From d523c8cc0ac0214963eee5323f717c67c647e540 Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Fri, 7 Aug 2020 16:14:21 -0700
Subject: [PATCH 01/18] Running LSTM for SAC

---
 .../mlagents/trainers/sac/optimizer_torch.py  | 100 +++++++++++++++---
 1 file changed, 85 insertions(+), 15 deletions(-)

diff --git a/ml-agents/mlagents/trainers/sac/optimizer_torch.py b/ml-agents/mlagents/trainers/sac/optimizer_torch.py
index 9c3ced80a7..84df073fc0 100644
--- a/ml-agents/mlagents/trainers/sac/optimizer_torch.py
+++ b/ml-agents/mlagents/trainers/sac/optimizer_torch.py
@@ -1,7 +1,8 @@
 import numpy as np
-from typing import Dict, List, Mapping, cast, Tuple
+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
@@ -56,10 +57,24 @@ def forward(
             self,
             vec_inputs: List[torch.Tensor],
             vis_inputs: List[torch.Tensor],
-            actions: torch.Tensor = None,
+            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)
-            q2_out, _ = self.q2_network(vec_inputs, vis_inputs, actions=actions)
+            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):
@@ -87,17 +102,28 @@ def __init__(self, policy: TorchPolicy, trainer_params: TrainerSettings):
             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,
-            policy_network_settings,
+            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,
-            policy_network_settings,
+            value_network_settings,
         )
         self.soft_update(self.policy.actor_critic.critic, self.target_network, 1.0)
 
@@ -232,7 +258,6 @@ def sac_value_loss(
                     v_backup = min_policy_qs[name] - torch.sum(
                         _ent_coef * log_probs, dim=1
                     )
-                # print(log_probs, v_backup, _ent_coef, loss_masks)
                 value_loss = 0.5 * torch.mean(
                     loss_masks * torch.nn.functional.mse_loss(values[name], v_backup)
                 )
@@ -369,12 +394,30 @@ def update(self, batch: AgentBuffer, num_sequences: int) -> Dict[str, float]:
         else:
             actions = ModelUtils.list_to_tensor(batch["actions"], dtype=torch.long)
 
-        memories = [
+        memories_list = [
             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)
+        # 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(
+            (memories.shape[0], memories.shape[1], memories.shape[2] // 2)
+        )
+
         vis_obs: List[torch.Tensor] = []
         next_vis_obs: List[torch.Tensor] = []
         if self.policy.use_vis_obs:
@@ -415,18 +458,45 @@ def update(self, batch: AgentBuffer, num_sequences: int) -> Dict[str, float]:
         )
         if self.policy.use_continuous_act:
             squeezed_actions = actions.squeeze(-1)
-            q1p_out, q2p_out = self.value_network(vec_obs, vis_obs, sampled_actions)
-            q1_out, q2_out = self.value_network(vec_obs, vis_obs, squeezed_actions)
+            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)
-            q1_out, q2_out = self.value_network(vec_obs, vis_obs)
+                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)
+            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.int32)
         use_discrete = not self.policy.use_continuous_act
         dones = ModelUtils.list_to_tensor(batch["done"])

From f2873b296fe60fefb86545405c6076e26d10af5e Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Mon, 10 Aug 2020 10:38:19 -0700
Subject: [PATCH 02/18] Use correct half of memories

---
 ml-agents/mlagents/trainers/sac/optimizer_torch.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ml-agents/mlagents/trainers/sac/optimizer_torch.py b/ml-agents/mlagents/trainers/sac/optimizer_torch.py
index 84df073fc0..3ba92e6c1b 100644
--- a/ml-agents/mlagents/trainers/sac/optimizer_torch.py
+++ b/ml-agents/mlagents/trainers/sac/optimizer_torch.py
@@ -402,7 +402,7 @@ def update(self, batch: AgentBuffer, num_sequences: int) -> Dict[str, float]:
         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]
+                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)
         ]

From b97b1e535af6006bd864249693c78460bc721877 Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Mon, 10 Aug 2020 17:40:47 -0700
Subject: [PATCH 03/18] Fix policy memory storinig

---
 ml-agents/mlagents/trainers/policy/torch_policy.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ml-agents/mlagents/trainers/policy/torch_policy.py b/ml-agents/mlagents/trainers/policy/torch_policy.py
index d6dd822646..9d935d4676 100644
--- a/ml-agents/mlagents/trainers/policy/torch_policy.py
+++ b/ml-agents/mlagents/trainers/policy/torch_policy.py
@@ -186,7 +186,7 @@ def evaluate(
         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().cpu().numpy()
+            run_out["memory_out"] = memories.detach().cpu().numpy().squeeze(0)
         return run_out
 
     def get_action(

From cd509ddbbdc3e4025040105f9d521034a37fdf3a Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Mon, 10 Aug 2020 18:12:27 -0700
Subject: [PATCH 04/18] Fix SeparateActorCritic and add test

---
 ml-agents/mlagents/trainers/tests/torch/test_networks.py | 6 +++++-
 ml-agents/mlagents/trainers/torch/networks.py            | 2 +-
 2 files changed, 6 insertions(+), 2 deletions(-)

diff --git a/ml-agents/mlagents/trainers/tests/torch/test_networks.py b/ml-agents/mlagents/trainers/tests/torch/test_networks.py
index ff5209b676..19030aeafc 100644
--- a/ml-agents/mlagents/trainers/tests/torch/test_networks.py
+++ b/ml-agents/mlagents/trainers/tests/torch/test_networks.py
@@ -203,7 +203,11 @@ def test_actor_critic(ac_type, lstm):
             assert value_out[stream].shape == (1,)
 
     # Test get_dist_and_value
-    dists, value_out, _ = actor.get_dist_and_value([sample_obs], [], memories=memories)
+    dists, value_out, mem_out = actor.get_dist_and_value(
+        [sample_obs], [], memories=memories
+    )
+    if mem_out is not None:
+        assert mem_out.shape == memories.shape
     for dist in dists:
         assert isinstance(dist, GaussianDistInstance)
     for stream in stream_names:
diff --git a/ml-agents/mlagents/trainers/torch/networks.py b/ml-agents/mlagents/trainers/torch/networks.py
index 15f0d92e2d..aff61ce0f6 100644
--- a/ml-agents/mlagents/trainers/torch/networks.py
+++ b/ml-agents/mlagents/trainers/torch/networks.py
@@ -463,7 +463,7 @@ def get_dist_and_value(
             vec_inputs, vis_inputs, memories=critic_mem, sequence_length=sequence_length
         )
         if self.use_lstm:
-            mem_out = torch.cat([actor_mem_outs, critic_mem_outs], dim=1)
+            mem_out = torch.cat([actor_mem_outs, critic_mem_outs], dim=-1)
         else:
             mem_out = None
         return dists, value_outputs, mem_out

From 07bb4c0818887dab74e75236f51f6ca1533bde80 Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Tue, 11 Aug 2020 19:15:30 -0700
Subject: [PATCH 05/18] Use loss masks in PPO.

---
 .../mlagents/trainers/ppo/optimizer_torch.py  | 44 +++++++++++++------
 1 file changed, 31 insertions(+), 13 deletions(-)

diff --git a/ml-agents/mlagents/trainers/ppo/optimizer_torch.py b/ml-agents/mlagents/trainers/ppo/optimizer_torch.py
index 9bbb7b51c8..13fdc12e01 100644
--- a/ml-agents/mlagents/trainers/ppo/optimizer_torch.py
+++ b/ml-agents/mlagents/trainers/ppo/optimizer_torch.py
@@ -61,12 +61,15 @@ def ppo_value_loss(
         old_values: Dict[str, torch.Tensor],
         returns: Dict[str, torch.Tensor],
         epsilon: float,
+        loss_masks: torch.Tensor,
     ) -> torch.Tensor:
         """
-        Creates training-specific Tensorflow ops for PPO models.
-        :param returns:
-        :param old_values:
-        :param values:
+        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():
@@ -77,18 +80,25 @@ def ppo_value_loss(
             )
             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))
+            masked_loss = torch.max(v_opt_a, v_opt_b) * loss_masks
+            value_loss = torch.mean(masked_loss)
             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):
+    def ppo_policy_loss(
+        self,
+        advantages: torch.Tensor,
+        log_probs: torch.Tensor,
+        old_log_probs: torch.Tensor,
+        loss_masks: torch.Tensor,
+    ) -> torch.Tensor:
         """
-        Creates training-specific Tensorflow ops for PPO models.
-        :param masks:
-        :param advantages:
+        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)
 
@@ -99,7 +109,8 @@ def ppo_policy_loss(self, advantages, log_probs, old_log_probs, masks):
         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))
+        masked_loss = torch.min(p_opt_a, p_opt_b) * loss_masks
+        policy_loss = -torch.mean(masked_loss)
         return policy_loss
 
     @timed
@@ -153,14 +164,21 @@ def update(self, batch: AgentBuffer, num_sequences: int) -> Dict[str, float]:
             memories=memories,
             seq_len=self.policy.sequence_length,
         )
-        value_loss = self.ppo_value_loss(values, old_values, returns, decay_eps)
+        loss_masks = ModelUtils.list_to_tensor(batch["masks"], dtype=torch.int32)
+        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"]),
-            ModelUtils.list_to_tensor(batch["masks"], dtype=torch.int32),
+            loss_masks,
+        )
+        loss = (
+            policy_loss
+            + 0.5 * value_loss
+            - decay_bet * torch.mean(entropy * loss_masks)
         )
-        loss = policy_loss + 0.5 * value_loss - decay_bet * torch.mean(entropy)
 
         # Set optimizer learning rate
         ModelUtils.update_learning_rate(self.optimizer, decay_lr)

From 0a3c795cee4197003d5c7f5e96b03925e14f9896 Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Tue, 11 Aug 2020 19:46:34 -0700
Subject: [PATCH 06/18] Proper shape of masks

---
 ml-agents/mlagents/trainers/ppo/optimizer_torch.py | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/ml-agents/mlagents/trainers/ppo/optimizer_torch.py b/ml-agents/mlagents/trainers/ppo/optimizer_torch.py
index 13fdc12e01..1232429f36 100644
--- a/ml-agents/mlagents/trainers/ppo/optimizer_torch.py
+++ b/ml-agents/mlagents/trainers/ppo/optimizer_torch.py
@@ -109,7 +109,7 @@ def ppo_policy_loss(
         p_opt_b = (
             torch.clamp(r_theta, 1.0 - decay_epsilon, 1.0 + decay_epsilon) * advantage
         )
-        masked_loss = torch.min(p_opt_a, p_opt_b) * loss_masks
+        masked_loss = torch.min(p_opt_a, p_opt_b).flatten() * loss_masks
         policy_loss = -torch.mean(masked_loss)
         return policy_loss
 
@@ -164,7 +164,7 @@ def update(self, batch: AgentBuffer, num_sequences: int) -> Dict[str, float]:
             memories=memories,
             seq_len=self.policy.sequence_length,
         )
-        loss_masks = ModelUtils.list_to_tensor(batch["masks"], dtype=torch.int32)
+        loss_masks = ModelUtils.list_to_tensor(batch["masks"], dtype=torch.float32)
         value_loss = self.ppo_value_loss(
             values, old_values, returns, decay_eps, loss_masks
         )
@@ -177,7 +177,7 @@ def update(self, batch: AgentBuffer, num_sequences: int) -> Dict[str, float]:
         loss = (
             policy_loss
             + 0.5 * value_loss
-            - decay_bet * torch.mean(entropy * loss_masks)
+            - decay_bet * torch.mean(entropy.flatten() * loss_masks)
         )
 
         # Set optimizer learning rate

From 2337d15c09e2cdd91ec5d0c08eb63a2ba5d5e297 Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Tue, 11 Aug 2020 20:04:48 -0700
Subject: [PATCH 07/18] Proper mask mean for PPO

---
 ml-agents/mlagents/trainers/ppo/optimizer_torch.py | 14 +++++++-------
 ml-agents/mlagents/trainers/torch/utils.py         | 10 ++++++++++
 2 files changed, 17 insertions(+), 7 deletions(-)

diff --git a/ml-agents/mlagents/trainers/ppo/optimizer_torch.py b/ml-agents/mlagents/trainers/ppo/optimizer_torch.py
index 1232429f36..b330bce0fb 100644
--- a/ml-agents/mlagents/trainers/ppo/optimizer_torch.py
+++ b/ml-agents/mlagents/trainers/ppo/optimizer_torch.py
@@ -80,8 +80,7 @@ def ppo_value_loss(
             )
             v_opt_a = (returns_tensor - head) ** 2
             v_opt_b = (returns_tensor - clipped_value_estimate) ** 2
-            masked_loss = torch.max(v_opt_a, v_opt_b) * loss_masks
-            value_loss = torch.mean(masked_loss)
+            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
@@ -109,8 +108,9 @@ def ppo_policy_loss(
         p_opt_b = (
             torch.clamp(r_theta, 1.0 - decay_epsilon, 1.0 + decay_epsilon) * advantage
         )
-        masked_loss = torch.min(p_opt_a, p_opt_b).flatten() * loss_masks
-        policy_loss = -torch.mean(masked_loss)
+        policy_loss = -1 * ModelUtils.masked_mean(
+            torch.min(p_opt_a, p_opt_b).flatten(), loss_masks
+        )
         return policy_loss
 
     @timed
@@ -138,7 +138,7 @@ def update(self, batch: AgentBuffer, num_sequences: int) -> Dict[str, float]:
         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)
+            actions = ModelUtils.list_to_tensor(batch["actions"], dtype=torch.bool)
 
         memories = [
             ModelUtils.list_to_tensor(batch["memory"][i])
@@ -164,7 +164,7 @@ def update(self, batch: AgentBuffer, num_sequences: int) -> Dict[str, float]:
             memories=memories,
             seq_len=self.policy.sequence_length,
         )
-        loss_masks = ModelUtils.list_to_tensor(batch["masks"], dtype=torch.float32)
+        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
         )
@@ -177,7 +177,7 @@ def update(self, batch: AgentBuffer, num_sequences: int) -> Dict[str, float]:
         loss = (
             policy_loss
             + 0.5 * value_loss
-            - decay_bet * torch.mean(entropy.flatten() * loss_masks)
+            - decay_bet * ModelUtils.masked_mean(entropy.flatten(), loss_masks)
         )
 
         # Set optimizer learning rate
diff --git a/ml-agents/mlagents/trainers/torch/utils.py b/ml-agents/mlagents/trainers/torch/utils.py
index baa99c98ef..ba6b7d57a0 100644
--- a/ml-agents/mlagents/trainers/torch/utils.py
+++ b/ml-agents/mlagents/trainers/torch/utils.py
@@ -284,3 +284,13 @@ def get_probs_and_entropy(
         else:
             all_probs = torch.cat(all_probs_list, dim=-1)
         return log_probs, entropies, all_probs
+
+    @staticmethod
+    def masked_mean(tensor: torch.Tensor, masks: torch.Tensor) -> torch.Tensor:
+        """
+        Returns the mean of the tensor but ignoring the values specified by masks.
+        Used for masking out loss functions.
+        :param tensor: Tensor which needs mean computation.
+        :param masks: Boolean tensor of masks with same dimension as tensor.
+        """
+        return (tensor * masks).sum() / masks.float().sum()

From 1f69102f45dd5bfd7a0c8e63fdee701af7a25632 Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Tue, 11 Aug 2020 20:17:47 -0700
Subject: [PATCH 08/18] Fix dtype for actions

---
 ml-agents/mlagents/trainers/ppo/optimizer_torch.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ml-agents/mlagents/trainers/ppo/optimizer_torch.py b/ml-agents/mlagents/trainers/ppo/optimizer_torch.py
index b330bce0fb..e162166481 100644
--- a/ml-agents/mlagents/trainers/ppo/optimizer_torch.py
+++ b/ml-agents/mlagents/trainers/ppo/optimizer_torch.py
@@ -138,7 +138,7 @@ def update(self, batch: AgentBuffer, num_sequences: int) -> Dict[str, float]:
         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.bool)
+            actions = ModelUtils.list_to_tensor(batch["actions"], dtype=torch.long)
 
         memories = [
             ModelUtils.list_to_tensor(batch["memory"][i])

From c0a77f76b6f9b1c6a3187b9c56e49a7228605e7b Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Wed, 12 Aug 2020 14:13:17 -0700
Subject: [PATCH 09/18] Proper initialization and SAC masking

---
 .../mlagents/trainers/sac/optimizer_torch.py  | 27 +++++++++----------
 ml-agents/mlagents/trainers/torch/layers.py   | 24 +++++++++++++++++
 ml-agents/mlagents/trainers/torch/networks.py | 13 +++++----
 3 files changed, 43 insertions(+), 21 deletions(-)

diff --git a/ml-agents/mlagents/trainers/sac/optimizer_torch.py b/ml-agents/mlagents/trainers/sac/optimizer_torch.py
index 3ba92e6c1b..6d110f99f5 100644
--- a/ml-agents/mlagents/trainers/sac/optimizer_torch.py
+++ b/ml-agents/mlagents/trainers/sac/optimizer_torch.py
@@ -194,11 +194,11 @@ def sac_q_loss(
                     * self.gammas[i]
                     * target_values[name]
                 )
-            _q1_loss = 0.5 * torch.mean(
-                loss_masks * torch.nn.functional.mse_loss(q_backup, q1_stream)
+            _q1_loss = 0.5 * ModelUtils.masked_mean(
+                torch.nn.functional.mse_loss(q_backup, q1_stream), loss_masks
             )
-            _q2_loss = 0.5 * torch.mean(
-                loss_masks * torch.nn.functional.mse_loss(q_backup, q2_stream)
+            _q2_loss = 0.5 * ModelUtils.masked_mean(
+                torch.nn.functional.mse_loss(q_backup, q2_stream), loss_masks
             )
 
             q1_losses.append(_q1_loss)
@@ -258,8 +258,8 @@ def sac_value_loss(
                     v_backup = min_policy_qs[name] - torch.sum(
                         _ent_coef * log_probs, dim=1
                     )
-                value_loss = 0.5 * torch.mean(
-                    loss_masks * torch.nn.functional.mse_loss(values[name], v_backup)
+                value_loss = 0.5 * ModelUtils.masked_mean(
+                    torch.nn.functional.mse_loss(values[name], v_backup), loss_masks
                 )
                 value_losses.append(value_loss)
         else:
@@ -278,9 +278,9 @@ def sac_value_loss(
                     v_backup = min_policy_qs[name] - torch.mean(
                         branched_ent_bonus, axis=0
                     )
-                value_loss = 0.5 * torch.mean(
-                    loss_masks
-                    * torch.nn.functional.mse_loss(values[name], v_backup.squeeze())
+                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))
@@ -300,7 +300,7 @@ def sac_policy_loss(
         if not discrete:
             mean_q1 = mean_q1.unsqueeze(1)
             batch_policy_loss = torch.mean(_ent_coef * log_probs - mean_q1, dim=1)
-            policy_loss = torch.mean(loss_masks * batch_policy_loss)
+            policy_loss = ModelUtils.masked_mean(batch_policy_loss, loss_masks)
         else:
             action_probs = log_probs.exp()
             branched_per_action_ent = ModelUtils.break_into_branches(
@@ -347,9 +347,8 @@ def sac_entropy_loss(
                 target_current_diff = torch.squeeze(
                     target_current_diff_branched, axis=2
                 )
-            entropy_loss = -torch.mean(
-                loss_masks
-                * torch.mean(self._log_ent_coef * target_current_diff, axis=1)
+            entropy_loss = -1 * ModelUtils.masked_mean(
+                torch.mean(self._log_ent_coef * target_current_diff, axis=1), loss_masks
             )
 
         return entropy_loss
@@ -497,7 +496,7 @@ def update(self, batch: AgentBuffer, num_sequences: int) -> Dict[str, float]:
                 memories=next_memories,
                 sequence_length=self.policy.sequence_length,
             )
-        masks = ModelUtils.list_to_tensor(batch["masks"], dtype=torch.int32)
+        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"])
 
diff --git a/ml-agents/mlagents/trainers/torch/layers.py b/ml-agents/mlagents/trainers/torch/layers.py
index 8dbb1cbcb4..d1c68887df 100644
--- a/ml-agents/mlagents/trainers/torch/layers.py
+++ b/ml-agents/mlagents/trainers/torch/layers.py
@@ -46,3 +46,27 @@ def linear_layer(
     layer.weight.data *= kernel_gain
     _init_methods[bias_init](layer.bias.data)
     return layer
+
+
+def lstm_layer(
+    input_size: int,
+    hidden_size: int,
+    num_layers: int = 1,
+    batch_first: bool = True,
+    forget_bias: float = 1.0,
+    kernel_init: Initialization = Initialization.XavierGlorotUniform,
+    bias_init: Initialization = Initialization.Zero,
+) -> torch.nn.Module:
+    """
+    Creates a torch.nn.LSTM and initializes its weights and biases. Provides a
+    forget_bias offset like is done in TensorFlow.
+    """
+    lstm = torch.nn.LSTM(input_size, hidden_size, num_layers, batch_first=batch_first)
+    # Add forget_bias to forget gate bias
+    for name, param in lstm.named_parameters():
+        if "weight" in name:
+            _init_methods[kernel_init](param.data)
+        elif "bias" in name:
+            _init_methods[bias_init](param.data)
+            param.data[hidden_size : 2 * hidden_size].add_(forget_bias)
+    return lstm
diff --git a/ml-agents/mlagents/trainers/torch/networks.py b/ml-agents/mlagents/trainers/torch/networks.py
index aff61ce0f6..b60cfbe543 100644
--- a/ml-agents/mlagents/trainers/torch/networks.py
+++ b/ml-agents/mlagents/trainers/torch/networks.py
@@ -14,6 +14,7 @@
 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
 
 ActivationFunction = Callable[[torch.Tensor], torch.Tensor]
 EncoderFunction = Callable[
@@ -50,7 +51,7 @@ def __init__(
         )
 
         if self.use_lstm:
-            self.lstm = nn.LSTM(self.h_size, self.m_size // 2, 1)
+            self.lstm = lstm_layer(self.h_size, self.m_size // 2, batch_first=True)
         else:
             self.lstm = None
 
@@ -101,13 +102,11 @@ def forward(
             raise Exception("No valid inputs to network.")
 
         if self.use_lstm:
-            encoding = encoding.view([sequence_length, -1, self.h_size])
+            # Resize to (batch, sequence length, encoding size)
+            encoding = encoding.reshape([-1, sequence_length, self.h_size])
             memories = torch.split(memories, self.m_size // 2, dim=-1)
-            encoding, memories = self.lstm(
-                encoding.contiguous(),
-                (memories[0].contiguous(), memories[1].contiguous()),
-            )
-            encoding = encoding.view([-1, self.m_size // 2])
+            encoding, memories = self.lstm(encoding, (memories[0], memories[1]))
+            encoding = encoding.reshape([-1, self.m_size // 2])
             memories = torch.cat(memories, dim=-1)
         return encoding, memories
 

From c7ea525591ce6a5469d2a3676074399fee9ac9f3 Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Wed, 12 Aug 2020 18:36:23 -0700
Subject: [PATCH 10/18] Redundant indexing

---
 ml-agents/mlagents/trainers/torch/networks.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ml-agents/mlagents/trainers/torch/networks.py b/ml-agents/mlagents/trainers/torch/networks.py
index b60cfbe543..0d76a85107 100644
--- a/ml-agents/mlagents/trainers/torch/networks.py
+++ b/ml-agents/mlagents/trainers/torch/networks.py
@@ -105,7 +105,7 @@ def forward(
             # Resize to (batch, sequence length, encoding size)
             encoding = encoding.reshape([-1, sequence_length, self.h_size])
             memories = torch.split(memories, self.m_size // 2, dim=-1)
-            encoding, memories = self.lstm(encoding, (memories[0], memories[1]))
+            encoding, memories = self.lstm(encoding, memories)
             encoding = encoding.reshape([-1, self.m_size // 2])
             memories = torch.cat(memories, dim=-1)
         return encoding, memories

From d8844301a218c1af432bc5b65488adb1c4f3ca9a Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Wed, 12 Aug 2020 18:41:31 -0700
Subject: [PATCH 11/18] Add test for lstm layer

---
 .../trainers/tests/torch/test_layers.py       | 22 ++++++++++++++++++-
 1 file changed, 21 insertions(+), 1 deletion(-)

diff --git a/ml-agents/mlagents/trainers/tests/torch/test_layers.py b/ml-agents/mlagents/trainers/tests/torch/test_layers.py
index 499d0de285..6d1132aa2e 100644
--- a/ml-agents/mlagents/trainers/tests/torch/test_layers.py
+++ b/ml-agents/mlagents/trainers/tests/torch/test_layers.py
@@ -1,6 +1,11 @@
 import torch
 
-from mlagents.trainers.torch.layers import Swish, linear_layer, Initialization
+from mlagents.trainers.torch.layers import (
+    Swish,
+    linear_layer,
+    lstm_layer,
+    Initialization,
+)
 
 
 def test_swish():
@@ -18,3 +23,18 @@ def test_initialization_layer():
     )
     assert torch.all(torch.eq(layer.weight.data, torch.zeros_like(layer.weight.data)))
     assert torch.all(torch.eq(layer.bias.data, torch.zeros_like(layer.bias.data)))
+
+
+def test_lstm_layer():
+    torch.manual_seed(0)
+    # Test zero for LSTM
+    layer = lstm_layer(
+        4, 4, kernel_init=Initialization.Zero, bias_init=Initialization.Zero
+    )
+    for name, param in layer.named_parameters():
+        if "weight" in name:
+            assert torch.all(torch.eq(param.data, torch.zeros_like(param.data)))
+        elif "bias" in name:
+            assert torch.all(
+                torch.eq(param.data[4:8], torch.ones_like(param.data[4:8]))
+            )

From 5ef36484cf13cb2c0572994d4bf51704c3b097c8 Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Wed, 12 Aug 2020 19:38:05 -0700
Subject: [PATCH 12/18] Fix per-block lstm initialization

---
 ml-agents/mlagents/trainers/torch/layers.py | 20 ++++++++++++++++----
 1 file changed, 16 insertions(+), 4 deletions(-)

diff --git a/ml-agents/mlagents/trainers/torch/layers.py b/ml-agents/mlagents/trainers/torch/layers.py
index d1c68887df..707d4748a5 100644
--- a/ml-agents/mlagents/trainers/torch/layers.py
+++ b/ml-agents/mlagents/trainers/torch/layers.py
@@ -64,9 +64,21 @@ def lstm_layer(
     lstm = torch.nn.LSTM(input_size, hidden_size, num_layers, batch_first=batch_first)
     # Add forget_bias to forget gate bias
     for name, param in lstm.named_parameters():
+        # Each weight and bias is a concatenation of 4 matrices
         if "weight" in name:
-            _init_methods[kernel_init](param.data)
-        elif "bias" in name:
-            _init_methods[bias_init](param.data)
-            param.data[hidden_size : 2 * hidden_size].add_(forget_bias)
+            for idx in range(4):
+                block_size = param.shape[0] // 4
+                _init_methods[kernel_init](
+                    param.data[idx * block_size : (idx + 1) * block_size]
+                )
+        if "bias" in name:
+            for idx in range(4):
+                block_size = param.shape[0] // 4
+                _init_methods[bias_init](
+                    param.data[idx * block_size : (idx + 1) * block_size]
+                )
+                if idx == 1:
+                    param.data[idx * block_size : (idx + 1) * block_size].add_(
+                        forget_bias
+                    )
     return lstm

From 767f399319a40dcfdff60d6a362ba2d62a505698 Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Wed, 12 Aug 2020 23:55:18 -0700
Subject: [PATCH 13/18] Add memories and sequence length to critic_pass

---
 .../trainers/optimizer/torch_optimizer.py     | 42 +++----------------
 ml-agents/mlagents/trainers/torch/networks.py | 25 +++++++----
 2 files changed, 21 insertions(+), 46 deletions(-)

diff --git a/ml-agents/mlagents/trainers/optimizer/torch_optimizer.py b/ml-agents/mlagents/trainers/optimizer/torch_optimizer.py
index 7b47173db8..422295d811 100644
--- a/ml-agents/mlagents/trainers/optimizer/torch_optimizer.py
+++ b/ml-agents/mlagents/trainers/optimizer/torch_optimizer.py
@@ -1,7 +1,6 @@
 from typing import Dict, 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
@@ -10,7 +9,6 @@
 from mlagents.trainers.policy.torch_policy import TorchPolicy
 from mlagents.trainers.optimizer import Optimizer
 from mlagents.trainers.settings import TrainerSettings
-from mlagents.trainers.trajectory import SplitObservations
 from mlagents.trainers.torch.utils import ModelUtils
 
 
@@ -42,35 +40,6 @@ def create_reward_signals(self, reward_signal_configs):
                 reward_signal, self.policy.behavior_spec, settings
             )
 
-    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 = 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 not self.reward_signals[k].ignore_done:
-                    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]]:
@@ -85,18 +54,17 @@ def get_trajectory_value_estimates(
         else:
             visual_obs = []
 
-        memory = torch.zeros([1, len(vector_obs[0]), self.policy.m_size])
+        memory = torch.zeros([1, 1, self.policy.m_size])
 
         next_obs = np.concatenate(next_obs, axis=-1)
         next_obs = [ModelUtils.list_to_tensor(next_obs).unsqueeze(0)]
-        next_memory = torch.zeros([1, 1, self.policy.m_size])
 
-        value_estimates = self.policy.actor_critic.critic_pass(
-            vector_obs, visual_obs, memory
+        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_obs, next_obs, next_memory
+        next_value_estimate, _ = self.policy.actor_critic.critic_pass(
+            next_obs, next_obs, next_memory, sequence_length=1
         )
 
         for name, estimate in value_estimates.items():
diff --git a/ml-agents/mlagents/trainers/torch/networks.py b/ml-agents/mlagents/trainers/torch/networks.py
index 0d76a85107..0e206ab4a6 100644
--- a/ml-agents/mlagents/trainers/torch/networks.py
+++ b/ml-agents/mlagents/trainers/torch/networks.py
@@ -209,7 +209,8 @@ def critic_pass(
         vec_inputs: List[torch.Tensor],
         vis_inputs: List[torch.Tensor],
         memories: Optional[torch.Tensor] = None,
-    ) -> Dict[str, torch.Tensor]:
+        sequence_length: int = 1,
+    ) -> Tuple[Dict[str, torch.Tensor], torch.Tensor]:
         """
         Get value outputs for the given obs.
         :param vec_inputs: List of vector inputs as tensors.
@@ -359,9 +360,12 @@ def critic_pass(
         vec_inputs: List[torch.Tensor],
         vis_inputs: List[torch.Tensor],
         memories: Optional[torch.Tensor] = None,
-    ) -> Dict[str, torch.Tensor]:
-        encoding, _ = self.network_body(vec_inputs, vis_inputs, memories=memories)
-        return self.value_heads(encoding)
+        sequence_length: int = 1,
+    ) -> Tuple[Dict[str, torch.Tensor], torch.Tensor]:
+        encoding, memories_out = self.network_body(
+            vec_inputs, vis_inputs, memories=memories, sequence_length=sequence_length
+        )
+        return self.value_heads(encoding), memories_out
 
     def get_dist_and_value(
         self,
@@ -426,16 +430,19 @@ def critic_pass(
         vec_inputs: List[torch.Tensor],
         vis_inputs: List[torch.Tensor],
         memories: Optional[torch.Tensor] = None,
-    ) -> Dict[str, torch.Tensor]:
+        sequence_length: int = 1,
+    ) -> Tuple[Dict[str, torch.Tensor], torch.Tensor]:
         if self.use_lstm:
             # Use only the back half of memories for critic
-            _, critic_mem = torch.split(memories, self.half_mem_size, -1)
+            actor_mem, critic_mem = torch.split(memories, self.half_mem_size, -1)
         else:
             critic_mem = None
-        value_outputs, _memories = self.critic(
-            vec_inputs, vis_inputs, memories=critic_mem
+        value_outputs, critic_mem_out = self.critic(
+            vec_inputs, vis_inputs, memories=critic_mem, sequence_length=sequence_length
         )
-        return value_outputs
+        # Make memories with the actor mem unchanged
+        memories_out = torch.cat([actor_mem, critic_mem_out], dim=-1)
+        return value_outputs, memories_out
 
     def get_dist_and_value(
         self,

From c104409413d6eb3ff4fff6b895d08a193dc4a8c6 Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Thu, 13 Aug 2020 10:03:37 -0700
Subject: [PATCH 14/18] Fix memory logic in SeparateActorCritic

---
 ml-agents/mlagents/trainers/torch/networks.py | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/ml-agents/mlagents/trainers/torch/networks.py b/ml-agents/mlagents/trainers/torch/networks.py
index 0e206ab4a6..bfe9c0ade3 100644
--- a/ml-agents/mlagents/trainers/torch/networks.py
+++ b/ml-agents/mlagents/trainers/torch/networks.py
@@ -432,16 +432,18 @@ def critic_pass(
         memories: Optional[torch.Tensor] = None,
         sequence_length: int = 1,
     ) -> Tuple[Dict[str, torch.Tensor], torch.Tensor]:
+        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)
-        else:
-            critic_mem = None
         value_outputs, critic_mem_out = self.critic(
             vec_inputs, vis_inputs, memories=critic_mem, sequence_length=sequence_length
         )
-        # Make memories with the actor mem unchanged
-        memories_out = torch.cat([actor_mem, critic_mem_out], dim=-1)
+        if actor_mem is not None:
+            # Make memories with the actor mem unchanged
+            memories_out = torch.cat([actor_mem, critic_mem_out], dim=-1)
+        else:
+            memories_out = None
         return value_outputs, memories_out
 
     def get_dist_and_value(

From ca77ea5564ac8ba1971e9d559fe0819f239bfcea Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Thu, 13 Aug 2020 11:19:25 -0700
Subject: [PATCH 15/18] Fix LSTM tests

---
 .../mlagents/trainers/tests/torch/test_networks.py    | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/ml-agents/mlagents/trainers/tests/torch/test_networks.py b/ml-agents/mlagents/trainers/tests/torch/test_networks.py
index ba954f8a97..06f8b1ab25 100644
--- a/ml-agents/mlagents/trainers/tests/torch/test_networks.py
+++ b/ml-agents/mlagents/trainers/tests/torch/test_networks.py
@@ -45,16 +45,16 @@ def test_networkbody_lstm():
     obs_size = 4
     seq_len = 16
     network_settings = NetworkSettings(
-        memory=NetworkSettings.MemorySettings(sequence_length=seq_len, memory_size=4)
+        memory=NetworkSettings.MemorySettings(sequence_length=seq_len, memory_size=12)
     )
     obs_shapes = [(obs_size,)]
 
     networkbody = NetworkBody(obs_shapes, network_settings)
-    optimizer = torch.optim.Adam(networkbody.parameters(), lr=3e-3)
+    optimizer = torch.optim.Adam(networkbody.parameters(), lr=3e-4)
     sample_obs = torch.ones((1, seq_len, obs_size))
 
-    for _ in range(100):
-        encoded, _ = networkbody([sample_obs], [], memories=torch.ones(1, seq_len, 4))
+    for _ in range(200):
+        encoded, _ = networkbody([sample_obs], [], memories=torch.ones(1, seq_len, 12))
         # Try to force output to 1
         loss = torch.nn.functional.mse_loss(encoded, torch.ones(encoded.shape))
         optimizer.zero_grad()
@@ -196,10 +196,11 @@ def test_actor_critic(ac_type, lstm):
         # memories isn't always set to None, the network should be able to
         # deal with that.
     # Test critic pass
-    value_out = actor.critic_pass([sample_obs], [], memories=memories)
+    value_out, memories_out = actor.critic_pass([sample_obs], [], memories=memories)
     for stream in stream_names:
         if lstm:
             assert value_out[stream].shape == (network_settings.memory.sequence_length,)
+            assert memories_out.shape == memories.shape
         else:
             assert value_out[stream].shape == (1,)
 

From fd1a4ff1227dd6534116c1cc88e2f20c7554bd0a Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Thu, 13 Aug 2020 16:04:44 -0700
Subject: [PATCH 16/18] Fix next_obs in get_trajectory_value_estimates

---
 .../mlagents/trainers/optimizer/torch_optimizer.py  | 13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)

diff --git a/ml-agents/mlagents/trainers/optimizer/torch_optimizer.py b/ml-agents/mlagents/trainers/optimizer/torch_optimizer.py
index 422295d811..8261b363d7 100644
--- a/ml-agents/mlagents/trainers/optimizer/torch_optimizer.py
+++ b/ml-agents/mlagents/trainers/optimizer/torch_optimizer.py
@@ -3,6 +3,7 @@
 import numpy as np
 
 from mlagents.trainers.buffer import AgentBuffer
+from mlagents.trainers.trajectory import SplitObservations
 from mlagents.trainers.components.bc.module import BCModule
 from mlagents.trainers.torch.components.reward_providers import create_reward_provider
 
@@ -56,15 +57,21 @@ def get_trajectory_value_estimates(
 
         memory = torch.zeros([1, 1, self.policy.m_size])
 
-        next_obs = np.concatenate(next_obs, axis=-1)
-        next_obs = [ModelUtils.list_to_tensor(next_obs).unsqueeze(0)]
+        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_obs, next_obs, next_memory, sequence_length=1
+            next_vec_obs, next_vis_obs, next_memory, sequence_length=1
         )
 
         for name, estimate in value_estimates.items():

From 545dd9cf010dcef66925c89bc552d5686af009f0 Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Thu, 13 Aug 2020 16:12:22 -0700
Subject: [PATCH 17/18] Fix and test for masked_mean

---
 .../mlagents/trainers/tests/torch/test_utils.py  | 16 ++++++++++++++++
 ml-agents/mlagents/trainers/torch/utils.py       |  2 +-
 2 files changed, 17 insertions(+), 1 deletion(-)

diff --git a/ml-agents/mlagents/trainers/tests/torch/test_utils.py b/ml-agents/mlagents/trainers/tests/torch/test_utils.py
index 70306a89f3..0275581d08 100644
--- a/ml-agents/mlagents/trainers/tests/torch/test_utils.py
+++ b/ml-agents/mlagents/trainers/tests/torch/test_utils.py
@@ -198,3 +198,19 @@ def test_get_probs_and_entropy():
     assert entropies.shape == (1, len(dist_list))
     # Make sure the first action has high probability than the others.
     assert log_probs.flatten()[0] > log_probs.flatten()[1]
+
+
+def test_masked_mean():
+    test_input = torch.tensor([1, 2, 3, 4, 5])
+    masks = torch.ones_like(test_input).bool()
+    mean = ModelUtils.masked_mean(test_input, masks=masks)
+    assert mean == 3.0
+
+    masks = torch.tensor([False, False, True, True, True])
+    mean = ModelUtils.masked_mean(test_input, masks=masks)
+    assert mean == 4.0
+
+    # Make sure it works if all masks are off
+    masks = torch.tensor([False, False, False, False, False])
+    mean = ModelUtils.masked_mean(test_input, masks=masks)
+    assert mean == 0.0
diff --git a/ml-agents/mlagents/trainers/torch/utils.py b/ml-agents/mlagents/trainers/torch/utils.py
index ba6b7d57a0..0e855ea79b 100644
--- a/ml-agents/mlagents/trainers/torch/utils.py
+++ b/ml-agents/mlagents/trainers/torch/utils.py
@@ -293,4 +293,4 @@ def masked_mean(tensor: torch.Tensor, masks: torch.Tensor) -> torch.Tensor:
         :param tensor: Tensor which needs mean computation.
         :param masks: Boolean tensor of masks with same dimension as tensor.
         """
-        return (tensor * masks).sum() / masks.float().sum()
+        return (tensor * masks).sum() / torch.clamp(masks.float().sum(), min=1.0)

From 790ce905cd25f4765bfff390853ff6e2244fca63 Mon Sep 17 00:00:00 2001
From: Ervin Teng <ervin@unity3d.com>
Date: Thu, 13 Aug 2020 16:17:50 -0700
Subject: [PATCH 18/18] Use zeros_like

---
 ml-agents/mlagents/trainers/sac/optimizer_torch.py | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/ml-agents/mlagents/trainers/sac/optimizer_torch.py b/ml-agents/mlagents/trainers/sac/optimizer_torch.py
index 6d110f99f5..2a26d715f5 100644
--- a/ml-agents/mlagents/trainers/sac/optimizer_torch.py
+++ b/ml-agents/mlagents/trainers/sac/optimizer_torch.py
@@ -413,9 +413,7 @@ def update(self, batch: AgentBuffer, num_sequences: int) -> Dict[str, float]:
             memories = None
             next_memories = None
         # Q network memories are 0'ed out, since we don't have them during inference.
-        q_memories = torch.zeros(
-            (memories.shape[0], memories.shape[1], memories.shape[2] // 2)
-        )
+        q_memories = torch.zeros_like(next_memories)
 
         vis_obs: List[torch.Tensor] = []
         next_vis_obs: List[torch.Tensor] = []