Adjust gluon tests to new version

pom.xml

@@ -8,7 +8,7 @@
 
   <groupId>de.monticore.lang.monticar</groupId>
   <artifactId>embedded-montiarc-emadl-generator</artifactId>
-  <version>0.3.2-SNAPSHOT</version>
+  <version>0.3.2-SNAPSHOT-NG</version>
 
   <!-- == PROJECT DEPENDENCIES ============================================= -->
 
@@ -19,7 +19,7 @@
     <CNNTrain.version>0.3.2-SNAPSHOT</CNNTrain.version>
     <cnnarch-mxnet-generator.version>0.2.15-SNAPSHOT</cnnarch-mxnet-generator.version>
     <cnnarch-caffe2-generator.version>0.2.11-SNAPSHOT</cnnarch-caffe2-generator.version>
-    <cnnarch-gluon-generator.version>0.2.1-SNAPSHOT</cnnarch-gluon-generator.version>
+    <cnnarch-gluon-generator.version>0.2.1-SNAPSHOT-NG</cnnarch-gluon-generator.version>
     <embedded-montiarc-math-opt-generator>0.1.4</embedded-montiarc-math-opt-generator>
 
         <!-- .. Libraries .................................................. -->

src/test/resources/target_code/gluon/reinforcementModel/cartpole/reinforcement_learning/agent.py

@@ -114,6 +114,8 @@ class Agent(object):
         agent_session_file = os.path.join(session_dir, 'agent.p')
         logger = self._logger
 
+        self._training_stats.save_stats(self._output_directory, episode=self._current_episode)
+
         self._make_pickle_ready(session_dir)
 
         with open(agent_session_file, 'wb') as f:
@@ -177,6 +179,9 @@ class Agent(object):
         return states, actions, rewards, next_states, terminals
 
     def evaluate(self, target=None, sample_games=100, verbose=True):
+        if sample_games <= 0:
+            return 0
+
         target = self._target_score if target is None else target
         if target:
             target_achieved = 0
@@ -268,8 +273,9 @@ class Agent(object):
 
     def _save_net(self, net, filename, filedir=None):
         filedir = self._output_directory if filedir is None else filedir
-        filename = os.path.join(filedir, filename + '.params')
-        net.save_parameters(filename)
+        filename = os.path.join(filedir, filename)
+        net.save_parameters(filename + '.params')
+        net.export(filename, epoch=0)
 
     def save_best_network(self, path, epoch=0):
         self._logger.info(
@@ -367,6 +373,8 @@ class DdpgAgent(Agent):
 
     def _make_pickle_ready(self, session_dir):
         super(DdpgAgent, self)._make_pickle_ready(session_dir)
+        self._save_net(self._actor, 'current_actor')
+
         self._save_net(self._actor, 'actor', session_dir)
         self._actor = None
         self._save_net(self._critic, 'critic', session_dir)
@@ -457,9 +465,9 @@ class DdpgAgent(Agent):
         else:
             self._training_stats = DdpgTrainingStats(episodes)
 
-        # Initialize target Q' and mu'
-        self._actor_target = self._copy_actor()
-        self._critic_target = self._copy_critic()
+            # Initialize target Q' and mu'
+            self._actor_target = self._copy_actor()
+            self._critic_target = self._copy_critic()
 
         # Initialize l2 loss for critic network
         l2_loss = gluon.loss.L2Loss()
@@ -732,6 +740,7 @@ class DqnAgent(Agent):
 
     def _make_pickle_ready(self, session_dir):
         super(DqnAgent, self)._make_pickle_ready(session_dir)
+        self._save_net(self._qnet, 'current_qnet')
         self._save_net(self._qnet, 'qnet', session_dir)
         self._qnet = None
         self._save_net(self._target_qnet, 'target_net', session_dir)

src/test/resources/target_code/gluon/reinforcementModel/cartpole/reinforcement_learning/strategy.py

@@ -168,5 +168,5 @@ class OrnsteinUhlenbeckStrategy(BaseStrategy):
 
     def select_action(self, values):
         noise = self._evolve_state()
-        action = values + (self.cur_eps * noise)
+        action = (1.0 - self.cur_eps) * values + (self.cur_eps * noise)
         return np.clip(action, self._action_low, self._action_high)

src/test/resources/target_code/gluon/reinforcementModel/cartpole/reinforcement_learning/util.py

@@ -127,13 +127,15 @@ class TrainingStats(object):
         else:
             return self._all_total_rewards[0]
 
-    def save(self, path):
-        np.save(os.path.join(path, 'total_rewards'), self._all_total_rewards)
-        np.save(os.path.join(path, 'eps'), self._all_eps)
-        np.save(os.path.join(path, 'time'), self._all_time)
+    def save(self, path, episode=None):
+        if episode is None:
+            episode = self._max_episodes
+        np.save(os.path.join(path, 'total_rewards'), self._all_total_rewards[:episode])
+        np.save(os.path.join(path, 'eps'), self._all_eps[:episode])
+        np.save(os.path.join(path, 'time'), self._all_time[:episode])
         np.save(
             os.path.join(path, 'mean_reward'),
-            self._all_mean_reward_last_100_episodes)
+            self._all_mean_reward_last_100_episodes[:episode])
 
     def _log_episode(self, episode, start_time, training_steps, eps, reward):
         self.add_eps(episode, eps)
@@ -170,33 +172,43 @@ class DqnTrainingStats(TrainingStats):
         self._logger.info(info)
         return avg_reward
 
-    def save_stats(self, path):
+    def save_stats(self, path, episode=None):
+        if episode is None:
+            episode = self._max_episodes
+
+        all_total_rewards = self._all_total_rewards[:episode]
+        all_avg_loss = self._all_avg_loss[:episode]
+        all_eps = self._all_eps[:episode]
+        all_mean_reward_last_100_episodes = self._all_mean_reward_last_100_episodes[:episode]
+
         fig = plt.figure(figsize=(20, 20))
 
         sub_rewards = fig.add_subplot(221)
         sub_rewards.set_title('Total Rewards per episode')
         sub_rewards.plot(
-            np.arange(self._max_episodes), self._all_total_rewards)
+            np.arange(episode), all_total_rewards)
 
         sub_loss = fig.add_subplot(222)
         sub_loss.set_title('Avg. Loss per episode')
-        sub_loss.plot(np.arange(self._max_episodes), self._all_avg_loss)
+        sub_loss.plot(np.arange(episode), all_avg_loss)
 
         sub_eps = fig.add_subplot(223)
         sub_eps.set_title('Epsilon per episode')
-        sub_eps.plot(np.arange(self._max_episodes), self._all_eps)
+        sub_eps.plot(np.arange(episode), all_eps)
 
         sub_rewards = fig.add_subplot(224)
         sub_rewards.set_title('Avg. mean reward of last 100 episodes')
-        sub_rewards.plot(np.arange(self._max_episodes),
-                         self._all_mean_reward_last_100_episodes)
+        sub_rewards.plot(np.arange(episode),
+                         all_mean_reward_last_100_episodes)
 
-        self.save(path)
+        self.save(path, episode=episode)
         plt.savefig(os.path.join(path, 'stats.pdf'))
 
-    def save(self, path):
-        super(DqnTrainingStats, self).save(path)
-        np.save(os.path.join(path, 'avg_loss'), self._all_avg_loss)
+    def save(self, path, episode=None):
+        if episode is None:
+            episode = self._max_episodes
+        super(DqnTrainingStats, self).save(path, episode=episode)
+        np.save(os.path.join(path, 'avg_loss'), self._all_avg_loss[:episode])
 
 
 class DdpgTrainingStats(TrainingStats):
@@ -233,44 +245,56 @@ class DdpgTrainingStats(TrainingStats):
         self.logger.info(info)
         return avg_reward
 
-    def save(self, path):
-        super(DdpgTrainingStats, self).save(path)
+    def save(self, path, episode=None):
+        if episode is None:
+            episode = self._max_episodes
+        super(DdpgTrainingStats, self).save(path, episode=episode)
         np.save(os.path.join(
-            path, 'avg_critic_loss'), self._all_avg_critic_loss)
-        np.save(os.path.join(path, 'avg_actor_loss'), self._all_avg_actor_loss)
-        np.save(os.path.join(path, 'avg_qvalues'), self._all_avg_qvalues)
+            path, 'avg_critic_loss'), self._all_avg_critic_loss[:episode])
+        np.save(os.path.join(path, 'avg_actor_loss'), self._all_avg_actor_loss[:episode])
+        np.save(os.path.join(path, 'avg_qvalues'), self._all_avg_qvalues[:episode])
+
+    def save_stats(self, path, episode=None):
+        if episode is None:
+            episode = self._max_episodes
+
+        all_total_rewards = self._all_total_rewards[:episode]
+        all_avg_actor_loss = self._all_avg_actor_loss[:episode]
+        all_avg_critic_loss = self._all_avg_critic_loss[:episode]
+        all_avg_qvalues = self._all_avg_qvalues[:episode]
+        all_eps = self._all_eps[:episode]
+        all_mean_reward_last_100_episodes = self._all_mean_reward_last_100_episodes[:episode]
 
-    def save_stats(self, path):
         fig = plt.figure(figsize=(120, 120))
 
         sub_rewards = fig.add_subplot(321)
         sub_rewards.set_title('Total Rewards per episode')
         sub_rewards.plot(
-            np.arange(self._max_episodes), self._all_total_rewards)
+            np.arange(episode), all_total_rewards)
 
         sub_actor_loss = fig.add_subplot(322)
         sub_actor_loss.set_title('Avg. Actor Loss per episode')
         sub_actor_loss.plot(
-            np.arange(self._max_episodes), self._all_avg_actor_loss)
+            np.arange(episode), all_avg_actor_loss)
 
         sub_critic_loss = fig.add_subplot(323)
         sub_critic_loss.set_title('Avg. Critic Loss per episode')
         sub_critic_loss.plot(
-            np.arange(self._max_episodes), self._all_avg_critic_loss)
+            np.arange(episode), all_avg_critic_loss)
 
         sub_qvalues = fig.add_subplot(324)
         sub_qvalues.set_title('Avg. QValues per episode')
         sub_qvalues.plot(
-            np.arange(self._max_episodes), self._all_avg_qvalues)
+            np.arange(episode), all_avg_qvalues)
 
         sub_eps = fig.add_subplot(325)
         sub_eps.set_title('Epsilon per episode')
-        sub_eps.plot(np.arange(self._max_episodes), self._all_eps)
+        sub_eps.plot(np.arange(episode), all_eps)
 
         sub_rewards = fig.add_subplot(326)
         sub_rewards.set_title('Avg. mean reward of last 100 episodes')
-        sub_rewards.plot(np.arange(self._max_episodes),
-                         self._all_mean_reward_last_100_episodes)
+        sub_rewards.plot(np.arange(episode),
+                         all_mean_reward_last_100_episodes)
 
-        self.save(path)
+        self.save(path, episode=episode)
         plt.savefig(os.path.join(path, 'stats.pdf'))

src/test/resources/target_code/gluon/reinforcementModel/mountaincar/reinforcement_learning/agent.py

@@ -114,6 +114,8 @@ class Agent(object):
         agent_session_file = os.path.join(session_dir, 'agent.p')
         logger = self._logger
 
+        self._training_stats.save_stats(self._output_directory, episode=self._current_episode)
+
         self._make_pickle_ready(session_dir)
 
         with open(agent_session_file, 'wb') as f:
@@ -177,6 +179,9 @@ class Agent(object):
         return states, actions, rewards, next_states, terminals
 
     def evaluate(self, target=None, sample_games=100, verbose=True):
+        if sample_games <= 0:
+            return 0
+
         target = self._target_score if target is None else target
         if target:
             target_achieved = 0
@@ -268,8 +273,9 @@ class Agent(object):
 
     def _save_net(self, net, filename, filedir=None):
         filedir = self._output_directory if filedir is None else filedir
-        filename = os.path.join(filedir, filename + '.params')
-        net.save_parameters(filename)
+        filename = os.path.join(filedir, filename)
+        net.save_parameters(filename + '.params')
+        net.export(filename, epoch=0)
 
     def save_best_network(self, path, epoch=0):
         self._logger.info(
@@ -367,6 +373,8 @@ class DdpgAgent(Agent):
 
     def _make_pickle_ready(self, session_dir):
         super(DdpgAgent, self)._make_pickle_ready(session_dir)
+        self._save_net(self._actor, 'current_actor')
+
         self._save_net(self._actor, 'actor', session_dir)
         self._actor = None
         self._save_net(self._critic, 'critic', session_dir)
@@ -457,9 +465,9 @@ class DdpgAgent(Agent):
         else:
             self._training_stats = DdpgTrainingStats(episodes)
 
-        # Initialize target Q' and mu'
-        self._actor_target = self._copy_actor()
-        self._critic_target = self._copy_critic()
+            # Initialize target Q' and mu'
+            self._actor_target = self._copy_actor()
+            self._critic_target = self._copy_critic()
 
         # Initialize l2 loss for critic network
         l2_loss = gluon.loss.L2Loss()
@@ -732,6 +740,7 @@ class DqnAgent(Agent):
 
     def _make_pickle_ready(self, session_dir):
         super(DqnAgent, self)._make_pickle_ready(session_dir)
+        self._save_net(self._qnet, 'current_qnet')
         self._save_net(self._qnet, 'qnet', session_dir)
         self._qnet = None
         self._save_net(self._target_qnet, 'target_net', session_dir)

src/test/resources/target_code/gluon/reinforcementModel/mountaincar/reinforcement_learning/strategy.py

@@ -168,5 +168,5 @@ class OrnsteinUhlenbeckStrategy(BaseStrategy):
 
     def select_action(self, values):
         noise = self._evolve_state()
-        action = values + (self.cur_eps * noise)
+        action = (1.0 - self.cur_eps) * values + (self.cur_eps * noise)
         return np.clip(action, self._action_low, self._action_high)

src/test/resources/target_code/gluon/reinforcementModel/mountaincar/reinforcement_learning/util.py

@@ -127,13 +127,15 @@ class TrainingStats(object):
         else:
             return self._all_total_rewards[0]
 
-    def save(self, path):
-        np.save(os.path.join(path, 'total_rewards'), self._all_total_rewards)
-        np.save(os.path.join(path, 'eps'), self._all_eps)
-        np.save(os.path.join(path, 'time'), self._all_time)
+    def save(self, path, episode=None):
+        if episode is None:
+            episode = self._max_episodes
+        np.save(os.path.join(path, 'total_rewards'), self._all_total_rewards[:episode])
+        np.save(os.path.join(path, 'eps'), self._all_eps[:episode])
+        np.save(os.path.join(path, 'time'), self._all_time[:episode])
         np.save(
             os.path.join(path, 'mean_reward'),
-            self._all_mean_reward_last_100_episodes)
+            self._all_mean_reward_last_100_episodes[:episode])
 
     def _log_episode(self, episode, start_time, training_steps, eps, reward):
         self.add_eps(episode, eps)
@@ -170,33 +172,43 @@ class DqnTrainingStats(TrainingStats):
         self._logger.info(info)
         return avg_reward
 
-    def save_stats(self, path):
+    def save_stats(self, path, episode=None):
+        if episode is None:
+            episode = self._max_episodes
+
+        all_total_rewards = self._all_total_rewards[:episode]
+        all_avg_loss = self._all_avg_loss[:episode]
+        all_eps = self._all_eps[:episode]
+        all_mean_reward_last_100_episodes = self._all_mean_reward_last_100_episodes[:episode]
+
         fig = plt.figure(figsize=(20, 20))
 
         sub_rewards = fig.add_subplot(221)
         sub_rewards.set_title('Total Rewards per episode')
         sub_rewards.plot(
-            np.arange(self._max_episodes), self._all_total_rewards)
+            np.arange(episode), all_total_rewards)
 
         sub_loss = fig.add_subplot(222)
         sub_loss.set_title('Avg. Loss per episode')
-        sub_loss.plot(np.arange(self._max_episodes), self._all_avg_loss)
+        sub_loss.plot(np.arange(episode), all_avg_loss)
 
         sub_eps = fig.add_subplot(223)
         sub_eps.set_title('Epsilon per episode')
-        sub_eps.plot(np.arange(self._max_episodes), self._all_eps)
+        sub_eps.plot(np.arange(episode), all_eps)
 
         sub_rewards = fig.add_subplot(224)
         sub_rewards.set_title('Avg. mean reward of last 100 episodes')
-        sub_rewards.plot(np.arange(self._max_episodes),
-                         self._all_mean_reward_last_100_episodes)
+        sub_rewards.plot(np.arange(episode),
+                         all_mean_reward_last_100_episodes)
 
-        self.save(path)
+        self.save(path, episode=episode)
         plt.savefig(os.path.join(path, 'stats.pdf'))
 
-    def save(self, path):
-        super(DqnTrainingStats, self).save(path)
-        np.save(os.path.join(path, 'avg_loss'), self._all_avg_loss)
+    def save(self, path, episode=None):
+        if episode is None:
+            episode = self._max_episodes
+        super(DqnTrainingStats, self).save(path, episode=episode)
+        np.save(os.path.join(path, 'avg_loss'), self._all_avg_loss[:episode])
 
 
 class DdpgTrainingStats(TrainingStats):
@@ -233,44 +245,56 @@ class DdpgTrainingStats(TrainingStats):
         self.logger.info(info)
         return avg_reward
 
-    def save(self, path):
-        super(DdpgTrainingStats, self).save(path)
+    def save(self, path, episode=None):
+        if episode is None:
+            episode = self._max_episodes
+        super(DdpgTrainingStats, self).save(path, episode=episode)
         np.save(os.path.join(
-            path, 'avg_critic_loss'), self._all_avg_critic_loss)
-        np.save(os.path.join(path, 'avg_actor_loss'), self._all_avg_actor_loss)
-        np.save(os.path.join(path, 'avg_qvalues'), self._all_avg_qvalues)
+            path, 'avg_critic_loss'), self._all_avg_critic_loss[:episode])
+        np.save(os.path.join(path, 'avg_actor_loss'), self._all_avg_actor_loss[:episode])
+        np.save(os.path.join(path, 'avg_qvalues'), self._all_avg_qvalues[:episode])
+
+    def save_stats(self, path, episode=None):
+        if episode is None:
+            episode = self._max_episodes
+
+        all_total_rewards = self._all_total_rewards[:episode]
+        all_avg_actor_loss = self._all_avg_actor_loss[:episode]
+        all_avg_critic_loss = self._all_avg_critic_loss[:episode]
+        all_avg_qvalues = self._all_avg_qvalues[:episode]
+        all_eps = self._all_eps[:episode]
+        all_mean_reward_last_100_episodes = self._all_mean_reward_last_100_episodes[:episode]
 
-    def save_stats(self, path):
         fig = plt.figure(figsize=(120, 120))
 
         sub_rewards = fig.add_subplot(321)
         sub_rewards.set_title('Total Rewards per episode')
         sub_rewards.plot(
-            np.arange(self._max_episodes), self._all_total_rewards)
+            np.arange(episode), all_total_rewards)
 
         sub_actor_loss = fig.add_subplot(322)
         sub_actor_loss.set_title('Avg. Actor Loss per episode')
         sub_actor_loss.plot(
-            np.arange(self._max_episodes), self._all_avg_actor_loss)
+            np.arange(episode), all_avg_actor_loss)
 
         sub_critic_loss = fig.add_subplot(323)
         sub_critic_loss.set_title('Avg. Critic Loss per episode')
         sub_critic_loss.plot(
-            np.arange(self._max_episodes), self._all_avg_critic_loss)
+            np.arange(episode), all_avg_critic_loss)
 
         sub_qvalues = fig.add_subplot(324)
         sub_qvalues.set_title('Avg. QValues per episode')
         sub_qvalues.plot(
-            np.arange(self._max_episodes), self._all_avg_qvalues)
+            np.arange(episode), all_avg_qvalues)
 
         sub_eps = fig.add_subplot(325)
         sub_eps.set_title('Epsilon per episode')
-        sub_eps.plot(np.arange(self._max_episodes), self._all_eps)
+        sub_eps.plot(np.arange(episode), all_eps)
 
         sub_rewards = fig.add_subplot(326)
         sub_rewards.set_title('Avg. mean reward of last 100 episodes')
-        sub_rewards.plot(np.arange(self._max_episodes),
-                         self._all_mean_reward_last_100_episodes)
+        sub_rewards.plot(np.arange(episode),
+                         all_mean_reward_last_100_episodes)
 
-        self.save(path)
+        self.save(path, episode=episode)
         plt.savefig(os.path.join(path, 'stats.pdf'))

src/test/resources/target_code/gluon/reinforcementModel/torcs/reinforcement_learning/agent.py

@@ -114,6 +114,8 @@ class Agent(object):
         agent_session_file = os.path.join(session_dir, 'agent.p')
         logger = self._logger
 
+        self._training_stats.save_stats(self._output_directory, episode=self._current_episode)
+
         self._make_pickle_ready(session_dir)
 
         with open(agent_session_file, 'wb') as f:
@@ -177,6 +179,9 @@ class Agent(object):
         return states, actions, rewards, next_states, terminals
 
     def evaluate(self, target=None, sample_games=100, verbose=True):
+        if sample_games <= 0:
+            return 0
+
         target = self._target_score if target is None else target
         if target:
             target_achieved = 0
@@ -268,8 +273,9 @@ class Agent(object):
 
     def _save_net(self, net, filename, filedir=None):
         filedir = self._output_directory if filedir is None else filedir
-        filename = os.path.join(filedir, filename + '.params')
-        net.save_parameters(filename)
+        filename = os.path.join(filedir, filename)
+        net.save_parameters(filename + '.params')
+        net.export(filename, epoch=0)
 
     def save_best_network(self, path, epoch=0):
         self._logger.info(
@@ -367,6 +373,8 @@ class DdpgAgent(Agent):
 
     def _make_pickle_ready(self, session_dir):
         super(DdpgAgent, self)._make_pickle_ready(session_dir)
+        self._save_net(self._actor, 'current_actor')
+
         self._save_net(self._actor, 'actor', session_dir)
         self._actor = None
         self._save_net(self._critic, 'critic', session_dir)
@@ -457,9 +465,9 @@ class DdpgAgent(Agent):
         else:
             self._training_stats = DdpgTrainingStats(episodes)
 
-        # Initialize target Q' and mu'
-        self._actor_target = self._copy_actor()
-        self._critic_target = self._copy_critic()
+            # Initialize target Q' and mu'
+            self._actor_target = self._copy_actor()
+            self._critic_target = self._copy_critic()
 
         # Initialize l2 loss for critic network
         l2_loss = gluon.loss.L2Loss()
@@ -732,6 +740,7 @@ class DqnAgent(Agent):
 
     def _make_pickle_ready(self, session_dir):
         super(DqnAgent, self)._make_pickle_ready(session_dir)
+        self._save_net(self._qnet, 'current_qnet')
         self._save_net(self._qnet, 'qnet', session_dir)
         self._qnet = None
         self._save_net(self._target_qnet, 'target_net', session_dir)

src/test/resources/target_code/gluon/reinforcementModel/torcs/reinforcement_learning/environment.py

@@ -47,7 +47,7 @@ from std_msgs.msg import Float32MultiArray, Bool, Int32, MultiArrayDimension, Fl
 class RosEnvironment(Environment):
     def __init__(self,
         ros_node_name='RosTrainingAgent',
-        timeout_in_s=3,
+        timeout_in_s=60,
         state_topic='state',
         action_topic='action',
         reset_topic='reset',
@@ -55,6 +55,7 @@ class RosEnvironment(Environment):
         reward_topic='reward'):
         super(RosEnvironment, self).__init__()
         self.__timeout_in_s = timeout_in_s
+        self.__in_reset = False
         self.__waiting_for_state_update = False
         self.__waiting_for_terminal_update = False
         self.__last_received_state = 0
@@ -82,13 +83,18 @@ class RosEnvironment(Environment):
         time.sleep(2)
 
     def reset(self):
+        self.__in_reset = True
         time.sleep(0.5)
         reset_message = Bool()
         reset_message.data = True
         self.__waiting_for_state_update = True
+        self.__waiting_for_terminal_update = False
+        self.__waiting_for_reward_update = False
         self.__reset_publisher.publish(reset_message)
+        self.__wait_for_new_state(self.__reset_publisher, reset_message)
         while self.__last_received_terminal:
-            self.__wait_for_new_state(self.__reset_publisher, reset_message)
+            pass
+        self.__in_reset = False
         return self.__last_received_state
 
     def step(self, action):
@@ -129,7 +135,7 @@ class RosEnvironment(Environment):
         rospy.signal_shutdown('Program ended!')
 
     def __state_callback(self, data):
-        self.__last_received_state = np.array(data.data, dtype='float32')
+        self.__last_received_state = np.array(data.data, dtype='float32').reshape((5,))
         rospy.logdebug('Received state: {}'.format(self.__last_received_state))
         self.__waiting_for_state_update = False
 

src/test/resources/target_code/gluon/reinforcementModel/torcs/reinforcement_learning/strategy.py

@@ -168,5 +168,5 @@ class OrnsteinUhlenbeckStrategy(BaseStrategy):
 
     def select_action(self, values):
         noise = self._evolve_state()
-        action = values + (self.cur_eps * noise)
+        action = (1.0 - self.cur_eps) * values + (self.cur_eps * noise)
         return np.clip(action, self._action_low, self._action_high)

src/test/resources/target_code/gluon/reinforcementModel/torcs/reinforcement_learning/util.py

@@ -127,13 +127,15 @@ class TrainingStats(object):
         else:
             return self._all_total_rewards[0]
 
-    def save(self, path):
-        np.save(os.path.join(path, 'total_rewards'), self._all_total_rewards)
-        np.save(os.path.join(path, 'eps'), self._all_eps)
-        np.save(os.path.join(path, 'time'), self._all_time)
+    def save(self, path, episode=None):
+        if episode is None:
+            episode = self._max_episodes
+        np.save(os.path.join(path, 'total_rewards'), self._all_total_rewards[:episode])
+        np.save(os.path.join(path, 'eps'), self._all_eps[:episode])
+        np.save(os.path.join(path, 'time'), self._all_time[:episode])
         np.save(
             os.path.join(path, 'mean_reward'),
-            self._all_mean_reward_last_100_episodes)
+            self._all_mean_reward_last_100_episodes[:episode])
 
     def _log_episode(self, episode, start_time, training_steps, eps, reward):
         self.add_eps(episode, eps)
@@ -170,33 +172,43 @@ class DqnTrainingStats(TrainingStats):
         self._logger.info(info)
         return avg_reward
 
-    def save_stats(self, path):
+    def save_stats(self, path, episode=None):
+        if episode is None:
+            episode = self._max_episodes
+
+        all_total_rewards = self._all_total_rewards[:episode]
+        all_avg_loss = self._all_avg_loss[:episode]
+        all_eps = self._all_eps[:episode]
+        all_mean_reward_last_100_episodes = self._all_mean_reward_last_100_episodes[:episode]
+
         fig = plt.figure(figsize=(20, 20))
 
         sub_rewards = fig.add_subplot(221)
         sub_rewards.set_title('Total Rewards per episode')
         sub_rewards.plot(