import abc
 import os
 import numpy as np
+
+from mlagents.model_serialization import SerializationSettings, export_policy_model
-from mlagents_envs.exception import UnityException
+from mlagents.trainers.policy.policy import UnityPolicyException
 from mlagents.trainers.trajectory import SplitObservations
 from mlagents.trainers.brain_conversion_utils import get_global_agent_id
 from mlagents_envs.base_env import DecisionSteps
 logger = get_logger(__name__)
 
 
-class UnityPolicyException(UnityException):
-    """
-    Related to errors with the Trainer.
-    """
-
-    pass
-
-
 class TFPolicy(Policy):
     """
     Contains a learning model, and the necessary
         :param brain: The corresponding Brain for this policy.
         :param trainer_parameters: The trainer parameters.
         """
+        super(TFPolicy, self).__init__(
+            brain=brain, seed=seed, trainer_params=trainer_parameters
+        )
         self._version_number_ = 2
         self.m_size = 0
 
         self.inference_dict = {}
         self.update_dict = {}
         self.sequence_length = 1
-        self.seed = seed
-        self.brain = brain
-        self.use_recurrent = trainer_parameters["use_recurrent"]
-        self.memory_dict: Dict[str, np.ndarray] = {}
-        self.num_branches = len(self.brain.vector_action_space_size)
-        self.previous_action_dict: Dict[str, np.array] = {}
-        self.normalize = trainer_parameters.get("normalize", False)
-        self.use_continuous_act = brain.vector_action_space_type == "continuous"
-        if self.use_continuous_act:
-            self.num_branches = self.brain.vector_action_space_size[0]
-        self.model_path = trainer_parameters["output_path"]
         self.initialize_path = trainer_parameters.get("init_path", None)
         self.keep_checkpoints = trainer_parameters.get("keep_checkpoints", 5)
         self.graph = tf.Graph()
         self.saver = None
         self.seed = seed
-        if self.use_recurrent:
-            self.m_size = trainer_parameters["memory_size"]
-            self.sequence_length = trainer_parameters["sequence_length"]
-            if self.m_size == 0:
-                raise UnityPolicyException(
-                    "The memory size for brain {0} is 0 even "
-                    "though the trainer uses recurrent.".format(brain.brain_name)
-                )
-            elif self.m_size % 2 != 0:
-                raise UnityPolicyException(
-                    "The memory size for brain {0} is {1} "
-                    "but it must be divisible by 2.".format(
-                        brain.brain_name, self.m_size
-                    )
-                )
         self._initialize_tensorflow_references()
         self.load = load
 
         """
         pass
 
+    def load_model(self, step=0):
+        reset_steps = not self.load
+        self._load_graph(self.model_path, reset_global_steps=reset_steps)
+
     def _initialize_graph(self):
         with self.graph.as_default():
             self.saver = tf.train.Saver(max_to_keep=self.keep_checkpoints)
         """
         raise UnityPolicyException("The evaluate function was not implemented.")
 
+    def export_model(self, step=0):
+        settings = SerializationSettings(self.model_path, self.brain.brain_name)
+        export_policy_model(settings, self.graph, self.sess)
+
     def get_action(
         self, decision_requests: DecisionSteps, worker_id: int = 0
     ) -> ActionInfo:
             feed_dict[self.action_masks] = mask
         return feed_dict
 
-    def make_empty_memory(self, num_agents):
-        """
-        Creates empty memory for use with RNNs
-        :param num_agents: Number of agents.
-        :return: Numpy array of zeros.
-        """
-        return np.zeros((num_agents, self.m_size), dtype=np.float32)
-
-    def save_memories(
-        self, agent_ids: List[str], memory_matrix: Optional[np.ndarray]
-    ) -> None:
-        if memory_matrix is None:
-            return
-        for index, agent_id in enumerate(agent_ids):
-            self.memory_dict[agent_id] = memory_matrix[index, :]
-
-    def retrieve_memories(self, agent_ids: List[str]) -> np.ndarray:
-        memory_matrix = np.zeros((len(agent_ids), self.m_size), dtype=np.float32)
-        for index, agent_id in enumerate(agent_ids):
-            if agent_id in self.memory_dict:
-                memory_matrix[index, :] = self.memory_dict[agent_id]
-        return memory_matrix
-
-    def remove_memories(self, agent_ids):
-        for agent_id in agent_ids:
-            if agent_id in self.memory_dict:
-                self.memory_dict.pop(agent_id)
-
-    def make_empty_previous_action(self, num_agents):
-        """
-        Creates empty previous action for use with RNNs and discrete control
-        :param num_agents: Number of agents.
-        :return: Numpy array of zeros.
-        """
-        return np.zeros((num_agents, self.num_branches), dtype=np.int)
-
-    def save_previous_action(
-        self, agent_ids: List[str], action_matrix: Optional[np.ndarray]
-    ) -> None:
-        if action_matrix is None:
-            return
-        for index, agent_id in enumerate(agent_ids):
-            self.previous_action_dict[agent_id] = action_matrix[index, :]
-
-    def retrieve_previous_action(self, agent_ids: List[str]) -> np.ndarray:
-        action_matrix = np.zeros((len(agent_ids), self.num_branches), dtype=np.int)
-        for index, agent_id in enumerate(agent_ids):
-            if agent_id in self.previous_action_dict:
-                action_matrix[index, :] = self.previous_action_dict[agent_id]
-        return action_matrix
-
-    def remove_previous_action(self, agent_ids):
-        for agent_id in agent_ids:
-            if agent_id in self.previous_action_dict:
-                self.previous_action_dict.pop(agent_id)
-
     def get_current_step(self):
         """
         Gets current model step.
         """
         return list(self.update_dict.keys())
 
-    def save_model(self, steps):
+    def save_model(self, step=0):
-        :param steps: The number of steps the model was trained for
+        :param step: The number of steps the model was trained for
-            last_checkpoint = os.path.join(self.model_path, f"model-{steps}.ckpt")
+            last_checkpoint = os.path.join(self.model_path, f"model-{step}.ckpt")
             self.saver.save(self.sess, last_checkpoint)
             tf.train.write_graph(
                 self.graph, self.model_path, "raw_graph_def.pb", as_text=False

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

-from abc import ABC, abstractmethod
+from abc import abstractmethod
+from typing import Dict, List, Optional
+import numpy as np
+from mlagents_envs.exception import UnityException
+
-class Policy(ABC):
-    @abstractmethod
+class UnityPolicyException(UnityException):
+    """
+    Related to errors with the Trainer.
+    """
+
+    pass
+
+
+class Policy(object):
+    def __init__(self, brain, seed, trainer_params):
+        self.brain = brain
+        self.seed = seed
+        self.model_path = None
+        self.use_continuous_act = brain.vector_action_space_type == "continuous"
+        if self.use_continuous_act:
+            self.num_branches = self.brain.vector_action_space_size[0]
+        else:
+            self.num_branches = len(self.brain.vector_action_space_size)
+        self.previous_action_dict: Dict[str, np.array] = {}
+        self.memory_dict: Dict[str, np.ndarray] = {}
+        self.normalize = trainer_params["normalize"]
+        self.use_recurrent = trainer_params["use_recurrent"]
+        self.model_path = trainer_params["output_path"]
+
+        if self.use_recurrent:
+            self.m_size = trainer_params["memory_size"]
+            self.sequence_length = trainer_params["sequence_length"]
+            if self.m_size == 0:
+                raise UnityPolicyException(
+                    "The memory size for brain {0} is 0 even "
+                    "though the trainer uses recurrent.".format(brain.brain_name)
+                )
+            elif self.m_size % 2 != 0:
+                raise UnityPolicyException(
+                    "The memory size for brain {0} is {1} "
+                    "but it must be divisible by 2.".format(
+                        brain.brain_name, self.m_size
+                    )
+                )
+
+    def make_empty_memory(self, num_agents):
+        """
+        Creates empty memory for use with RNNs
+        :param num_agents: Number of agents.
+        :return: Numpy array of zeros.
+        """
+        return np.zeros((num_agents, self.m_size), dtype=np.float32)
+
+    def save_memories(
+        self, agent_ids: List[str], memory_matrix: Optional[np.ndarray]
+    ) -> None:
+        if memory_matrix is None:
+            return
+        for index, agent_id in enumerate(agent_ids):
+            self.memory_dict[agent_id] = memory_matrix[index, :]
+
+    def retrieve_memories(self, agent_ids: List[str]) -> np.ndarray:
+        memory_matrix = np.zeros((len(agent_ids), self.m_size), dtype=np.float32)
+        for index, agent_id in enumerate(agent_ids):
+            if agent_id in self.memory_dict:
+                memory_matrix[index, :] = self.memory_dict[agent_id]
+        return memory_matrix
+
+    def remove_memories(self, agent_ids):
+        for agent_id in agent_ids:
+            if agent_id in self.memory_dict:
+                self.memory_dict.pop(agent_id)
+
+    def make_empty_previous_action(self, num_agents):
+        """
+        Creates empty previous action for use with RNNs and discrete control
+        :param num_agents: Number of agents.
+        :return: Numpy array of zeros.
+        """
+        return np.zeros((num_agents, self.num_branches), dtype=np.int)
+
+    def save_previous_action(
+        self, agent_ids: List[str], action_matrix: Optional[np.ndarray]
+    ) -> None:
+        if action_matrix is None:
+            return
+        for index, agent_id in enumerate(agent_ids):
+            self.previous_action_dict[agent_id] = action_matrix[index, :]
+
+    def retrieve_previous_action(self, agent_ids: List[str]) -> np.ndarray:
+        action_matrix = np.zeros((len(agent_ids), self.num_branches), dtype=np.int)
+        for index, agent_id in enumerate(agent_ids):
+            if agent_id in self.previous_action_dict:
+                action_matrix[index, :] = self.previous_action_dict[agent_id]
+        return action_matrix
+
+    def remove_previous_action(self, agent_ids):
+        for agent_id in agent_ids:
+            if agent_id in self.previous_action_dict:
+                self.previous_action_dict.pop(agent_id)
+
+        raise NotImplementedError
+
+    @abstractmethod
+    def update_normalization(self, vector_obs: np.ndarray) -> None:
+        pass
+
+    @abstractmethod
+    def export_model(self, step=0):
+        pass
+
+    @abstractmethod
+    def save_model(self, step=0):
+        pass
+
+    @abstractmethod
+    def load_model(self, step=0):
+        pass
+
+    @abstractmethod
+    def increment_step(self, n_steps):
+        pass
+
+    @abstractmethod
+    def get_current_step(self):
         pass

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

 from mlagents.trainers.buffer import AgentBuffer
 
 
-class PPOOptimizer(TFOptimizer):
+class TFPPOOptimizer(TFOptimizer):
     def __init__(self, policy: TFPolicy, trainer_params: Dict[str, Any]):
         """
         Takes a Policy and a Dict of trainer parameters and creates an Optimizer around the policy.
             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

 # Contains an implementation of PPO as described in: https://arxiv.org/abs/1707.06347
 
 from collections import defaultdict
-
+import time
+from mlagents.trainers.policy import Policy
-from mlagents.trainers.policy.nn_policy import NNPolicy
-from mlagents.trainers.policy.tf_policy import TFPolicy
-from mlagents.trainers.ppo.optimizer import PPOOptimizer
+from mlagents.trainers.policy.torch_policy import TorchPolicy
+from mlagents.trainers.policy.nn_policy import NNPolicy
+from mlagents.trainers.ppo.optimizer_torch import TorchPPOOptimizer
+from mlagents.trainers.ppo.optimizer_tf import TFPPOOptimizer
+logger = get_logger(__name__)
-logger = get_logger(__name__)
+TIMINGS = []
 
 
 class PPOTrainer(RLTrainer):
         self._check_param_keys()
         self.load = load
         self.seed = seed
-        self.policy: NNPolicy = None  # type: ignore
+        self.framework = "torch"
+        self.policy: Policy = None  # type: ignore
+        self.update_times = []
 
     def _check_param_keys(self):
         super()._check_param_keys()
             trajectory.next_obs,
             trajectory.done_reached and not trajectory.interrupted,
         )
+
         for name, v in value_estimates.items():
             agent_buffer_trajectory["{}_value_estimates".format(name)].extend(v)
             self._stats_reporter.add_stat(
             local_value_estimates = agent_buffer_trajectory[
                 "{}_value_estimates".format(name)
             ].get_batch()
+
             local_advantage = get_gae(
                 rewards=local_rewards,
                 value_estimates=local_value_estimates,
             buffer = self.update_buffer
             max_num_batch = buffer_length // batch_size
             for i in range(0, max_num_batch * batch_size, batch_size):
+                t1 = time.perf_counter()
+                t2 = time.perf_counter()
+                TIMINGS.append(t2 - t1)
                 for stat_name, value in update_stats.items():
                     batch_update_stats[stat_name].append(value)
 
 
     def create_policy(
         self, parsed_behavior_id: BehaviorIdentifiers, brain_parameters: BrainParameters
-    ) -> TFPolicy:
+    ) -> Policy:
+        if self.framework == "torch":
+            return self.create_torch_policy(parsed_behavior_id, brain_parameters)
+        else:
+            return self.create_tf_policy(parsed_behavior_id, brain_parameters)
+
+    def create_tf_policy(
+        self, parsed_behavior_id: BehaviorIdentifiers, brain_parameters: BrainParameters
+    ) -> NNPolicy:
+        :param parsed_behavior_id:
         :param brain_parameters: specifications for policy construction
         :return policy
         """
             self.is_training,
             self.load,
             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, brain_parameters: BrainParameters
+    ) -> 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,
+            brain_parameters,
+            self.trainer_parameters,
+            self.is_training,
+            self.load,
+            condition_sigma_on_obs=False,  # Faster training for PPO
+        )
-        self, parsed_behavior_id: BehaviorIdentifiers, policy: TFPolicy
+        self, parsed_behavior_id: BehaviorIdentifiers, policy: Policy
     ) -> None:
         """
         Adds policy to trainer.
                     self.__class__.__name__
                 )
             )
-        if not isinstance(policy, NNPolicy):
+        if not isinstance(policy, Policy):
-        self.optimizer = PPOOptimizer(self.policy, self.trainer_parameters)
+        if self.framework == "torch":
+            self.optimizer = TorchPPOOptimizer(  # type: ignore
+                self.policy, self.trainer_parameters  # type: ignore
+            )  # type: ignore
+        else:
+            self.optimizer = TFPPOOptimizer(  # type: ignore
+                self.policy, self.trainer_parameters  # type: ignore
+            )  # type: ignore
         for _reward_signal in self.optimizer.reward_signals.keys():
             self.collected_rewards[_reward_signal] = defaultdict(lambda: 0)
         # Needed to resume loads properly
-    def get_policy(self, name_behavior_id: str) -> TFPolicy:
+    def get_policy(self, name_behavior_id: str) -> Policy:
         """
         Gets policy from trainer associated with name_behavior_id
         :param name_behavior_id: full identifier of policy

ml-agents/mlagents/trainers/distributions_torch.py

+import torch
+from torch import nn
+from torch import distributions
+import numpy as np
+
+EPSILON = 1e-7  # Small value to avoid divide by zero
+
+
+class GaussianDistribution(nn.Module):
+    def __init__(self, hidden_size, num_outputs, conditional_sigma=False, **kwargs):
+        super(GaussianDistribution, self).__init__(**kwargs)
+        self.conditional_sigma = conditional_sigma
+        self.mu = nn.Linear(hidden_size, num_outputs)
+        nn.init.xavier_uniform_(self.mu.weight, gain=0.01)
+        if conditional_sigma:
+            self.log_sigma = nn.Linear(hidden_size, num_outputs)
+            nn.init.xavier_uniform(self.log_sigma.weight, gain=0.01)
+        else:
+            self.log_sigma = nn.Parameter(
+                torch.zeros(1, num_outputs, requires_grad=True)
+            )
+
+    def forward(self, inputs):
+        mu = self.mu(inputs)
+        if self.conditional_sigma:
+            log_sigma = self.log_sigma(inputs)
+        else:
+            log_sigma = self.log_sigma
+        return [distributions.normal.Normal(loc=mu, scale=torch.exp(log_sigma))]
+
+
+class MultiCategoricalDistribution(nn.Module):
+    def __init__(self, hidden_size, act_sizes):
+        super(MultiCategoricalDistribution, self).__init__()
+        self.act_sizes = act_sizes
+        self.branches = self.create_policy_branches(hidden_size)
+
+    def create_policy_branches(self, hidden_size):
+        branches = []
+        for size in self.act_sizes:
+            branch_output_layer = nn.Linear(hidden_size, size)
+            nn.init.xavier_uniform_(branch_output_layer.weight, gain=0.01)
+            branches.append(branch_output_layer)
+        return nn.ModuleList(branches)
+
+    def mask_branch(self, logits, mask):
+        raw_probs = torch.nn.functional.softmax(logits, dim=-1) * mask
+        normalized_probs = raw_probs / torch.sum(raw_probs, dim=-1).unsqueeze(-1)
+        normalized_logits = torch.log(normalized_probs + EPSILON)
+        return normalized_logits
+
+    def split_masks(self, masks):
+        split_masks = []
+        for idx, _ in enumerate(self.act_sizes):
+            start = int(np.sum(self.act_sizes[:idx]))
+            end = int(np.sum(self.act_sizes[: idx + 1]))
+            split_masks.append(masks[:, start:end])
+        return split_masks
+
+    def forward(self, inputs, masks):
+        # Todo - Support multiple branches in mask code
+        branch_distributions = []
+        masks = self.split_masks(masks)
+        for idx, branch in enumerate(self.branches):
+            logits = branch(inputs)
+            norm_logits = self.mask_branch(logits, masks[idx])
+            distribution = distributions.categorical.Categorical(logits=norm_logits)
+            branch_distributions.append(distribution)
+        return branch_distributions

ml-agents/mlagents/trainers/models_torch.py

+from enum import Enum
+from typing import Callable, NamedTuple
+
+import torch
+from torch import nn
+
+from mlagents.trainers.distributions_torch import (
+    GaussianDistribution,
+    MultiCategoricalDistribution,
+)
+from mlagents.trainers.exception import UnityTrainerException
+
+ActivationFunction = Callable[[torch.Tensor], torch.Tensor]
+EncoderFunction = Callable[
+    [torch.Tensor, int, ActivationFunction, int, str, bool], torch.Tensor
+]
+
+EPSILON = 1e-7
+
+
+class EncoderType(Enum):
+    SIMPLE = "simple"
+    NATURE_CNN = "nature_cnn"
+    RESNET = "resnet"
+
+
+class ActionType(Enum):
+    DISCRETE = "discrete"
+    CONTINUOUS = "continuous"
+
+    @staticmethod
+    def from_str(label):
+        if label in "continuous":
+            return ActionType.CONTINUOUS
+        elif label in "discrete":
+            return ActionType.DISCRETE
+        else:
+            raise NotImplementedError
+
+
+class LearningRateSchedule(Enum):
+    CONSTANT = "constant"
+    LINEAR = "linear"
+
+
+class NormalizerTensors(NamedTuple):
+    steps: torch.Tensor
+    running_mean: torch.Tensor
+    running_variance: torch.Tensor
+
+
+class NetworkBody(nn.Module):
+    def __init__(
+        self,
+        vector_sizes,
+        visual_sizes,
+        h_size,
+        normalize,
+        num_layers,
+        m_size,
+        vis_encode_type,
+        use_lstm,
+    ):
+        super(NetworkBody, self).__init__()
+        self.normalize = normalize
+        self.visual_encoders = []
+        self.vector_encoders = []
+        self.vector_normalizers = []
+        self.use_lstm = use_lstm
+        self.h_size = h_size
+        self.m_size = m_size
+
+        visual_encoder = ModelUtils.get_encoder_for_type(vis_encode_type)
+        for vector_size in vector_sizes:
+            if vector_size != 0:
+                self.vector_normalizers.append(Normalizer(vector_size))
+                self.vector_encoders.append(
+                    VectorEncoder(vector_size, h_size, num_layers)
+                )
+        for visual_size in visual_sizes:
+            self.visual_encoders.append(
+                visual_encoder(
+                    visual_size.height,
+                    visual_size.width,
+                    visual_size.num_channels,
+                    h_size,
+                )
+            )
+
+        self.vector_encoders = nn.ModuleList(self.vector_encoders)
+        self.visual_encoders = nn.ModuleList(self.visual_encoders)
+        if use_lstm:
+            self.lstm = nn.LSTM(h_size, m_size // 2, 1)
+
+    def update_normalization(self, vec_inputs):
+        if self.normalize:
+            for idx, vec_input in enumerate(vec_inputs):
+                self.vector_normalizers[idx].update(vec_input)
+
+    def forward(self, vec_inputs, vis_inputs, memories=None, sequence_length=1):
+        vec_embeds = []
+        for idx, encoder in enumerate(self.vector_encoders):
+            vec_input = vec_inputs[idx]
+            if self.normalize:
+                vec_input = self.vector_normalizers[idx](vec_input)
+            hidden = encoder(vec_input)
+            vec_embeds.append(hidden)
+
+        vis_embeds = []
+        for idx, encoder in enumerate(self.visual_encoders):
+            vis_input = vis_inputs[idx]
+            vis_input = vis_input.permute([0, 3, 1, 2])
+            hidden = encoder(vis_input)
+            vis_embeds.append(hidden)
+
+        if len(vec_embeds) > 0:
+            vec_embeds = torch.cat(vec_embeds)
+        if len(vis_embeds) > 0:
+            vis_embeds = torch.cat(vis_embeds)
+        if len(vec_embeds) > 0 and len(vis_embeds) > 0:
+            embedding = torch.cat([vec_embeds, vis_embeds])
+        elif len(vec_embeds) > 0:
+            embedding = vec_embeds
+        else:
+            embedding = vis_embeds
+
+        if self.use_lstm:
+            embedding = embedding.reshape([sequence_length, -1, self.h_size])
+            memories = torch.split(memories, self.m_size // 2, dim=-1)
+            embedding, memories = self.lstm(embedding, memories)
+            embedding = embedding.reshape([-1, self.m_size // 2])
+            memories = torch.cat(memories, dim=-1)
+        return embedding, memories
+
+
+class ActorCritic(nn.Module):
+    def __init__(
+        self,
+        h_size,
+        vector_sizes,
+        visual_sizes,
+        act_size,
+        normalize,
+        num_layers,
+        m_size,
+        vis_encode_type,
+        act_type,
+        use_lstm,
+        stream_names,
+        separate_critic,
+    ):
+        super(ActorCritic, self).__init__()
+        self.act_type = ActionType.from_str(act_type)
+        self.act_size = act_size
+        self.separate_critic = separate_critic
+        self.network_body = NetworkBody(
+            vector_sizes,
+            visual_sizes,
+            h_size,
+            normalize,
+            num_layers,
+            m_size,
+            vis_encode_type,
+            use_lstm,
+        )
+        if use_lstm:
+            embedding_size = m_size // 2
+        else:
+            embedding_size = h_size
+        if self.act_type == ActionType.CONTINUOUS:
+            self.distribution = GaussianDistribution(embedding_size, act_size[0])
+        else:
+            self.distribution = MultiCategoricalDistribution(embedding_size, act_size)
+        if separate_critic:
+            self.critic = Critic(
+                stream_names,
+                h_size,
+                vector_sizes,
+                visual_sizes,
+                normalize,
+                num_layers,
+                m_size,
+                vis_encode_type,
+            )
+        else:
+            self.stream_names = stream_names
+            self.value_heads = ValueHeads(stream_names, embedding_size)
+
+    def update_normalization(self, vector_obs):
+        self.network_body.update_normalization(vector_obs)
+        if self.separate_critic:
+            self.critic.network_body.update_normalization(vector_obs)
+
+    def critic_pass(self, vec_inputs, vis_inputs, memories=None):
+        if self.separate_critic:
+            return self.critic(vec_inputs, vis_inputs)
+        else:
+            embedding, _ = self.network_body(vec_inputs, vis_inputs, memories=memories)
+            return self.value_heads(embedding)
+
+    def sample_action(self, dists):
+        actions = []
+        for action_dist in dists:
+            action = action_dist.sample()
+            actions.append(action)
+        actions = torch.stack(actions, dim=-1)
+        return actions
+
+    def get_probs_and_entropy(self, actions, dists):
+        log_probs = []
+        entropies = []
+        for idx, action_dist in enumerate(dists):
+            action = actions[..., idx]
+            log_probs.append(action_dist.log_prob(action))
+            entropies.append(action_dist.entropy())
+        log_probs = torch.stack(log_probs, dim=-1)
+        entropies = torch.stack(entropies, dim=-1)
+        if self.act_type == ActionType.CONTINUOUS:
+            log_probs = log_probs.squeeze(-1)
+            entropies = entropies.squeeze(-1)
+        return log_probs, entropies
+
+    def get_dist_and_value(
+        self, vec_inputs, vis_inputs, masks=None, memories=None, sequence_length=1
+    ):
+        embedding, memories = self.network_body(
+            vec_inputs, vis_inputs, memories, sequence_length
+        )
+        if self.act_type == ActionType.CONTINUOUS:
+            dists = self.distribution(embedding)
+        else:
+            dists = self.distribution(embedding, masks=masks)
+        if self.separate_critic:
+            value_outputs = self.critic(vec_inputs, vis_inputs)
+        else:
+            value_outputs = self.value_heads(embedding)
+        return dists, value_outputs, memories
+
+    def forward(
+        self, vec_inputs, vis_inputs, masks=None, memories=None, sequence_length=1
+    ):
+        dists, value_outputs, memories = self.get_dist_and_value(
+            vec_inputs, vis_inputs, masks, memories, sequence_length
+        )
+        sampled_actions = self.sample_action(dists)
+        return sampled_actions, memories
+
+
+class Critic(nn.Module):
+    def __init__(
+        self,
+        stream_names,
+        h_size,
+        vector_sizes,
+        visual_sizes,
+        normalize,
+        num_layers,
+        m_size,
+        vis_encode_type,
+    ):
+        super(Critic, self).__init__()
+        self.network_body = NetworkBody(
+            vector_sizes,
+            visual_sizes,
+            h_size,
+            normalize,
+            num_layers,
+            m_size,
+            vis_encode_type,
+            False,
+        )
+        self.stream_names = stream_names
+        self.value_heads = ValueHeads(stream_names, h_size)
+
+    def forward(self, vec_inputs, vis_inputs):
+        embedding, _ = self.network_body(vec_inputs, vis_inputs)
+        return self.value_heads(embedding)
+
+
+class Normalizer(nn.Module):
+    def __init__(self, vec_obs_size, **kwargs):
+        super(Normalizer, self).__init__(**kwargs)
+        self.normalization_steps = torch.tensor(1)
+        self.running_mean = torch.zeros(vec_obs_size)
+        self.running_variance = torch.ones(vec_obs_size)
+
+    def forward(self, inputs):
+        normalized_state = torch.clamp(
+            (inputs - self.running_mean)
+            / torch.sqrt(self.running_variance / self.normalization_steps),
+            -5,
+            5,
+        )
+        return normalized_state
+
+    def update(self, vector_input):
+        steps_increment = vector_input.size()[0]
+        total_new_steps = self.normalization_steps + steps_increment
+
+        input_to_old_mean = vector_input - self.running_mean
+        new_mean = self.running_mean + (input_to_old_mean / total_new_steps).sum(0)
+
+        input_to_new_mean = vector_input - new_mean
+        new_variance = self.running_variance + (
+            input_to_new_mean * input_to_old_mean
+        ).sum(0)
+        self.running_mean = new_mean
+        self.running_variance = new_variance
+        self.normalization_steps = total_new_steps
+
+
+class ValueHeads(nn.Module):
+    def __init__(self, stream_names, input_size):
+        super(ValueHeads, self).__init__()
+        self.stream_names = stream_names
+        self.value_heads = {}
+
+        for name in stream_names:
+            value = nn.Linear(input_size, 1)
+            self.value_heads[name] = value
+        self.value_heads = nn.ModuleDict(self.value_heads)
+
+    def forward(self, hidden):
+        value_outputs = {}
+        for stream_name, _ in self.value_heads.items():
+            value_outputs[stream_name] = self.value_heads[stream_name](hidden).squeeze(
+                -1
+            )
+        return (
+            value_outputs,
+            torch.mean(torch.stack(list(value_outputs.values())), dim=0),
+        )
+
+
+class VectorEncoder(nn.Module):
+    def __init__(self, input_size, hidden_size, num_layers, **kwargs):
+        super(VectorEncoder, self).__init__(**kwargs)
+        self.layers = [nn.Linear(input_size, hidden_size)]
+        for _ in range(num_layers - 1):
+            self.layers.append(nn.Linear(hidden_size, hidden_size))
+            self.layers.append(nn.ReLU())
+        self.layers = nn.ModuleList(self.layers)
+
+    def forward(self, inputs):
+        x = inputs
+        for layer in self.layers:
+            x = layer(x)
+        return x
+
+
+def conv_output_shape(h_w, kernel_size=1, stride=1, pad=0, dilation=1):
+    from math import floor
+
+    if type(kernel_size) is not tuple:
+        kernel_size = (kernel_size, kernel_size)
+    h = floor(
+        ((h_w[0] + (2 * pad) - (dilation * (kernel_size[0] - 1)) - 1) / stride) + 1
+    )
+    w = floor(
+        ((h_w[1] + (2 * pad) - (dilation * (kernel_size[1] - 1)) - 1) / stride) + 1
+    )
+    return h, w
+
+
+class SimpleVisualEncoder(nn.Module):
+    def __init__(self, height, width, initial_channels, output_size):
+        super(SimpleVisualEncoder, self).__init__()
+        self.h_size = output_size
+        conv_1_hw = conv_output_shape((height, width), 8, 4)
+        conv_2_hw = conv_output_shape(conv_1_hw, 4, 2)
+        self.final_flat = conv_2_hw[0] * conv_2_hw[1] * 32
+
+        self.conv1 = nn.Conv2d(initial_channels, 16, [8, 8], [4, 4])
+        self.conv2 = nn.Conv2d(16, 32, [4, 4], [2, 2])
+        self.dense = nn.Linear(self.final_flat, self.h_size)
+
+    def forward(self, visual_obs):
+        conv_1 = torch.relu(self.conv1(visual_obs))
+        conv_2 = torch.relu(self.conv2(conv_1))
+        hidden = self.dense(conv_2.reshape([-1, self.final_flat]))
+        return hidden
+
+
+class NatureVisualEncoder(nn.Module):
+    def __init__(self, height, width, initial_channels, output_size):
+        super(NatureVisualEncoder, self).__init__()
+        self.h_size = output_size
+        conv_1_hw = conv_output_shape((height, width), 8, 4)
+        conv_2_hw = conv_output_shape(conv_1_hw, 4, 2)
+        conv_3_hw = conv_output_shape(conv_2_hw, 3, 1)
+        self.final_flat = conv_3_hw[0] * conv_3_hw[1] * 64
+
+        self.conv1 = nn.Conv2d(initial_channels, 32, [8, 8], [4, 4])
+        self.conv2 = nn.Conv2d(43, 64, [4, 4], [2, 2])
+        self.conv3 = nn.Conv2d(64, 64, [3, 3], [1, 1])
+        self.dense = nn.Linear(self.final_flat, self.h_size)
+
+    def forward(self, visual_obs):
+        conv_1 = torch.relu(self.conv1(visual_obs))
+        conv_2 = torch.relu(self.conv2(conv_1))
+        conv_3 = torch.relu(self.conv3(conv_2))
+        hidden = self.dense(conv_3.reshape([-1, self.final_flat]))
+        return hidden
+
+
+class GlobalSteps(nn.Module):
+    def __init__(self):
+        super(GlobalSteps, self).__init__()
+        self.global_step = torch.Tensor([0])
+
+    def increment(self, value):
+        self.global_step += value
+
+
+class LearningRate(nn.Module):
+    def __init__(self, lr):
+        # Todo: add learning rate decay
+        super(LearningRate, self).__init__()
+        self.learning_rate = torch.Tensor([lr])
+
+
+class ResNetVisualEncoder(nn.Module):
+    def __init__(self, initial_channels):
+        super(ResNetVisualEncoder, self).__init__()
+        n_channels = [16, 32, 32]  # channel for each stack
+        n_blocks = 2  # number of residual blocks
+        self.layers = []
+        for _, channel in enumerate(n_channels):
+            self.layers.append(nn.Conv2d(initial_channels, channel, [3, 3], [1, 1]))
+            self.layers.append(nn.MaxPool2d([3, 3], [2, 2]))
+            for _ in range(n_blocks):
+                self.layers.append(self.make_block(channel))
+        self.layers.append(nn.ReLU())
+
+    @staticmethod
+    def make_block(channel):
+        block_layers = [
+            nn.ReLU(),
+            nn.Conv2d(channel, channel, [3, 3], [1, 1]),
+            nn.ReLU(),
+            nn.Conv2d(channel, channel, [3, 3], [1, 1]),
+        ]
+        return block_layers
+
+    @staticmethod
+    def forward_block(input_hidden, block_layers):
+        hidden = input_hidden
+        for layer in block_layers:
+            hidden = layer(hidden)
+        return hidden + input_hidden
+
+    def forward(self, visual_obs):
+        hidden = visual_obs
+        for layer in self.layers:
+            if layer is nn.Module:
+                hidden = layer(hidden)
+            elif layer is list:
+                hidden = self.forward_block(hidden, layer)
+        return hidden.flatten()
+
+
+class ModelUtils:
+    # Minimum supported side for each encoder type. If refactoring an encoder, please
+    # adjust these also.
+    MIN_RESOLUTION_FOR_ENCODER = {
+        EncoderType.SIMPLE: 20,
+        EncoderType.NATURE_CNN: 36,
+        EncoderType.RESNET: 15,
+    }
+
+    @staticmethod
+    def swish(input_activation: torch.Tensor) -> torch.Tensor:
+        """Swish activation function. For more info: https://arxiv.org/abs/1710.05941"""
+        return torch.mul(input_activation, torch.sigmoid(input_activation))
+
+    @staticmethod
+    def get_encoder_for_type(encoder_type: EncoderType) -> nn.Module:
+        ENCODER_FUNCTION_BY_TYPE = {
+            EncoderType.SIMPLE: SimpleVisualEncoder,
+            EncoderType.NATURE_CNN: NatureVisualEncoder,
+            EncoderType.RESNET: ResNetVisualEncoder,
+        }
+        return ENCODER_FUNCTION_BY_TYPE.get(encoder_type)
+
+    @staticmethod
+    def _check_resolution_for_encoder(
+        vis_in: torch.Tensor, vis_encoder_type: EncoderType
+    ) -> None:
+        min_res = ModelUtils.MIN_RESOLUTION_FOR_ENCODER[vis_encoder_type]
+        height = vis_in.shape[1]
+        width = vis_in.shape[2]
+        if height < min_res or width < min_res:
+            raise UnityTrainerException(
+                f"Visual observation resolution ({width}x{height}) is too small for"
+                f"the provided EncoderType ({vis_encoder_type.value}). The min dimension is {min_res}"
+            )

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

+from typing import Dict, Any, Optional, Tuple, List
+import torch
+import numpy as np
+from mlagents_envs.base_env import DecisionSteps
+
+from mlagents.trainers.buffer import AgentBuffer
+from mlagents.trainers.components.bc.module import BCModule
+from mlagents.trainers.components.reward_signals.extrinsic.signal import (
+    ExtrinsicRewardSignal,
+)
+from mlagents.trainers.policy.torch_policy import TorchPolicy
+from mlagents.trainers.optimizer import Optimizer
+from mlagents.trainers.trajectory import SplitObservations
+
+
+class TorchOptimizer(Optimizer):  # pylint: disable=W0223
+    def __init__(self, policy: TorchPolicy, trainer_params: Dict[str, Any]):
+        super(TorchOptimizer, self).__init__()
+        self.policy = policy
+        self.trainer_params = trainer_params
+        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_params["reward_signals"])
+
+    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.
+        """
+        extrinsic_signal = ExtrinsicRewardSignal(
+            self.policy, **reward_signal_configs["extrinsic"]
+        )
+        self.reward_signals = {"extrinsic": extrinsic_signal}
+        # Create reward signals
+        # for reward_signal, config in reward_signal_configs.items():
+        #    self.reward_signals[reward_signal] = create_reward_signal(
+        #        self.policy, reward_signal, config
+        #    )
+        #    self.update_dict.update(self.reward_signals[reward_signal].update_dict)
+
+    def get_value_estimates(
+        self, decision_requests: DecisionSteps, idx: int, done: bool
+    ) -> Dict[str, float]:
+        """
+        Generates value estimates for bootstrapping.
+        :param decision_requests:
+        :param idx: Index in BrainInfo of agent.
+        :param done: Whether or not this is the last element of the episode,
+        in which case the value estimate will be 0.
+        :return: The value estimate dictionary with key being the name of the reward signal
+        and the value the corresponding value estimate.
+        """
+        vec_vis_obs = SplitObservations.from_observations(decision_requests.obs)
+
+        value_estimates, mean_value = self.policy.actor_critic.critic_pass(
+            np.expand_dims(vec_vis_obs.vector_observations[idx], 0),
+            np.expand_dims(vec_vis_obs.visual_observations[idx], 0),
+        )
+
+        value_estimates = {k: float(v) for k, v in value_estimates.items()}
+
+        # If we're done, reassign all of the value estimates that need terminal states.
+        if done:
+            for k in value_estimates:
+                if self.reward_signals[k].use_terminal_states:
+                    value_estimates[k] = 0.0
+
+        return value_estimates
+
+    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 = [torch.as_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 = torch.as_tensor(batch["visual_obs%d" % idx])
+                visual_obs.append(visual_ob)
+        else:
+            visual_obs = []
+
+        memory = torch.zeros([1, len(vector_obs[0]), self.policy.m_size])
+
+        next_obs = np.concatenate(next_obs, axis=-1)
+        next_obs = [torch.as_tensor(next_obs).unsqueeze(0)]
+        next_memory = torch.zeros([1, 1, self.policy.m_size])
+
+        value_estimates, mean_value = self.policy.actor_critic.critic_pass(
+            vector_obs, visual_obs, memory
+        )
+
+        next_value_estimate, next_value = self.policy.actor_critic.critic_pass(
+            next_obs, next_obs, next_memory
+        )
+
+        for name, estimate in value_estimates.items():
+            value_estimates[name] = estimate.detach().numpy()
+            next_value_estimate[name] = next_value_estimate[name].detach().numpy()
+
+        if done:
+            for k in next_value_estimate:
+                if self.reward_signals[k].use_terminal_states:
+                    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
+import numpy as np
+import torch
+import os
+from torch import onnx
+from mlagents.trainers.action_info import ActionInfo
+from mlagents.trainers.brain_conversion_utils import get_global_agent_id
+
+from mlagents.trainers.policy import Policy
+from mlagents_envs.base_env import DecisionSteps
+from mlagents.tf_utils import tf
+from mlagents_envs.timers import timed
+
+from mlagents.trainers.policy.policy import UnityPolicyException
+from mlagents.trainers.trajectory import SplitObservations
+from mlagents.trainers.brain import BrainParameters
+from mlagents.trainers.models_torch import EncoderType, ActorCritic
+
+EPSILON = 1e-7  # Small value to avoid divide by zero
+
+torch.set_num_threads(1)
+
+
+class TorchPolicy(Policy):
+    def __init__(
+        self,
+        seed: int,
+        brain: BrainParameters,
+        trainer_params: Dict[str, Any],
+        load: bool,
+        tanh_squash: bool = False,
+        reparameterize: bool = False,
+        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_params: 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(TorchPolicy, self).__init__(brain, seed, trainer_params)
+        self.grads = None
+        num_layers = trainer_params["num_layers"]
+        self.h_size = trainer_params["hidden_units"]
+        self.seed = seed
+        self.brain = brain
+        self.global_step = 0
+        self.m_size = 0
+
+        self.act_size = brain.vector_action_space_size
+        self.act_type = brain.vector_action_space_type
+        self.sequence_length = 1
+        if self.use_recurrent:
+            self.m_size = trainer_params["memory_size"]
+            self.sequence_length = trainer_params["sequence_length"]
+            if self.m_size == 0:
+                raise UnityPolicyException(
+                    "The memory size for brain {0} is 0 even "
+                    "though the trainer uses recurrent.".format(brain.brain_name)
+                )
+            elif self.m_size % 2 != 0:
+                raise UnityPolicyException(
+                    "The memory size for brain {0} is {1} "
+                    "but it must be divisible by 2.".format(
+                        brain.brain_name, self.m_size
+                    )
+                )
+
+        if num_layers < 1:
+            num_layers = 1
+        self.num_layers = num_layers
+        self.vis_encode_type = EncoderType(
+            trainer_params.get("vis_encode_type", "simple")
+        )
+        self.tanh_squash = tanh_squash
+        self.reparameterize = reparameterize
+        self.condition_sigma_on_obs = condition_sigma_on_obs
+
+        # Non-exposed parameters; these aren't exposed because they don't have a
+        # good explanation and usually shouldn't be touched.
+        self.log_std_min = -20
+        self.log_std_max = 2
+
+        self.inference_dict: Dict[str, tf.Tensor] = {}
+        self.update_dict: Dict[str, tf.Tensor] = {}
+        # TF defaults to 32-bit, so we use the same here.
+        torch.set_default_tensor_type(torch.FloatTensor)
+
+        reward_signal_configs = trainer_params["reward_signals"]
+        self.stats_name_to_update_name = {
+            "Losses/Value Loss": "value_loss",
+            "Losses/Policy Loss": "policy_loss",
+        }
+
+        self.actor_critic = ActorCritic(
+            h_size=int(trainer_params["hidden_units"]),
+            act_type=self.act_type,
+            vector_sizes=[brain.vector_observation_space_size],
+            act_size=brain.vector_action_space_size,
+            normalize=trainer_params["normalize"],
+            num_layers=int(trainer_params["num_layers"]),
+            m_size=trainer_params["memory_size"],
+            use_lstm=self.use_recurrent,
+            visual_sizes=brain.camera_resolutions,
+            vis_encode_type=EncoderType(
+                trainer_params.get("vis_encode_type", "simple")
+            ),
+            stream_names=list(reward_signal_configs.keys()),
+            separate_critic=self.use_continuous_act,
+        )
+
+    def split_decision_step(self, decision_requests):
+        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.brain.vector_action_space_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.vector_observations, vec_vis_obs.visual_observations, 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, vis_obs, masks=None, memories=None, seq_len=1):
+        dists, (
+            value_heads,
+            mean_value,
+        ), memories = self.actor_critic.get_dist_and_value(
+            vec_obs, vis_obs, masks, memories, seq_len
+        )
+
+        actions = self.actor_critic.sample_action(dists)
+        log_probs, entropies = self.actor_critic.get_probs_and_entropy(actions, dists)
+        if self.act_type == "continuous":
+            actions.squeeze_(-1)
+
+        return actions, log_probs, entropies, value_heads, memories
+
+    def evaluate_actions(
+        self, vec_obs, vis_obs, actions, masks=None, memories=None, seq_len=1
+    ):
+        dists, (value_heads, mean_value), _ = self.actor_critic.get_dist_and_value(
+            vec_obs, vis_obs, masks, memories, seq_len
+        )
+
+        log_probs, entropies = self.actor_critic.get_probs_and_entropy(actions, 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_obs, vis_obs, masks = self.split_decision_step(decision_requests)
+        vec_obs = [torch.as_tensor(vec_obs)]
+        vis_obs = [torch.as_tensor(vis_ob) for vis_ob in vis_obs]
+        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().numpy()
+        run_out["pre_action"] = action.detach().numpy()
+        # Todo - make pre_action difference
+        run_out["log_probs"] = log_probs.detach().numpy()
+        run_out["entropy"] = entropy.detach().numpy()
+        run_out["value_heads"] = {
+            name: t.detach().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["memories"] = memories.detach().numpy()
+        self.actor_critic.update_normalization(vec_obs)
+        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),
+        )
+
+    def save_model(self, step=0):
+        """
+        Saves the model
+        :param step: The number of steps the model was trained for
+        """
+        if not os.path.exists(self.model_path):
+            os.makedirs(self.model_path)
+        save_path = self.model_path + "/model-" + str(step) + ".pt"
+        torch.save(self.actor_critic.state_dict(), save_path)
+
+    def load_model(self, step=0):
+        load_path = self.model_path + "/model-" + str(step) + ".pt"
+        self.actor_critic.load_state_dict(torch.load(load_path))
+
+    def export_model(self, step=0):
+        fake_vec_obs = [torch.zeros([1] + [self.vec_obs_size])]
+        fake_vis_obs = [
+            torch.zeros(
+                [1] + [camera_res.height, camera_res.width, camera_res.num_channels]
+            )
+            for camera_res in self.brain.camera_resolutions
+        ]
+        if self.use_continuous_act:
+            fake_masks = None
+        else:
+            fake_masks = torch.ones([1] + [int(np.sum(self.act_size))])
+        fake_memories = torch.zeros([1] + [self.m_size])
+        export_path = self.model_path + "/model-" + str(step) + ".onnx"
+        output_names = ["action", "value_estimates", "memories"]
+        onnx.export(
+            self.actor_critic,
+            (fake_vec_obs, fake_vis_obs, fake_masks, fake_memories, 1),
+            export_path,
+            verbose=True,
+            output_names=output_names,
+        )
+
+    @property
+    def vis_obs_size(self):
+        return self.brain.number_visual_observations
+
+    @property
+    def vec_obs_size(self):
+        return self.brain.vector_observation_space_size
+
+    @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.
+        """
+        step = self.global_step
+        return step
+
+    def increment_step(self, n_steps):
+        """
+        Increments model step.
+        """
+        self.global_step += n_steps
+        return self.get_current_step()

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

+from typing import Any, Dict
+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
+
+
+class TorchPPOOptimizer(TorchOptimizer):
+    def __init__(self, policy: TorchPolicy, trainer_params: Dict[str, Any]):
+        """
+        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(TorchPPOOptimizer, self).__init__(policy, trainer_params)
+        params = list(self.policy.actor_critic.parameters())
+
+        self.optimizer = torch.optim.Adam(
+            params, lr=self.trainer_params["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, old_values, returns):
+        """
+        Creates training-specific Tensorflow ops for PPO models.
+        :param returns:
+        :param old_values:
+        :param values:
+        """
+
+        decay_epsilon = self.trainer_params["epsilon"]
+
+        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, -decay_epsilon, decay_epsilon
+            )
+            v_opt_a = (returns_tensor - head) ** 2
+            v_opt_b = (returns_tensor - clipped_value_estimate) ** 2
+            value_loss = torch.mean(torch.max(v_opt_a, v_opt_b))
+            value_losses.append(value_loss)
+        value_loss = torch.mean(torch.stack(value_losses))
+        return value_loss
+
+    def ppo_policy_loss(self, advantages, log_probs, old_log_probs, masks):
+        """
+        Creates training-specific Tensorflow ops for PPO models.
+        :param masks:
+        :param advantages:
+        :param log_probs: Current policy probabilities
+        :param old_log_probs: Past policy probabilities
+        """
+        advantage = advantages.unsqueeze(-1)
+
+        decay_epsilon = self.trainer_params["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 = -torch.mean(torch.min(p_opt_a, p_opt_b))
+        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.
+        """
+        returns = {}
+        old_values = {}
+        for name in self.reward_signals:
+            old_values[name] = torch.as_tensor(batch["{}_value_estimates".format(name)])
+            returns[name] = torch.as_tensor(batch["{}_returns".format(name)])
+
+        vec_obs = [torch.as_tensor(batch["vector_obs"])]
+        act_masks = torch.as_tensor(batch["action_mask"])
+        if self.policy.use_continuous_act:
+            actions = torch.as_tensor(batch["actions"]).unsqueeze(-1)
+        else:
+            actions = torch.as_tensor(batch["actions"])
+
+        memories = [
+            torch.as_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 = torch.as_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,
+        )
+        value_loss = self.ppo_value_loss(values, old_values, returns)
+        policy_loss = self.ppo_policy_loss(
+            torch.as_tensor(batch["advantages"]),
+            log_probs,
+            torch.as_tensor(batch["action_probs"]),
+            torch.as_tensor(batch["masks"], dtype=torch.int32),
+        )
+        loss = (
+            policy_loss
+            + 0.5 * value_loss
+            - self.trainer_params["beta"] * torch.mean(entropy)
+        )
+        self.optimizer.zero_grad()
+        loss.backward()
+
+        self.optimizer.step()
+        update_stats = {
+            "Losses/Policy Loss": abs(policy_loss.detach().numpy()),
+            "Losses/Value Loss": value_loss.detach().numpy(),
+        }
+
+        return update_stats

test_env_perf.py

+from mlagents.trainers.tests.test_simple_rl import (
+    _check_environment_trains,
+    PPO_CONFIG,
+    generate_config,
+)
+from mlagents.trainers.tests.simple_test_envs import SimpleEnvironment
+from mlagents.trainers.ppo.trainer import TIMINGS
+import matplotlib.pyplot as plt
+import numpy as np
+
+BRAIN_NAME = "1D"
+
+if __name__ == "__main__":
+    env = SimpleEnvironment([BRAIN_NAME], use_discrete=False)
+    config = generate_config(
+        PPO_CONFIG,
+        override_vals={"batch_size": 256, "max_steps": 20000, "buffer_size": 1024},
+    )
+    try:
+        _check_environment_trains(env, config)
+    except Exception:
+        pass
+    print(f"Mean update time {np.mean(TIMINGS)}")
+    plt.plot(TIMINGS)
+    plt.ylim((0, 0.006))
+    plt.title("PyTorch w/ 3DBall Running, batch size 256, 32 hidden units, 1 layer")
+    plt.ylabel("Update Time (s)")
+    plt.ylabel("Update #")
+    plt.show()