update

codes/DQN/main.py

@@ -5,68 +5,78 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-12 00:48:57
 @LastEditor: John
-LastEditTime: 2021-04-13 19:03:39
+LastEditTime: 2021-04-18 14:44:45
 @Discription: 
 @Environment: python 3.7.7
 '''
-import sys,os
+from common.utils import save_results, make_dir, del_empty_dir
-curr_path = os.path.dirname(__file__)
-parent_path=os.path.dirname(curr_path) 
-sys.path.append(parent_path) # add current terminal path to sys.path
-
-import gym
-import torch
-import datetime
-from DQN.agent import DQN
 from common.plot import plot_rewards
-from common.utils import save_results,make_dir,del_empty_dir
+from DQN.agent import DQN
+import datetime
+import torch
+import gym
+import sys
+import os
+curr_path = os.path.dirname(__file__)
+parent_path = os.path.dirname(curr_path)
+sys.path.append(parent_path)  # add current terminal path to sys.path
+
+
+curr_time = datetime.datetime.now().strftime(
+    "%Y%m%d-%H%M%S")  # obtain current time
 
-curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # obtain current time
 
 class DQNConfig:
     def __init__(self):
         self.algo = "DQN"  # name of algo
         self.env = 'CartPole-v0'
-        self.result_path = curr_path+"/results/" +self.env+'/'+curr_time+'/'  # path to save results
+        self.result_path = curr_path+"/outputs/" + self.env + \
+            '/'+curr_time+'/results/'  # path to save results
+        self.model_path = curr_path+"/outputs/" + self.env + \
+            '/'+curr_time+'/models/'  # path to save results
         self.gamma = 0.95
-        self.epsilon_start = 1 # e-greedy策略的初始epsilon
+        self.epsilon_start = 1  # e-greedy策略的初始epsilon
         self.epsilon_end = 0.01
         self.epsilon_decay = 500
-        self.lr = 0.0001 # learning rate
+        self.lr = 0.0001  # learning rate
-        self.memory_capacity = 10000 # Replay Memory容量
+        self.memory_capacity = 10000  # Replay Memory容量
         self.batch_size = 32
-        self.train_eps = 10 # 训练的episode数目
+        self.train_eps = 300  # 训练的episode数目
-        self.target_update = 2 # target net的更新频率
+        self.target_update = 2  # target net的更新频率
-        self.eval_eps = 20 # 测试的episode数目
+        self.eval_eps = 20  # 测试的episode数目
-        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测gpu
+        self.device = torch.device(
-        self.hidden_dim = 256 # 神经网络隐藏层维度
+            "cuda" if torch.cuda.is_available() else "cpu")  # 检测gpu
- 
+        self.hidden_dim = 256  # 神经网络隐藏层维度
-def train(cfg,env,agent):
+
+
+def train(cfg, env, agent):
     print('Start to train !')
+    print(f'Env:{cfg.env}, Algorithm:{cfg.algo}, Device:{cfg.device}')
     rewards = []
-    ma_rewards = [] # moveing average reward
+    ma_rewards = []  # moveing average reward
     for i_episode in range(cfg.train_eps):
-        state = env.reset() 
+        state = env.reset()
         done = False
         ep_reward = 0
         while not done:
-            action = agent.choose_action(state) 
+            action = agent.choose_action(state)
-            next_state, reward, done, _ = env.step(action) 
+            next_state, reward, done, _ = env.step(action)
             ep_reward += reward
-            agent.memory.push(state, action, reward, next_state, done) 
+            agent.memory.push(state, action, reward, next_state, done)
-            state = next_state 
+            state = next_state
-            agent.update() 
+            agent.update()
         if i_episode % cfg.target_update == 0:
             agent.target_net.load_state_dict(agent.policy_net.state_dict())
-        print('Episode:{}/{}, Reward:{}'.format(i_episode+1,cfg.train_eps,ep_reward))
+        print('Episode:{}/{}, Reward:{}'.format(i_episode+1, cfg.train_eps, ep_reward))
         rewards.append(ep_reward)
         # 计算滑动窗口的reward
         if ma_rewards:
             ma_rewards.append(0.9*ma_rewards[-1]+0.1*ep_reward)
         else:
-            ma_rewards.append(ep_reward)   
+            ma_rewards.append(ep_reward)
     print('Complete training！')
-    return rewards,ma_rewards
+    return rewards, ma_rewards
+
 
 if __name__ == "__main__":
     cfg = DQNConfig()
@@ -74,9 +84,10 @@ if __name__ == "__main__":
     env.seed(1)
     state_dim = env.observation_space.shape[0]
     action_dim = env.action_space.n
-    agent = DQN(state_dim,action_dim,cfg)
+    agent = DQN(state_dim, action_dim, cfg)
-    rewards,ma_rewards = train(cfg,env,agent)
+    rewards, ma_rewards = train(cfg, env, agent)
-    make_dir(cfg.result_path)
+    make_dir(cfg.result_path, cfg.model_path)
-    agent.save(path=cfg.result_path)
+    agent.save(path=cfg.model_path)
-    save_results(rewards,ma_rewards,tag='train',path=cfg.result_path)
+    save_results(rewards, ma_rewards, tag='train', path=cfg.result_path)
-    plot_rewards(rewards,ma_rewards,tag="train",algo = cfg.algo,path=cfg.result_path)
+    plot_rewards(rewards, ma_rewards, tag="train",
+                 algo=cfg.algo, path=cfg.result_path)

codes/DQN/task1.py

@@ -1,88 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-@Author: John
-@Email: johnjim0816@gmail.com
-@Date: 2020-06-12 00:48:57
-@LastEditor: John
-LastEditTime: 2021-04-13 18:49:44
-@Discription: 
-@Environment: python 3.7.7
-'''
-import sys,os
-curr_path = os.path.dirname(__file__)
-parent_path=os.path.dirname(curr_path) 
-sys.path.append(parent_path) # add current terminal path to sys.path
-
-import gym
-import torch
-import datetime
-from DQN.agent import DQN
-from common.plot import plot_rewards
-from common.utils import save_results,make_dir,del_empty_dir
-
-SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # obtain current time
-SAVED_MODEL_PATH = curr_path+"/saved_model/"+SEQUENCE+'/' # path to save model
-RESULT_PATH = curr_path+"/results/"+SEQUENCE+'/' # path to save rewards
-make_dir(curr_path+"/saved_model/",curr_path+"/results/")
-del_empty_dir(curr_path+"/saved_model/",curr_path+"/results/")
-
-class DQNConfig:
-    def __init__(self):
-        self.env = 'LunarLander-v2'
-        self.algo = "DQN"  # name of algo
-        self.gamma = 0.95
-        self.epsilon_start = 1 # e-greedy策略的初始epsilon
-        self.epsilon_end = 0.01
-        self.epsilon_decay = 500
-        self.lr = 0.0001 # learning rate
-        self.memory_capacity = 1000000 # Replay Memory容量
-        self.batch_size = 64
-        self.train_eps = 300 # 训练的episode数目
-        self.train_steps = 1000
-        self.target_update = 2 # target net的更新频率
-        self.eval_eps = 20 # 测试的episode数目
-        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测gpu
-        self.hidden_dim = 256 # 神经网络隐藏层维度
- 
-def train(cfg,env,agent):
-    print('Start to train !')
-    rewards = []
-    ma_rewards = [] # moveing average reward
-    for i_episode in range(cfg.train_eps):
-        state = env.reset() 
-        ep_reward = 0
-        for i_step in range(cfg.train_steps):
-            action = agent.choose_action(state) 
-            next_state, reward, done, _ = env.step(action) 
-            ep_reward += reward
-            agent.memory.push(state, action, reward, next_state, done) 
-            state = next_state 
-            agent.update()
-            if done:
-                break 
-        if i_episode % cfg.target_update == 0:
-            agent.target_net.load_state_dict(agent.policy_net.state_dict())
-        print('Episode:{}/{}, Reward:{}'.format(i_episode+1,cfg.train_eps,ep_reward))
-        rewards.append(ep_reward)
-        # 计算滑动窗口的reward
-        if ma_rewards:
-            ma_rewards.append(0.9*ma_rewards[-1]+0.1*ep_reward)
-        else:
-            ma_rewards.append(ep_reward)   
-    print('Complete training！')
-    return rewards,ma_rewards
-
-if __name__ == "__main__":
-    cfg = DQNConfig()
-    env = gym.make(cfg.env)
-    env.seed(1)
-    state_dim = env.observation_space.shape[0]
-    action_dim = env.action_space.n
-    agent = DQN(state_dim,action_dim,cfg)
-    rewards,ma_rewards = train(cfg,env,agent)
-    make_dir(SAVED_MODEL_PATH,RESULT_PATH)
-    agent.save(path=SAVED_MODEL_PATH)
-    save_results(rewards,ma_rewards,tag='train',path=RESULT_PATH)
-    plot_rewards(rewards,ma_rewards,tag="train",algo = cfg.algo,path=RESULT_PATH)
-    del_empty_dir(SAVED_MODEL_PATH,RESULT_PATH)

codes/PPO/agent.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2021-03-23 15:17:42
 LastEditor: John
-LastEditTime: 2021-04-11 01:24:24
+LastEditTime: 2021-04-28 10:11:09
 Discription: 
 Environment: 
 '''
@@ -17,7 +17,6 @@ from PPO.model import Actor,Critic
 from PPO.memory import PPOMemory
 class PPO:
     def __init__(self, state_dim, action_dim,cfg):
-        self.env = cfg.env
         self.gamma = cfg.gamma
         self.policy_clip = cfg.policy_clip
         self.n_epochs = cfg.n_epochs
@@ -84,13 +83,13 @@ class PPO:
                 self.critic_optimizer.step()
         self.memory.clear()  
     def save(self,path):
-        actor_checkpoint = os.path.join(path, self.env+'_actor.pt')
+        actor_checkpoint = os.path.join(path, 'ppo_actor.pt')
-        critic_checkpoint= os.path.join(path, self.env+'_critic.pt')
+        critic_checkpoint= os.path.join(path, 'ppo_critic.pt')
         torch.save(self.actor.state_dict(), actor_checkpoint)
         torch.save(self.critic.state_dict(), critic_checkpoint)
     def load(self,path):
-        actor_checkpoint = os.path.join(path, self.env+'_actor.pt')
+        actor_checkpoint = os.path.join(path, 'ppo_actor.pt')
-        critic_checkpoint= os.path.join(path, self.env+'_critic.pt')
+        critic_checkpoint= os.path.join(path, 'ppo_critic.pt')
         self.actor.load_state_dict(torch.load(actor_checkpoint)) 
         self.critic.load_state_dict(torch.load(critic_checkpoint))  
 

codes/PPO/main.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2021-03-22 16:18:10
 LastEditor: John
-LastEditTime: 2021-04-11 01:24:41
+LastEditTime: 2021-04-28 10:13:00
 Discription: 
 Environment: 
 '''
@@ -19,24 +19,16 @@ import torch
 import datetime
 from PPO.agent import PPO
 from common.plot import plot_rewards
-from common.utils import save_results
+from common.utils import save_results,make_dir
 
-SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # 获取当前时间
+curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # obtain current time
-SAVED_MODEL_PATH = os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"+SEQUENCE+'/' # 生成保存的模型路径
-if not os.path.exists(os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"): # 检测是否存在文件夹
-    os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/saved_model/")
-if not os.path.exists(SAVED_MODEL_PATH): # 检测是否存在文件夹
-    os.mkdir(SAVED_MODEL_PATH)
-RESULT_PATH = os.path.split(os.path.abspath(__file__))[0]+"/results/"+SEQUENCE+'/' # 存储reward的路径
-if not os.path.exists(os.path.split(os.path.abspath(__file__))[0]+"/results/"): # 检测是否存在文件夹
-    os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/results/")
-if not os.path.exists(RESULT_PATH): # 检测是否存在文件夹
-    os.mkdir(RESULT_PATH)
 
 class PPOConfig:
     def __init__(self) -> None:
         self.env = 'CartPole-v0'
         self.algo = 'PPO'
+        self.result_path = curr_path+"/results/" +self.env+'/'+curr_time+'/results/'  # path to save results
+        self.model_path = curr_path+"/results/" +self.env+'/'+curr_time+'/models/'  # path to save models
         self.batch_size = 5
         self.gamma=0.99
         self.n_epochs = 4
@@ -50,12 +42,10 @@ class PPOConfig:
         self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # check gpu
         
 def train(cfg,env,agent):
-    best_reward = env.reward_range[0]
     rewards= []
     ma_rewards = [] # moving average rewards
-    avg_reward = 0
     running_steps = 0
-    for i_episode in range(cfg.train_eps):
+    for i_ep in range(cfg.train_eps):
         state = env.reset()
         done = False
         ep_reward = 0
@@ -74,21 +64,18 @@ def train(cfg,env,agent):
                 0.9*ma_rewards[-1]+0.1*ep_reward)
         else:
             ma_rewards.append(ep_reward)
-        avg_reward = np.mean(rewards[-100:])
+        print(f"Episode:{i_ep+1}/{cfg.train_eps}, Reward:{ep_reward:.3f}")
-        if avg_rewardself.actor_lr = 0.002
-        self.critic_lr = 0.005 > best_reward:
-            best_reward = avg_reward
-            agent.save(path=SAVED_MODEL_PATH)
-        print('Episode:{}/{}, Reward:{:.1f}, avg reward:{:.1f}, Done:{}'.format(i_episode+1,cfg.train_eps,ep_reward,avg_reward,done))
     return rewards,ma_rewards
 
 if __name__ == '__main__':
-    cfg = PPOConfig()
+    cfg  = PPOConfig()
     env = gym.make(cfg.env)
-    env.seed(1)
+    env.seed(1) # Set seeds
     state_dim=env.observation_space.shape[0]
     action_dim=env.action_space.n
     agent = PPO(state_dim,action_dim,cfg)
     rewards,ma_rewards = train(cfg,env,agent)
-    save_results(rewards,ma_rewards,tag='train',path=RESULT_PATH)
+    make_dir(cfg.result_path,cfg.model_path)
-    plot_rewards(rewards,ma_rewards,tag="train",algo = cfg.algo,path=RESULT_PATH)
+    agent.save(path=cfg.model_path)
+    save_results(rewards,ma_rewards,tag='train',path=cfg.result_path)
+    plot_rewards(rewards,ma_rewards,tag="train",env=cfg.env,algo = cfg.algo,path=cfg.result_path)

codes/PolicyGradient/README.md

@@ -1,14 +1,18 @@
 # Policy Gradient
-实现的是Policy Gradient最基本的REINFORCE方法
-## 使用说明
-直接运行```main.py```即可
-## 原理讲解
 
-参考我的博客[Policy Gradient算法实战](https://blog.csdn.net/JohnJim0/article/details/110236851)
 
-## 环境
+Policy-based方法是强化学习中与Value-based(比如Q-learning)相对的方法，其目的是对策略本身进行梯度下降，相关基础知识参考[Datawhale-Policy Gradient](https://datawhalechina.github.io/leedeeprl-notes/#/chapter4/chapter4)。
-python 3.7.9、pytorch 1.6.0
+其中REINFORCE是一个最基本的Policy Gradient方法，主要解决策略梯度无法直接计算的问题，具体原理参考[CSDN-REINFORCE和Reparameterization Trick](https://blog.csdn.net/JohnJim0/article/details/110230703)
-## 程序运行方法
+
+## 伪代码
+
+结合REINFORCE原理，其伪代码如下：
+
+![img](assets/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L0pvaG5KaW0w,size_16,color_FFFFFF,t_70-20210428001336032.png)
+
+## 实现
+
+
 
 ## 参考
 

codes/README.md

@@ -22,7 +22,9 @@ python 3.7、pytorch 1.6.0-1.7.1、gym 0.17.0-0.18.0
 
 ## 使用说明
 
-运行```main.py```或者```main.ipynb```，或者包含```task```名的文件(比如```task1.py```)
+运行带有```train```的py文件或ipynb文件进行训练，如果前面带有```task```如```task0_train.py```，表示对task0任务训练
+类似的带有```eval```即为测试。
+
 ## 算法进度
 
 |                 算法名称                 |                         相关论文材料                         | 环境                                  |                备注                |
@@ -45,11 +47,8 @@ python 3.7、pytorch 1.6.0-1.7.1、gym 0.17.0-0.18.0
 
 
 
-
 ## Refs
 
-
 [RL-Adventure-2](https://github.com/higgsfield/RL-Adventure-2)
 
 [RL-Adventure](https://github.com/higgsfield/RL-Adventure)
-

codes/README_en.md

@@ -19,10 +19,14 @@ Note that ```model.py```,```memory.py```,```plot.py``` shall be utilized in diff
 
 ## Runnig Environment
 
-python 3.7.9、pytorch 1.6.0、gym 0.18.0
+python 3.7、pytorch 1.6.0-1.7.1、gym 0.17.0-0.18.0
 ## Usage
+运行带有```train```的py文件或ipynb文件进行训练，如果前面带有```task```如```task0_train.py```，表示对task0任务训练
+类似的带有```eval```即为测试。
 
-run ```main.py``` or ```main.ipynb```, or run files with ```task```(like ```task1.py```)
+run python scripts or jupyter notebook file with ```train``` to train the agent, if there is a ```task``` like ```task0_train.py```, it means to train with task 0.
+
+similar to file with ```eval```, which means to evaluate the agent.
 
 ## Schedule
 
@@ -51,5 +55,3 @@ run ```main.py``` or ```main.ipynb```, or run files with ```task```(like ```task
 [RL-Adventure-2](https://github.com/higgsfield/RL-Adventure-2)
 
 [RL-Adventure](https://github.com/higgsfield/RL-Adventure)
-
-

codes/RandomPolicy/main.py

@@ -0,0 +1,108 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: JiangJi
+Email: johnjim0816@gmail.com
+Date: 2021-04-21 11:07:57
+LastEditor: JiangJi
+LastEditTime: 2021-04-21 11:15:00
+Discription: 
+Environment: 
+'''
+import sys,os
+curr_path = os.path.dirname(__file__)
+parent_path=os.path.dirname(curr_path) 
+sys.path.append(parent_path) # add current terminal path to sys.path
+
+import torch
+import gym
+import numpy as np
+import datetime
+
+from common.plot import plot_rewards
+from common.utils import save_results,make_dir
+
+curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # obtain current time
+	
+
+class TD3Config:
+	def __init__(self) -> None:
+		self.algo = 'TD3'
+		self.env = 'HalfCheetah-v2'
+		self.seed = 0
+		self.result_path = curr_path+"/results/" +self.env+'/'+curr_time+'/results/'  # path to save results
+		self.model_path = curr_path+"/results/" +self.env+'/'+curr_time+'/models/'  # path to save models
+		self.eval_freq = 5e3 # How often (time steps) we evaluate
+		# self.train_eps = 800
+		self.max_timestep = 4000000 # Max time steps to run environment
+		
+# Runs policy for X episodes and returns average reward
+# A fixed seed is used for the eval environment
+def eval(env_name,seed, eval_episodes=10):
+	eval_env = gym.make(env_name)
+	eval_env.seed(seed + 100)
+	avg_reward = 0.
+	for _ in range(eval_episodes):
+		state, done = eval_env.reset(), False
+		while not done:
+			# eval_env.render()
+			action = eval_env.action_space.sample()
+			state, reward, done, _ = eval_env.step(action)
+			avg_reward += reward
+	avg_reward /= eval_episodes
+	print("---------------------------------------")
+	print(f"Evaluation over {eval_episodes} episodes: {avg_reward:.3f}")
+	print("---------------------------------------")
+	return avg_reward
+
+def train(cfg,env):
+	# Evaluate untrained policy
+	evaluations = [eval(cfg.env, cfg.seed)]
+	state, done = env.reset(), False
+	ep_reward = 0
+	ep_timesteps = 0
+	episode_num = 0
+	rewards = []
+	ma_rewards = [] # moveing average reward
+	for t in range(int(cfg.max_timestep)):
+		ep_timesteps += 1
+		# Select action randomly
+		action = env.action_space.sample()
+		# Perform action
+		next_state, reward, done, _ = env.step(action) 
+		state = next_state
+		ep_reward += reward
+		if done: 
+			# +1 to account for 0 indexing. +0 on ep_timesteps since it will increment +1 even if done=True
+			print(f"Episode:{episode_num+1}, Episode T:{ep_timesteps}, Reward:{ep_reward:.3f}")
+			# Reset environment
+			state, done = env.reset(), False
+			rewards.append(ep_reward)
+			# 计算滑动窗口的reward
+			if ma_rewards:
+				ma_rewards.append(0.9*ma_rewards[-1]+0.1*ep_reward)
+			else:
+				ma_rewards.append(ep_reward) 
+			ep_reward = 0
+			ep_timesteps = 0
+			episode_num += 1 
+		# Evaluate episode
+		if (t + 1) % cfg.eval_freq == 0:
+			evaluations.append(eval(cfg.env, cfg.seed))
+	return rewards, ma_rewards
+
+if __name__ == "__main__":
+	cfg  = TD3Config()
+	env = gym.make(cfg.env)
+	env.seed(cfg.seed) # Set seeds
+	torch.manual_seed(cfg.seed)
+	np.random.seed(cfg.seed)
+	rewards,ma_rewards = train(cfg,env)
+	make_dir(cfg.result_path)
+	save_results(rewards,ma_rewards,tag='train',path=cfg.result_path)
+	plot_rewards(rewards,ma_rewards,tag="train",env=cfg.env,algo = cfg.algo,path=cfg.result_path)
+	# cfg.result_path = './TD3/results/HalfCheetah-v2/20210416-130341/'
+	# agent.load(cfg.result_path)
+	# eval(cfg.env,agent, cfg.seed)
+	
+		

codes/TD3/agent.py

@@ -92,14 +92,10 @@ class TD3(object):
 		self.critic_optimizer = torch.optim.Adam(self.critic.parameters(), lr=3e-4)
 		self.memory = ReplayBuffer(state_dim, action_dim)
 
-		
-
-
 	def choose_action(self, state):
 		state = torch.FloatTensor(state.reshape(1, -1)).to(self.device)
 		return self.actor(state).cpu().data.numpy().flatten()
 
-
 	def update(self):
 		self.total_it += 1
 
@@ -167,4 +163,4 @@ class TD3(object):
 		self.actor.load_state_dict(torch.load(path + "td3_actor"))
 		self.actor_optimizer.load_state_dict(torch.load(path + "td3_actor_optimizer"))
 		self.actor_target = copy.deepcopy(self.actor)
-		
+		

codes/TD3/task0_eval.py

@@ -0,0 +1,89 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: JiangJi
+Email: johnjim0816@gmail.com
+Date: 2021-04-23 20:36:23
+LastEditor: JiangJi
+LastEditTime: 2021-04-23 20:37:22
+Discription: 
+Environment: 
+'''
+import sys,os
+curr_path = os.path.dirname(__file__)
+parent_path=os.path.dirname(curr_path) 
+sys.path.append(parent_path) # add current terminal path to sys.path
+
+import torch
+import gym
+import numpy as np
+import datetime
+
+
+from TD3.agent import TD3
+from common.plot import plot_rewards
+from common.utils import save_results,make_dir
+
+curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # obtain current time
+
+class TD3Config:
+	def __init__(self) -> None:
+		self.algo = 'TD3 and Random'
+		self.env = 'HalfCheetah-v2'
+		self.seed = 0
+		self.result_path = curr_path+"/results/" +self.env+'/'+curr_time+'/results/'  # path to save results
+		self.model_path = curr_path+"/results/" +self.env+'/'+curr_time+'/models/'  # path to save models
+		self.start_timestep = 25e3 # Time steps initial random policy is used
+		self.eval_freq = 5e3 # How often (time steps) we evaluate
+		self.max_timestep = 200000 # Max time steps to run environment
+		self.expl_noise = 0.1 # Std of Gaussian exploration noise
+		self.batch_size = 256 # Batch size for both actor and critic
+		self.gamma = 0.99 # gamma factor
+		self.lr = 0.0005 # Target network update rate 
+		self.policy_noise = 0.2 # Noise added to target policy during critic update
+		self.noise_clip = 0.5  # Range to clip target policy noise
+		self.policy_freq = 2 # Frequency of delayed policy updates
+		self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+		
+# Runs policy for X episodes and returns average reward
+# A fixed seed is used for the eval environment
+def eval(env_name,agent, seed, eval_episodes=50):
+	eval_env = gym.make(env_name)
+	eval_env.seed(seed + 100)
+	rewards,ma_rewards =[],[]
+	for i_episode in range(eval_episodes):
+		ep_reward = 0
+		state, done = eval_env.reset(), False
+		while not done:
+			eval_env.render()
+			action = agent.choose_action(np.array(state))
+			state, reward, done, _ = eval_env.step(action)
+			ep_reward += reward
+		print(f"Episode:{i_episode+1}, Reward:{ep_reward:.3f}")
+		rewards.append(ep_reward)
+		# 计算滑动窗口的reward
+		if ma_rewards:
+			ma_rewards.append(0.9*ma_rewards[-1]+0.1*ep_reward)
+		else:
+			ma_rewards.append(ep_reward) 
+	return rewards,ma_rewards
+
+if __name__ == "__main__":
+	cfg  = TD3Config()
+	env = gym.make(cfg.env)
+	env.seed(cfg.seed) # Set seeds
+	torch.manual_seed(cfg.seed)
+	np.random.seed(cfg.seed)
+	state_dim = env.observation_space.shape[0]
+	action_dim = env.action_space.shape[0] 
+	max_action = float(env.action_space.high[0])
+	td3= TD3(state_dim,action_dim,max_action,cfg)
+	cfg.model_path = './TD3/results/HalfCheetah-v2/20210416-130341/models/'
+	td3.load(cfg.model_path)
+	td3_rewards,td3_ma_rewards = eval(cfg.env,td3,cfg.seed)
+	make_dir(cfg.result_path,cfg.model_path)
+	save_results(td3_rewards,td3_ma_rewards,tag='eval',path=cfg.result_path)
+	plot_rewards({'td3_rewards':td3_rewards,'td3_ma_rewards':td3_ma_rewards,},tag="eval",env=cfg.env,algo = cfg.algo,path=cfg.result_path)
+	# cfg.result_path = './TD3/results/HalfCheetah-v2/20210416-130341/'
+	# agent.load(cfg.result_path)
+	# eval(cfg.env,agent, cfg.seed)

codes/TD3/task0_train.py

@@ -21,11 +21,12 @@ class TD3Config:
 		self.algo = 'TD3'
 		self.env = 'HalfCheetah-v2'
 		self.seed = 0
-		self.result_path = curr_path+"/results/" +self.env+'/'+curr_time+'/'  # path to save results
+		self.result_path = curr_path+"/results/" +self.env+'/'+curr_time+'/results/'  # path to save results
+		self.model_path = curr_path+"/results/" +self.env+'/'+curr_time+'/models/'  # path to save models
 		self.start_timestep = 25e3 # Time steps initial random policy is used
 		self.eval_freq = 5e3 # How often (time steps) we evaluate
 		# self.train_eps = 800
-		self.max_timestep = 1600000 # Max time steps to run environment
+		self.max_timestep = 4000000 # Max time steps to run environment
 		self.expl_noise = 0.1 # Std of Gaussian exploration noise
 		self.batch_size = 256 # Batch size for both actor and critic
 		self.gamma = 0.99 # gamma factor
@@ -161,9 +162,12 @@ if __name__ == "__main__":
 	max_action = float(env.action_space.high[0])
 	agent = TD3(state_dim,action_dim,max_action,cfg)
 	rewards,ma_rewards = train(cfg,env,agent)
-	make_dir(cfg.result_path)
+	make_dir(cfg.result_path,cfg.model_path)
-	agent.save(path=cfg.result_path)
+	agent.save(path=cfg.model_path)
 	save_results(rewards,ma_rewards,tag='train',path=cfg.result_path)
 	plot_rewards(rewards,ma_rewards,tag="train",env=cfg.env,algo = cfg.algo,path=cfg.result_path)
+	# cfg.result_path = './TD3/results/HalfCheetah-v2/20210416-130341/'
+	# agent.load(cfg.result_path)
+	# eval(cfg.env,agent, cfg.seed)
 	
 		

codes/TD3/task1_eval.py

@@ -0,0 +1,83 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: JiangJi
+Email: johnjim0816@gmail.com
+Date: 2021-04-23 20:36:23
+LastEditor: JiangJi
+LastEditTime: 2021-04-28 10:14:33
+Discription: 
+Environment: 
+'''
+import sys,os
+curr_path = os.path.dirname(__file__)
+parent_path=os.path.dirname(curr_path) 
+sys.path.append(parent_path) # add current terminal path to sys.path
+
+import torch
+import gym
+import numpy as np
+import datetime
+
+
+from TD3.agent import TD3
+from common.plot import plot_rewards
+from common.utils import save_results,make_dir
+
+curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # obtain current time
+
+class TD3Config:
+	def __init__(self) -> None:
+		self.algo = 'TD3'
+		self.env = 'Pendulum-v0'
+		self.seed = 0
+		self.result_path = curr_path+"/results/" +self.env+'/'+curr_time+'/results/'  # path to save results
+		self.model_path = curr_path+"/results/" +self.env+'/'+curr_time+'/models/'  # path to save models
+		self.batch_size = 256 # Batch size for both actor and critic
+		self.gamma = 0.99 # gamma factor
+		self.lr = 0.0005 # Target network update rate 
+		self.policy_noise = 0.2 # Noise added to target policy during critic update
+		self.noise_clip = 0.5  # Range to clip target policy noise
+		self.policy_freq = 2 # Frequency of delayed policy updates
+		self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+		
+# Runs policy for X episodes and returns average reward
+# A fixed seed is used for the eval environment
+def eval(env_name,agent, seed, eval_episodes=50):
+	eval_env = gym.make(env_name)
+	eval_env.seed(seed + 100)
+	rewards,ma_rewards =[],[]
+	for i_episode in range(eval_episodes):
+		ep_reward = 0
+		state, done = eval_env.reset(), False
+		while not done:
+			# eval_env.render()
+			action = agent.choose_action(np.array(state))
+			state, reward, done, _ = eval_env.step(action)
+			ep_reward += reward
+		print(f"Episode:{i_episode+1}, Reward:{ep_reward:.3f}")
+		rewards.append(ep_reward)
+		# 计算滑动窗口的reward
+		if ma_rewards:
+			ma_rewards.append(0.9*ma_rewards[-1]+0.1*ep_reward)
+		else:
+			ma_rewards.append(ep_reward) 
+	return rewards,ma_rewards
+
+if __name__ == "__main__":
+	cfg  = TD3Config()
+	env = gym.make(cfg.env)
+	env.seed(cfg.seed) # Set seeds
+	torch.manual_seed(cfg.seed)
+	np.random.seed(cfg.seed)
+	state_dim = env.observation_space.shape[0]
+	action_dim = env.action_space.shape[0] 
+	max_action = float(env.action_space.high[0])
+	td3= TD3(state_dim,action_dim,max_action,cfg)
+	cfg.model_path = './TD3/results/Pendulum-v0/20210428-092059/models/'
+	cfg.result_path = './TD3/results/Pendulum-v0/20210428-092059/results/'
+	td3.load(cfg.model_path)
+	rewards,ma_rewards = eval(cfg.env,td3,cfg.seed)
+	make_dir(cfg.result_path,cfg.model_path)
+	save_results(rewards,ma_rewards,tag='eval',path=cfg.result_path)
+	plot_rewards(rewards,ma_rewards,tag="train",env=cfg.env,algo = cfg.algo,path=cfg.result_path)

codes/TD3/task1_train.py

@@ -0,0 +1,112 @@
+import sys,os
+curr_path = os.path.dirname(__file__)
+parent_path=os.path.dirname(curr_path) 
+sys.path.append(parent_path) # add current terminal path to sys.path
+
+import torch
+import gym
+import numpy as np
+import datetime
+
+
+from TD3.agent import TD3
+from common.plot import plot_rewards
+from common.utils import save_results,make_dir
+
+curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # obtain current time
+	
+
+class TD3Config:
+	def __init__(self) -> None:
+		self.algo = 'TD3'
+		self.env = 'Pendulum-v0'
+		self.seed = 0
+		self.result_path = curr_path+"/results/" +self.env+'/'+curr_time+'/results/'  # path to save results
+		self.model_path = curr_path+"/results/" +self.env+'/'+curr_time+'/models/'  # path to save models
+		self.start_timestep = 25e3 # Time steps initial random policy is used
+		self.start_ep = 50 # Episodes initial random policy is used
+		self.eval_freq = 10 # How often (episodes) we evaluate
+		self.train_eps = 600
+		self.max_timestep = 100000 # Max time steps to run environment
+		self.expl_noise = 0.1 # Std of Gaussian exploration noise
+		self.batch_size = 256 # Batch size for both actor and critic
+		self.gamma = 0.9 # gamma factor
+		self.lr = 0.0005 # Target network update rate 
+		self.policy_noise = 0.2 # Noise added to target policy during critic update
+		self.noise_clip = 0.3  # Range to clip target policy noise
+		self.policy_freq = 2 # Frequency of delayed policy updates
+		self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+# Runs policy for X episodes and returns average reward
+# A fixed seed is used for the eval environment
+def eval(env,agent, seed, eval_episodes=10):
+	eval_env = gym.make(env)
+	eval_env.seed(seed + 100)
+	avg_reward = 0.
+	for _ in range(eval_episodes):
+		state, done = eval_env.reset(), False
+		while not done:
+			# eval_env.render()
+			action = agent.choose_action(np.array(state))
+			state, reward, done, _ = eval_env.step(action)
+			avg_reward += reward
+	avg_reward /= eval_episodes
+	print("---------------------------------------")
+	print(f"Evaluation over {eval_episodes} episodes: {avg_reward:.3f}")
+	print("---------------------------------------")
+	return avg_reward
+
+def train(cfg,env,agent):
+	rewards = []
+	ma_rewards = [] # moveing average reward
+	for i_ep in range(int(cfg.train_eps)):
+		ep_reward = 0
+		ep_timesteps = 0
+		state, done = env.reset(), False
+		while not done:
+			ep_timesteps += 1
+			# Select action randomly or according to policy
+			if i_ep < cfg.start_ep:
+				action = env.action_space.sample()
+			else:
+				action = (
+					agent.choose_action(np.array(state))
+					+ np.random.normal(0, max_action * cfg.expl_noise, size=action_dim)
+				).clip(-max_action, max_action)
+			# Perform action
+			next_state, reward, done, _ = env.step(action) 
+			done_bool = float(done) if ep_timesteps < env._max_episode_steps else 0
+			# Store data in replay buffer
+			agent.memory.push(state, action, next_state, reward, done_bool)
+			state = next_state
+			ep_reward += reward
+			# Train agent after collecting sufficient data
+			if i_ep+1 >= cfg.start_ep:
+				agent.update()
+		print(f"Episode:{i_ep+1}/{cfg.train_eps}, Step:{ep_timesteps}, Reward:{ep_reward:.3f}")
+		rewards.append(ep_reward)
+		# 计算滑动窗口的reward
+		if ma_rewards:
+			ma_rewards.append(0.9*ma_rewards[-1]+0.1*ep_reward)
+		else:
+			ma_rewards.append(ep_reward) 	
+	return rewards, ma_rewards
+
+
+if __name__ == "__main__":
+	cfg  = TD3Config()
+	env = gym.make(cfg.env)
+	env.seed(cfg.seed) # Set seeds
+	torch.manual_seed(cfg.seed)
+	np.random.seed(cfg.seed)
+	state_dim = env.observation_space.shape[0]
+	action_dim = env.action_space.shape[0] 
+	max_action = float(env.action_space.high[0])
+	agent = TD3(state_dim,action_dim,max_action,cfg)
+	rewards,ma_rewards = train(cfg,env,agent)
+	make_dir(cfg.result_path,cfg.model_path)
+	agent.save(path=cfg.model_path)
+	save_results(rewards,ma_rewards,tag='train',path=cfg.result_path)
+	plot_rewards(rewards,ma_rewards,tag="train",env=cfg.env,algo = cfg.algo,path=cfg.result_path)
+
+		

codes/common/plot.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2020-10-07 20:57:11
 LastEditor: John
-LastEditTime: 2021-04-08 21:45:09
+LastEditTime: 2021-04-28 10:13:21
 Discription: 
 Environment: 
 '''
@@ -16,12 +16,21 @@ def plot_rewards(rewards,ma_rewards,tag="train",env='CartPole-v0',algo = "DQN",s
     plt.title("average learning curve of {} for {}".format(algo,env))
     plt.xlabel('epsiodes')
     plt.plot(rewards,label='rewards')
-    plt.plot(ma_rewards,label='moving average rewards')
+    plt.plot(ma_rewards,label='ma rewards')
     plt.legend()
     if save:
         plt.savefig(path+"rewards_curve_{}".format(tag))
     plt.show()
-
+# def plot_rewards(dic,tag="train",env='CartPole-v0',algo = "DQN",save=True,path='./'):
+#     sns.set()
+#     plt.title("average learning curve of {} for {}".format(algo,env))
+#     plt.xlabel('epsiodes')
+#     for key, value in dic.items():
+#         plt.plot(value,label=key)
+#     plt.legend()
+#     if save:
+#         plt.savefig(path+algo+"_rewards_curve_{}".format(tag))
+#     plt.show()
 def plot_losses(losses,algo = "DQN",save=True,path='./'):
     sns.set()
     plt.title("loss curve of {}".format(algo))