A single thread DDPG

utils/config.py

@@ -11,7 +11,7 @@ class Config:
         self.task_fn = None
         self.optimizer_fn = None
         self.actor_optimizer_fn = None
-        self.critic_optimizer_fn = None
+        self.critic_optimizer_fn = Nonea
         self.network_fn = None
         self.actor_network_fn = None
         self.critic_network_fn = None
@@ -50,3 +50,4 @@ class Config:
         self.save_interval = 0
         self.max_steps = 0
         self.success_threshold = float('inf')
+        self.render_episode_freq = 0

utils/misc.py

@@ -7,6 +7,7 @@
 import numpy as np
 import pickle
 import os
+import gym.monitoring
 
 def run_episodes(agent):
     config = agent.config
@@ -30,6 +31,12 @@ def run_episodes(agent):
                     agent_type, config.tag, agent.task.name), 'wb') as f:
                 pickle.dump([steps, rewards], f)
 
+        if config.render_episode_freq and ep % config.render_episode_freq == 0:
+            video_recoder = gym.monitoring.VideoRecorder(
+                env=agent.task.env, base_path='./data/video/%s-%s-%s-%d' % (agent_type, config.tag, agent.task.name, ep))
+            agent.episode(True, video_recoder)
+            video_recoder.close()
+
         if config.episode_limit and ep > config.episode_limit:
             break