logger saver

train.py

@@ -37,7 +37,7 @@
 
 from agent.ddpg import ddpgAgent
 
-NUM_EPISODES_ = 5000
+NUM_EPISODES_ = 20000
 
 def model_train(pretrained_):
 	# Create Environments
@@ -74,6 +74,7 @@ def model_train(pretrained_):
 	try:
 		act_range = env.action_space.high
 		rewards = []; critic_losses = []
+		max_reward = 0
 		for epi in range(NUM_EPISODES_):
 			print("=========EPISODE # %d =========="%epi)
 			obs = env.reset()
@@ -105,13 +106,14 @@ def model_train(pretrained_):
 					print("Episode#%d, steps:%d, rewards:%f"%(epi,t,epi_reward))
 					agent.replay(1)
 
-					# save weights at every 50 iters
-					if epi%50 == 0:
+					# save weights at the new records performance
+					if epi_reward > max_reward:
+						max_reward = epi_reward
 						dir_path = "%s/weights"%os.getcwd()
 						if not os.path.isdir(dir_path):
 							os.mkdir(dir_path)
 						path = dir_path+'/'+'gym_ddpg_'
-						agent.save_weights(path + 'ep%d'%epi)
+						agent.save_weights(path + 'ep%d_%f'%(epi,max_reward))
 
 
 					# save reward logs
@@ -132,12 +134,12 @@ def model_train(pretrained_):
 		if not os.path.isdir(dir_path):
 			os.mkdir(dir_path)
 		path = dir_path+'/'+'gym_ddpg_'
-		agent.save_weights(path +'temp')
+		agent.save_weights(path +'lastest')
 		env.close()
 
 		# log saver
 		import pickle
-		pickle.dump(open(path+'%s.log'%time.Time.now(),'wb'))
+		pickle.dump(logger,open(path+'%s.log'%time.time(),'wb'))
 
 
 argparser = argparse.ArgumentParser(