Remove step in SAC._train_step arglist, fixes issue of redundant wrap…

…ping of HER replay_buffer
hill-a · Sep 21, 2020 · 8123b35 · 8123b35
1 parent b7e0f40
commit 8123b35
Show file tree

Hide file tree

Showing 3 changed files with 8 additions and 5 deletions.
diff --git a/docs/misc/changelog.rst b/docs/misc/changelog.rst
@@ -65,7 +65,8 @@ Bug Fixes:
 - Fixed a bug in CloudPickleWrapper's (used by VecEnvs) ``__setstate___`` where loading was incorrectly using ``pickle.loads`` (@shwang).
 - Fixed a bug in ``SAC`` and ``TD3`` where the log timesteps was not correct(@YangRui2015)
 - Fixed a bug where the environment was reset twice when using ``evaluate_policy``
-- Fixed a bug where ``SAC`` uses wrong step to log to tensorboard after multiple calls to ``SAC.learn(..., reset_num_timesteps=True)``
+- Fixed a bug where ``SAC`` uses wrong step to log to tensorboard after multiple calls to ``SAC.learn(..., reset_num_timesteps=True)`` (@krishpop)
+- Fixed issue where HER replay buffer wrapper is used multiple times after multiple calls to ``HER.learn`` (@krishpop)
 
 Deprecations:
 ^^^^^^^^^^^^^

diff --git a/stable_baselines/her/her.py b/stable_baselines/her/her.py
@@ -70,6 +70,7 @@ def _create_replay_wrapper(self, env):
                                                 n_sampled_goal=self.n_sampled_goal,
                                                 goal_selection_strategy=self.goal_selection_strategy,
                                                 wrapped_env=self.env)
+        self.wrapped_buffer = False
 
     def set_env(self, env):
         assert not isinstance(env, VecEnvWrapper), "HER does not support VecEnvWrapper"
@@ -108,9 +109,11 @@ def setup_model(self):
 
     def learn(self, total_timesteps, callback=None, log_interval=100, tb_log_name="HER",
               reset_num_timesteps=True):
+        replay_wrapper = self.replay_wrapper if not self.wrapped_buffer else None
+        self.wrapped_buffer = True
         return self.model.learn(total_timesteps, callback=callback, log_interval=log_interval,
                                 tb_log_name=tb_log_name, reset_num_timesteps=reset_num_timesteps,
-                                replay_wrapper=self.replay_wrapper)
+                                replay_wrapper=replay_wrapper)
 
     def _check_obs(self, observation):
         if isinstance(observation, dict):

diff --git a/stable_baselines/sac/sac.py b/stable_baselines/sac/sac.py
@@ -313,8 +313,7 @@ def setup_model(self):
 
                 self.summary = tf.summary.merge_all()
 
-    def _train_step(self, step, writer, learning_rate):
-        del step
+    def _train_step(self, writer, learning_rate):
         # Sample a batch from the replay buffer
         batch = self.replay_buffer.sample(self.batch_size, env=self._vec_normalize_env)
         batch_obs, batch_actions, batch_rewards, batch_next_obs, batch_dones = batch
@@ -461,7 +460,7 @@ def learn(self, total_timesteps, callback=None,
                         frac = 1.0 - step / total_timesteps
                         current_lr = self.learning_rate(frac)
                         # Update policy and critics (q functions)
-                        mb_infos_vals.append(self._train_step(step, writer, current_lr))
+                        mb_infos_vals.append(self._train_step(writer, current_lr))
                         # Update target network
                         if (step + grad_step) % self.target_update_interval == 0:
                             # Update target network