update

codes/A2C/.vscode/settings.json

@@ -1,3 +0,0 @@
-{
-    "python.pythonPath": "/Users/jj/anaconda3/envs/py37/bin/python"
-}

codes/A2C/agent.py

@@ -5,19 +5,18 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2020-11-03 20:47:09
 LastEditor: John
-LastEditTime: 2020-11-08 22:16:29
+LastEditTime: 2021-03-20 17:41:21
 Discription: 
 Environment: 
 '''
-from model import ActorCritic
+from A2C.model import ActorCritic
 import torch.optim as optim
 
 class A2C:
-    def __init__(self,n_states, n_actions, hidden_dim=256,device="cpu",lr = 3e-4):
-        self.device = device
+    def __init__(self,n_states, n_actions, cfg):
         self.gamma = 0.99
-        self.model = ActorCritic(n_states, n_actions, hidden_dim=hidden_dim).to(device)
-        self.optimizer = optim.Adam(self.model.parameters(),lr=lr)
+        self.model = ActorCritic(n_states, n_actions, hidden_dim=cfg.hidden_dim).to(cfg.device)
+        self.optimizer = optim.Adam(self.model.parameters(),lr=cfg.lr)
     def choose_action(self, state):
         dist, value = self.model(state)
         action = dist.sample()

codes/A2C/env.py

@@ -5,13 +5,13 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2020-10-30 15:39:37
 LastEditor: John
-LastEditTime: 2020-11-03 20:52:07
+LastEditTime: 2021-03-17 20:19:14
 Discription: 
 Environment: 
 '''
 
 import gym
-from common.multiprocessing_env import SubprocVecEnv
+from A2C.multiprocessing_env import SubprocVecEnv
 
 # num_envs = 16
 # env_name = "Pendulum-v0"

codes/A2C/main.py

@@ -5,94 +5,73 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-11 20:58:21
 @LastEditor: John
-LastEditTime: 2020-11-08 22:19:56
+LastEditTime: 2021-03-20 16:58:04
 @Discription: 
 @Environment: python 3.7.9
 '''
+import sys,os
+sys.path.append(os.getcwd()) # add current terminal path
 import torch
 import gym
-import os
-import numpy as np
-import argparse
-from torch.utils.tensorboard import SummaryWriter
-
-from agent import A2C
-from env import make_envs
-from utils import SEQUENCE, SAVED_MODEL_PATH, RESULT_PATH
-from utils import save_model,save_results
-
-def get_args():
-    '''模型建立好之后只需要在这里调参
-    '''
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--train", default=1, type=int)  # 1 表示训练，0表示只进行eval
-    parser.add_argument("--gamma", default=0.99,
-                        type=float)  # reward 折扣因子
-    parser.add_argument("--lr", default=3e-4, type=float)  # critic学习率
-    parser.add_argument("--actor_lr", default=1e-4, type=float)
-    parser.add_argument("--memory_capacity", default=10000,
-                        type=int, help="capacity of Replay Memory")
-    parser.add_argument("--batch_size", default=128, type=int,
-                        help="batch size of memory sampling")
-    parser.add_argument("--train_eps", default=4000, type=int)
-    parser.add_argument("--train_steps", default=5, type=int)
-    parser.add_argument("--eval_eps", default=200, type=int)  # 训练的最大episode数目
-    parser.add_argument("--eval_steps", default=200,
-                        type=int)  # 训练每个episode的长度
-    parser.add_argument("--target_update", default=4, type=int,
-                        help="when(every default 10 eisodes) to update target net ")
-    config = parser.parse_args()
-    return config
-
-def test_env(agent,device='cpu'):
-    env = gym.make("CartPole-v0")
-    state = env.reset()
-    ep_reward=0
-    for _ in range(200):
-        state = torch.FloatTensor(state).unsqueeze(0).to(device)
-        dist, value = agent.model(state)
-        action = dist.sample()
-        next_state, reward, done, _ = env.step(action.cpu().numpy()[0])
-        state = next_state
-        ep_reward += reward
-        if done:
-            break
-    return ep_reward
+import datetime
+from A2C.agent import A2C
 
 
-def train(cfg):
-    print('Start to train ! \n')
-    envs = make_envs(num_envs=16,env_name="CartPole-v0")
-    n_states = envs.observation_space.shape[0]
-    n_actions = envs.action_space.n
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-    agent = A2C(n_states, n_actions, hidden_dim=256)
-    # moving_average_rewards = []
-    # ep_steps = []
-    log_dir=os.path.split(os.path.abspath(__file__))[0]+"/logs/train/" + SEQUENCE
-    writer = SummaryWriter(log_dir)
-    state = envs.reset()
-    for i_episode in range(1, cfg.train_eps+1):
+
+SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # 获取当前时间
+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/"): 
+    os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/saved_model/")
+if not os.path.exists(SAVED_MODEL_PATH): 
+    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/"): 
+    os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/results/")
+if not os.path.exists(RESULT_PATH): 
+    os.mkdir(RESULT_PATH)
+
+class A2CConfig:
+    def __init__(self):
+        self.gamma = 0.99
+        self.lr = 3e-4 # learnning rate
+        self.actor_lr = 1e-4 # learnning rate of actor network
+        self.memory_capacity = 10000 # capacity of replay memory
+        self.batch_size = 128
+        self.train_eps = 200
+        self.train_steps = 200
+        self.eval_eps = 200
+        self.eval_steps = 200
+        self.target_update = 4
+        self.hidden_dim=256
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    
+def train(cfg,env,agent):
+    print('Start to train ! ')
+    for i_episode in range(cfg.train_eps):
+        state = env.reset()
         log_probs = []
         values    = []
         rewards = []
         masks     = []
         entropy = 0
-        for i_step in range(1, cfg.train_steps+1):
-            state = torch.FloatTensor(state).to(device)
+        ep_reward = 0
+        for i_step in range(cfg.train_steps):
+            state = torch.FloatTensor(state).to(cfg.device)
             dist, value = agent.model(state)
             action = dist.sample()
-            next_state, reward, done, _ = envs.step(action.cpu().numpy())
+            next_state, reward, done, _ = env.step(action.cpu().numpy())
+            ep_reward+=reward
             state = next_state                                                  
             log_prob = dist.log_prob(action)
             entropy += dist.entropy().mean()
             log_probs.append(log_prob)
             values.append(value)
-            rewards.append(torch.FloatTensor(reward).unsqueeze(1).to(device))
-            masks.append(torch.FloatTensor(1 - done).unsqueeze(1).to(device))
-        if i_episode%20 == 0:
-            print("reward",test_env(agent,device='cpu'))
-        next_state = torch.FloatTensor(next_state).to(device)
+            rewards.append(torch.FloatTensor(reward).unsqueeze(1).to(cfg.device))
+            masks.append(torch.FloatTensor(1 - done).unsqueeze(1).to(cfg.device))
+            if done:
+                break
+        print('Episode:{}/{}, Reward:{}, Steps:{}, Done:{}'.format(i_episode+1,cfg.train_eps,ep_reward,i_step+1,done))
+        next_state = torch.FloatTensor(next_state).to(cfg.device)
         _, next_value =agent.model(next_state)
         returns = agent.compute_returns(next_value, rewards, masks)
 
@@ -107,80 +86,17 @@ def train(cfg):
         agent.optimizer.zero_grad()
         loss.backward()
         agent.optimizer.step()
-    for _ in range(100):
-        print("test_reward",test_env(agent,device='cpu'))
-        
-        # print('Episode:', i_episode, ' Reward: %i' %
-        #       int(ep_reward[0]), 'n_steps:', i_step)
-        # ep_steps.append(i_step)
-        # rewards.append(ep_reward)
-        # if i_episode == 1:
-        #     moving_average_rewards.append(ep_reward[0])
-        # else:
-        #     moving_average_rewards.append(
-        #         0.9*moving_average_rewards[-1]+0.1*ep_reward[0])
-        # writer.add_scalars('rewards',{'raw':rewards[-1], 'moving_average': moving_average_rewards[-1]}, i_episode)
-        # writer.add_scalar('steps_of_each_episode',
-        #                   ep_steps[-1], i_episode)
-    writer.close()
+ 
     print('Complete training！')
-    ''' 保存模型 '''
-    # save_model(agent,model_path=SAVED_MODEL_PATH)
-    # '''存储reward等相关结果'''
-    # save_results(rewards,moving_average_rewards,ep_steps,tag='train',result_path=RESULT_PATH)
 
-# def eval(cfg, saved_model_path = SAVED_MODEL_PATH):
-#     print('start to eval ! \n')
-#     env = NormalizedActions(gym.make("Pendulum-v0"))
-#     n_states = env.observation_space.shape[0]
-#     n_actions = env.action_space.shape[0]
-#     agent = DDPG(n_states, n_actions, critic_lr=1e-3,
-#                  actor_lr=1e-4, gamma=0.99, soft_tau=1e-2, memory_capacity=100000, batch_size=128)
-#     agent.load_model(saved_model_path+'checkpoint.pth')
-#     rewards = []
-#     moving_average_rewards = []
-#     ep_steps = []
-#     log_dir=os.path.split(os.path.abspath(__file__))[0]+"/logs/eval/" + SEQUENCE
-#     writer = SummaryWriter(log_dir)
-#     for i_episode in range(1, cfg.eval_eps+1):
-#         state = env.reset()  # reset环境状态
-#         ep_reward = 0
-#         for i_step in range(1, cfg.eval_steps+1):
-#             action = agent.choose_action(state)  # 根据当前环境state选择action
-#             next_state, reward, done, _ = env.step(action)  # 更新环境参数
-#             ep_reward += reward
-#             state = next_state  # 跳转到下一个状态
-#             if done:
-#                 break
-#         print('Episode:', i_episode, ' Reward: %i' %
-#               int(ep_reward), 'n_steps:', i_step, 'done: ', done)
-#         ep_steps.append(i_step)
-#         rewards.append(ep_reward)
-#         # 计算滑动窗口的reward
-#         if i_episode == 1:
-#             moving_average_rewards.append(ep_reward)
-#         else:
-#             moving_average_rewards.append(
-#                 0.9*moving_average_rewards[-1]+0.1*ep_reward)
-#         writer.add_scalars('rewards',{'raw':rewards[-1], 'moving_average': moving_average_rewards[-1]}, i_episode)
-#         writer.add_scalar('steps_of_each_episode',
-#                           ep_steps[-1], i_episode)
-#     writer.close()
-#     '''存储reward等相关结果'''
-#     if not os.path.exists(RESULT_PATH): # 检测是否存在文件夹
-#         os.mkdir(RESULT_PATH)
-#     np.save(RESULT_PATH+'rewards_eval.npy', rewards)
-#     np.save(RESULT_PATH+'moving_average_rewards_eval.npy', moving_average_rewards)
-#     np.save(RESULT_PATH+'steps_eval.npy', ep_steps)
+
 
 if __name__ == "__main__":
-    cfg = get_args()
-    train(cfg)
-    
-    # cfg = get_args()
-    # if cfg.train:
-    #     train(cfg)
-    #     eval(cfg)
-    # else:
-    #     model_path = os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"
-    #     eval(cfg,saved_model_path=model_path)
+    cfg = A2CConfig()
+    env = gym.make('CartPole-v0')
+    env.seed(1) # set random seed for env
+    n_states = env.observation_space.shape[0]
+    n_actions = env.action_space.n
+    agent = A2C(n_states, n_actions, cfg)
+    train(cfg,env,agent)
+

codes/A2C/model.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2020-11-03 20:45:25
 LastEditor: John
-LastEditTime: 2020-11-07 18:49:09
+LastEditTime: 2021-03-20 17:41:33
 Discription: 
 Environment: 
 '''
@@ -13,7 +13,7 @@ import torch.nn as nn
 from torch.distributions import Categorical
 
 class ActorCritic(nn.Module):
-    def __init__(self, n_states, n_actions, hidden_dim=256, std=0.0):
+    def __init__(self, n_states, n_actions, hidden_dim=256):
         super(ActorCritic, self).__init__()
         self.critic = nn.Sequential(
             nn.Linear(n_states, hidden_dim),
@@ -30,6 +30,7 @@ class ActorCritic(nn.Module):
         
     def forward(self, x):
         value = self.critic(x)
+        print(x)
         probs = self.actor(x)
         dist  = Categorical(probs)
         return dist, value

codes/A2C/test.py

@@ -0,0 +1,162 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2021-03-20 17:43:17
+LastEditor: John
+LastEditTime: 2021-03-20 19:36:24
+Discription: 
+Environment: 
+'''
+import sys
+import torch  
+import gym
+import numpy as np  
+import torch.nn as nn
+import torch.optim as optim
+import torch.nn.functional as F
+from torch.autograd import Variable
+import matplotlib.pyplot as plt
+import pandas as pd
+
+
+learning_rate = 3e-4
+
+# Constants
+GAMMA = 0.99
+
+class A2CConfig:
+    ''' hyperparameters
+    '''
+    def __init__(self):
+        self.gamma = 0.99
+        self.lr = 3e-4 # learnning rate
+        self.actor_lr = 1e-4 # learnning rate of actor network
+        self.memory_capacity = 10000 # capacity of replay memory
+        self.batch_size = 128
+        self.train_eps = 3000
+        self.train_steps = 200
+        self.eval_eps = 200
+        self.eval_steps = 200
+        self.target_update = 4
+        self.hidden_dim=256
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    
+    
+class ActorCritic(nn.Module):
+    def __init__(self, n_states, n_actions, hidden_dim, learning_rate=3e-4):
+        super(ActorCritic, self).__init__()
+
+        self.n_actions = n_actions
+        self.critic_linear1 = nn.Linear(n_states, hidden_dim)
+        self.critic_linear2 = nn.Linear(hidden_dim, 1)
+
+        self.actor_linear1 = nn.Linear(n_states, hidden_dim)
+        self.actor_linear2 = nn.Linear(hidden_dim, n_actions)
+    
+    def forward(self, state):
+        state = Variable(torch.from_numpy(state).float().unsqueeze(0))
+        value = F.relu(self.critic_linear1(state))
+        value = self.critic_linear2(value)
+        policy_dist = F.relu(self.actor_linear1(state))
+        policy_dist = F.softmax(self.actor_linear2(policy_dist), dim=1)
+
+        return value, policy_dist
+
+class A2C:
+    def __init__(self,n_states,n_actions,cfg):
+        self.model = ActorCritic(n_states, n_actions, cfg.hidden_dim)
+        self.optimizer = optim.Adam(self.model.parameters(), lr=cfg.lr)
+    def choose_action(self,state):
+        pass
+    def update(self):
+        pass
+    
+def train(cfg,env,agent):
+    n_states = env.observation_space.shape[0]
+    n_actions = env.action_space.n
+    actor_critic = ActorCritic(n_states, n_actions, hidden_dim)
+    ac_optimizer = optim.Adam(actor_critic.parameters(), lr=learning_rate)
+
+    all_lengths = []
+    average_lengths = []
+    all_rewards = []
+    entropy_term = 0
+
+    for episode in range(cfg.train_eps):
+        log_probs = []
+        values = []
+        rewards = []
+        state = env.reset()
+        for steps in range(cfg.train_steps):
+            value, policy_dist = actor_critic.forward(state)
+            value = value.detach().numpy()[0,0]
+            dist = policy_dist.detach().numpy() 
+
+            action = np.random.choice(n_actions, p=np.squeeze(dist))
+            log_prob = torch.log(policy_dist.squeeze(0)[action])
+            entropy = -np.sum(np.mean(dist) * np.log(dist))
+            new_state, reward, done, _ = env.step(action)
+
+            rewards.append(reward)
+            values.append(value)
+            log_probs.append(log_prob)
+            entropy_term += entropy
+            state = new_state
+            
+            if done or steps == cfg.train_steps-1:
+                Qval, _ = actor_critic.forward(new_state)
+                Qval = Qval.detach().numpy()[0,0]
+                all_rewards.append(np.sum(rewards))
+                all_lengths.append(steps)
+                average_lengths.append(np.mean(all_lengths[-10:]))
+                if episode % 10 == 0:                    
+                    sys.stdout.write("episode: {}, reward: {}, total length: {}, average length: {} \n".format(episode, np.sum(rewards), steps, average_lengths[-1]))
+                break
+        
+        # compute Q values
+        Qvals = np.zeros_like(values)
+        for t in reversed(range(len(rewards))):
+            Qval = rewards[t] + GAMMA * Qval
+            Qvals[t] = Qval
+  
+        #update actor critic
+        values = torch.FloatTensor(values)
+        Qvals = torch.FloatTensor(Qvals)
+        log_probs = torch.stack(log_probs)
+        
+        advantage = Qvals - values
+        actor_loss = (-log_probs * advantage).mean()
+        critic_loss = 0.5 * advantage.pow(2).mean()
+        ac_loss = actor_loss + critic_loss + 0.001 * entropy_term
+
+        ac_optimizer.zero_grad()
+        ac_loss.backward()
+        ac_optimizer.step()
+
+        
+    
+    # Plot results
+    smoothed_rewards = pd.Series.rolling(pd.Series(all_rewards), 10).mean()
+    smoothed_rewards = [elem for elem in smoothed_rewards]
+    plt.plot(all_rewards)
+    plt.plot(smoothed_rewards)
+    plt.plot()
+    plt.xlabel('Episode')
+    plt.ylabel('Reward')
+    plt.show()
+
+    plt.plot(all_lengths)
+    plt.plot(average_lengths)
+    plt.xlabel('Episode')
+    plt.ylabel('Episode length')
+    plt.show()
+
+if __name__ == "__main__":
+    cfg = A2CConfig
+    env = gym.make("CartPole-v0")
+    n_states = env.observation_space.shape[0]
+    n_actions = env.action_space.n
+    agent = A2C(n_states,n_actions,cfg)
+    train(cfg,env,agent)    

codes/A2C/utils.py

@@ -15,7 +15,7 @@ import datetime
 
 SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
 SAVED_MODEL_PATH = os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"+SEQUENCE+'/'
-RESULT_PATH = os.path.split(os.path.abspath(__file__))[0]+"/result/"+SEQUENCE+'/'
+RESULT_PATH = os.path.split(os.path.abspath(__file__))[0]+"/results/"+SEQUENCE+'/'
 
 
 def save_results(rewards,moving_average_rewards,ep_steps,path=RESULT_PATH):

codes/DQN/main.py

@@ -5,7 +5,7 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-12 00:48:57
 @LastEditor: John
-LastEditTime: 2021-03-13 14:56:50
+LastEditTime: 2021-03-17 20:35:37
 @Discription: 
 @Environment: python 3.7.7
 '''
@@ -68,7 +68,7 @@ def train(cfg,env,agent):
         # 更新target network，复制DQN中的所有weights and biases
         if i_episode % cfg.target_update == 0:
             agent.target_net.load_state_dict(agent.policy_net.state_dict())
-        print('Episode:{}/{}, Reward:{}, Steps:{}, Done:{}'.format(i_episode+1,cfg.train_eps,ep_reward,i_step,done))
+        print('Episode:{}/{}, Reward:{}, Steps:{}, Done:{}'.format(i_episode+1,cfg.train_eps,ep_reward,i_step+1,done))
         ep_steps.append(i_step)
         rewards.append(ep_reward)
         # 计算滑动窗口的reward

codes/DoubleDQN/README.md

@@ -1,33 +0,0 @@
-## 思路
-
-见[博客](https://blog.csdn.net/JohnJim0/article/details/111552545)
-
-## 环境
-
-python 3.7.9
-
-pytorch 1.6.0
-
-tensorboard 2.3.0 
-
-torchvision 0.7.0 
-
-## 使用
-
-
-train: 
-
-```python
-python main.py 
-```
-
-eval: 
-
-```python
-python main.py --train 0 
-```
-可视化
-
-```python
-tensorboard --logdir logs 
-```

codes/DoubleDQN/agent.py

@@ -5,7 +5,7 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-12 00:50:49
 @LastEditor: John
-LastEditTime: 2020-12-22 16:20:35
+LastEditTime: 2021-03-13 15:01:27
 @Discription: 
 @Environment: python 3.7.7
 '''
@@ -20,65 +20,51 @@ import torch.nn.functional as F
 import random
 import math
 import numpy as np
-from memory import ReplayBuffer
-from model import FCN
-class DQN:
-    def __init__(self, n_states, n_actions, gamma=0.99, epsilon_start=0.9, epsilon_end=0.05, epsilon_decay=200, memory_capacity=10000, policy_lr=0.01, batch_size=128, device="cpu"):
-        self.actions_count = 0
+from common.memory import ReplayBuffer
+from common.model import MLP2
+class DoubleDQN:
+    def __init__(self, n_states, n_actions, cfg):
+        
         self.n_actions = n_actions  # 总的动作个数
-        self.device = device  # 设备，cpu或gpu等
-        self.gamma = gamma
+        self.device = cfg.device  # 设备，cpu或gpu等
+        self.gamma = cfg.gamma
         # e-greedy策略相关参数
-        self.epsilon = 0
-        self.epsilon_start = epsilon_start
-        self.epsilon_end = epsilon_end
-        self.epsilon_decay = epsilon_decay
-        self.batch_size = batch_size
-        self.policy_net = FCN(n_states, n_actions).to(self.device)
-        self.target_net = FCN(n_states, n_actions).to(self.device)
+        self.actions_count = 0
+        self.epsilon_start = cfg.epsilon_start
+        self.epsilon_end = cfg.epsilon_end
+        self.epsilon_decay = cfg.epsilon_decay
+        self.batch_size = cfg.batch_size
+        self.policy_net = MLP2(n_states, n_actions,hidden_dim=cfg.hidden_dim).to(self.device)
+        self.target_net = MLP2(n_states, n_actions,hidden_dim=cfg.hidden_dim).to(self.device)
         # target_net的初始模型参数完全复制policy_net
         self.target_net.load_state_dict(self.policy_net.state_dict())
         self.target_net.eval()  # 不启用 BatchNormalization 和 Dropout
         # 可查parameters()与state_dict()的区别，前者require_grad=True
-        self.optimizer = optim.Adam(self.policy_net.parameters(), lr=policy_lr)
+        self.optimizer = optim.Adam(self.policy_net.parameters(), lr=cfg.lr)
         self.loss = 0
-        self.memory = ReplayBuffer(memory_capacity)
+        self.memory = ReplayBuffer(cfg.memory_capacity)
 
-    def choose_action(self, state, train=True):
+    def choose_action(self, state):
         '''选择动作
         '''
-        if train:
-            self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * \
-                math.exp(-1. * self.actions_count / self.epsilon_decay)
-            self.actions_count += 1
-            if random.random() > self.epsilon:
-                with torch.no_grad():
-                    # 先转为张量便于丢给神经网络,state元素数据原本为float64
-                    # 注意state=torch.tensor(state).unsqueeze(0)跟state=torch.tensor([state])等价
-                    state = torch.tensor(
-                        [state], device=self.device, dtype=torch.float32)
-                    # 如tensor([[-0.0798, -0.0079]], grad_fn=<AddmmBackward>)
-                    q_value = self.policy_net(state)
-                    # tensor.max(1)返回每行的最大值以及对应的下标，
-                    # 如torch.return_types.max(values=tensor([10.3587]),indices=tensor([0]))
-                    # 所以tensor.max(1)[1]返回最大值对应的下标，即action
-                    action = q_value.max(1)[1].item()  
-            else:
-                action = random.randrange(self.n_actions)
-            return action
-        else: 
+        self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * \
+            math.exp(-1. * self.actions_count / self.epsilon_decay)
+        self.actions_count += 1
+        if random.random() > self.epsilon:
             with torch.no_grad():
-                    # 先转为张量便于丢给神经网络,state元素数据原本为float64
-                    # 注意state=torch.tensor(state).unsqueeze(0)跟state=torch.tensor([state])等价
-                    state = torch.tensor(
-                        [state], device='cpu', dtype=torch.float32)
-                    # 如tensor([[-0.0798, -0.0079]], grad_fn=<AddmmBackward>)
-                    q_value = self.target_net(state)
-                    # tensor.max(1)返回每行的最大值以及对应的下标，
-                    # 如torch.return_types.max(values=tensor([10.3587]),indices=tensor([0]))
-                    # 所以tensor.max(1)[1]返回最大值对应的下标，即action
-                    action = q_value.max(1)[1].item() 
-            return action
+                # 先转为张量便于丢给神经网络,state元素数据原本为float64
+                # 注意state=torch.tensor(state).unsqueeze(0)跟state=torch.tensor([state])等价
+                state = torch.tensor(
+                    [state], device=self.device, dtype=torch.float32)
+                # 如tensor([[-0.0798, -0.0079]], grad_fn=<AddmmBackward>)
+                q_value = self.policy_net(state)
+                # tensor.max(1)返回每行的最大值以及对应的下标，
+                # 如torch.return_types.max(values=tensor([10.3587]),indices=tensor([0]))
+                # 所以tensor.max(1)[1]返回最大值对应的下标，即action
+                action = q_value.max(1)[1].item()  
+        else:
+            action = random.randrange(self.n_actions)
+        return action
     def update(self):
 
         if len(self.memory) < self.batch_size:
@@ -86,8 +72,7 @@ class DQN:
         # 从memory中随机采样transition
         state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(
             self.batch_size)
-        # 转为张量
-        # 例如tensor([[-4.5543e-02, -2.3910e-01,  1.8344e-02,  2.3158e-01],...,[-1.8615e-02, -2.3921e-01, -1.1791e-02,  2.3400e-01]])
+        ### 转为张量 ###
         state_batch = torch.tensor(
             state_batch, device=self.device, dtype=torch.float)
         action_batch = torch.tensor(action_batch, device=self.device).unsqueeze(
@@ -96,6 +81,7 @@ class DQN:
             reward_batch, device=self.device, dtype=torch.float)  # tensor([1., 1.,...,1])
         next_state_batch = torch.tensor(
             next_state_batch, device=self.device, dtype=torch.float)
+        
         done_batch = torch.tensor(np.float32(
             done_batch), device=self.device).unsqueeze(1)  # 将bool转为float然后转为张量
 
@@ -112,7 +98,7 @@ class DQN:
         # 对于终止状态，此时done_batch[0]=1, 对应的expected_q_value等于reward
         q_target = reward_batch + self.gamma * next_q_state_value * (1-done_batch[0])
         '''
-        '''以下是Double DQNq_target计算方式，与NatureDQN稍有不同'''
+        '''以下是Double DQN q_target计算方式，与NatureDQN稍有不同'''
         next_target_values = self.target_net(
             next_state_batch)
         # 选出Q(s_t‘, a)对应的action，代入到next_target_values获得target net对应的next_q_value，即Q’(s_t|a=argmax Q(s_t‘, a))
@@ -127,8 +113,8 @@ class DQN:
             param.grad.data.clamp_(-1, 1)
         self.optimizer.step()  # 更新模型
 
-    def save_model(self,path):
-        torch.save(self.target_net.state_dict(), path)
+    def save(self,path):
+        torch.save(self.target_net.state_dict(), path+'DoubleDQN_checkpoint.pth')
 
-    def load_model(self,path):
-        self.target_net.load_state_dict(torch.load(path))  
+    def load(self,path):
+        self.target_net.load_state_dict(torch.load(path+'DoubleDQN_checkpoint.pth'))    

codes/DoubleDQN/main.py

@@ -5,37 +5,58 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-12 00:48:57
 @LastEditor: John
-LastEditTime: 2020-12-22 15:39:46
+LastEditTime: 2021-03-17 20:11:19
 @Discription: 
 @Environment: python 3.7.7
 '''
+import sys,os
+sys.path.append(os.getcwd()) # add current terminal path
 import gym
 import torch
-from torch.utils.tensorboard import SummaryWriter
-import os
-from agent import DQN
-from params import SEQUENCE,SAVED_MODEL_PATH,RESULT_PATH
-from params import get_args
-from utils import save_results
+import datetime
+from DoubleDQN.agent import DoubleDQN
+from common.plot import plot_rewards
+from common.utils import save_results
 
-def train(cfg):
+SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # 获取当前时间
+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/"): 
+    os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/saved_model/")
+if not os.path.exists(SAVED_MODEL_PATH): 
+    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/"): 
+    os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/results/")
+if not os.path.exists(RESULT_PATH): 
+    os.mkdir(RESULT_PATH)
+
+class DoubleDQNConfig:
+    def __init__(self):
+        self.algo = "Double DQN" # 算法名称
+        self.gamma = 0.99
+        self.epsilon_start = 0.9 # e-greedy策略的初始epsilon
+        self.epsilon_end = 0.01
+        self.epsilon_decay = 200
+        self.lr = 0.01 # 学习率
+        self.memory_capacity = 10000 # Replay Memory容量
+        self.batch_size = 128
+        self.train_eps = 250 # 训练的episode数目
+        self.train_steps = 200 # 训练每个episode的最大长度
+        self.target_update = 2 # target net的更新频率
+        self.eval_eps = 20 # 测试的episode数目
+        self.eval_steps = 200 # 测试每个episode的最大长度
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测gpu
+        self.hidden_dim = 128 # 神经网络隐藏层维度
+ 
+
+def train(cfg,env,agent):
     print('Start to train !')
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测gpu
-    env = gym.make('CartPole-v0').unwrapped # 可google为什么unwrapped gym，此处一般不需要
-    env.seed(1) # 设置env随机种子
-    n_states = env.observation_space.shape[0]
-    n_actions = env.action_space.n
-    agent = DQN(n_states=n_states, n_actions=n_actions, device=device, gamma=cfg.gamma, epsilon_start=cfg.epsilon_start,
-                epsilon_end=cfg.epsilon_end, epsilon_decay=cfg.epsilon_decay, policy_lr=cfg.policy_lr, memory_capacity=cfg.memory_capacity, batch_size=cfg.batch_size)
-    rewards = []
-    moving_average_rewards = []
+    rewards,ma_rewards = [],[]
     ep_steps = []
-    log_dir=os.path.split(os.path.abspath(__file__))[0]+"/logs/train/" + SEQUENCE
-    writer = SummaryWriter(log_dir)
-    for i_episode in range(1, cfg.train_eps+1):
+    for i_episode in range(cfg.train_eps):
         state = env.reset() # reset环境状态
         ep_reward = 0
-        for i_step in range(1, cfg.train_steps+1):
+        for i_step in range(cfg.train_steps):
             action = agent.choose_action(state) # 根据当前环境state选择action
             next_state, reward, done, _ = env.step(action) # 更新环境参数
             ep_reward += reward
@@ -47,80 +68,26 @@ def train(cfg):
         # 更新target network，复制DQN中的所有weights and biases
         if i_episode % cfg.target_update == 0:
             agent.target_net.load_state_dict(agent.policy_net.state_dict())
-        print('Episode:', i_episode, ' Reward: %i' %
-              int(ep_reward), 'n_steps:', i_step, 'done: ', done,' Explore: %.2f' % agent.epsilon)
+        print('Episode:{}/{}, Reward:{}, Steps:{}, Done:{}'.format(i_episode+1,cfg.train_eps,ep_reward,i_step,done))
         ep_steps.append(i_step)
         rewards.append(ep_reward)
         # 计算滑动窗口的reward
-        if i_episode == 1:
-            moving_average_rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(
+                0.9*ma_rewards[-1]+0.1*ep_reward)
         else:
-            moving_average_rewards.append(
-                0.9*moving_average_rewards[-1]+0.1*ep_reward)
-        writer.add_scalars('rewards',{'raw':rewards[-1], 'moving_average': moving_average_rewards[-1]}, i_episode)
-        writer.add_scalar('steps_of_each_episode',
-                          ep_steps[-1], i_episode)
-    writer.close()
+            ma_rewards.append(ep_reward)   
     print('Complete training！')
-    ''' 保存模型 '''
-    if not os.path.exists(SAVED_MODEL_PATH): # 检测是否存在文件夹
-        os.mkdir(SAVED_MODEL_PATH)
-    agent.save_model(SAVED_MODEL_PATH+'checkpoint.pth')
-    print('model saved！')
-    '''存储reward等相关结果'''
-    save_results(rewards,moving_average_rewards,ep_steps,tag='train',result_path=RESULT_PATH)
+    return rewards,ma_rewards
     
-
-def eval(cfg, saved_model_path = SAVED_MODEL_PATH):
-    print('start to eval !')
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测gpu
+if __name__ == "__main__":
+    cfg = DoubleDQNConfig()
     env = gym.make('CartPole-v0').unwrapped # 可google为什么unwrapped gym，此处一般不需要
     env.seed(1) # 设置env随机种子
     n_states = env.observation_space.shape[0]
     n_actions = env.action_space.n
-    agent = DQN(n_states=n_states, n_actions=n_actions, device=device, gamma=cfg.gamma, epsilon_start=cfg.epsilon_start,
-                epsilon_end=cfg.epsilon_end, epsilon_decay=cfg.epsilon_decay, policy_lr=cfg.policy_lr, memory_capacity=cfg.memory_capacity, batch_size=cfg.batch_size)
-    agent.load_model(saved_model_path+'checkpoint.pth')
-    rewards = []
-    moving_average_rewards = []
-    ep_steps = []
-    log_dir=os.path.split(os.path.abspath(__file__))[0]+"/logs/eval/" + SEQUENCE
-    writer = SummaryWriter(log_dir)
-    for i_episode in range(1, cfg.eval_eps+1):
-        state = env.reset()  # reset环境状态
-        ep_reward = 0
-        for i_step in range(1, cfg.eval_steps+1):
-            action = agent.choose_action(state,train=False)  # 根据当前环境state选择action
-            next_state, reward, done, _ = env.step(action)  # 更新环境参数
-            ep_reward += reward
-            state = next_state  # 跳转到下一个状态
-            if done:
-                break
-        print('Episode:', i_episode, ' Reward: %i' %
-              int(ep_reward), 'n_steps:', i_step, 'done: ', done)
-              
-        ep_steps.append(i_step)
-        rewards.append(ep_reward)
-        # 计算滑动窗口的reward
-        if i_episode == 1:
-            moving_average_rewards.append(ep_reward)
-        else:
-            moving_average_rewards.append(
-                0.9*moving_average_rewards[-1]+0.1*ep_reward)
-                
-        writer.add_scalars('rewards',{'raw':rewards[-1], 'moving_average': moving_average_rewards[-1]}, i_episode)
-        writer.add_scalar('steps_of_each_episode',
-                          ep_steps[-1], i_episode)
-    writer.close()
-    '''存储reward等相关结果'''
-    save_results(rewards,moving_average_rewards,ep_steps,tag='eval',result_path=RESULT_PATH)
-    print('Complete evaling！')
-    
-if __name__ == "__main__":
-    cfg = get_args()
-    if cfg.train:
-        train(cfg)
-        eval(cfg)
-    else:
-        model_path = os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"
-        eval(cfg,saved_model_path=model_path)
+    agent = DoubleDQN(n_states,n_actions,cfg)
+    rewards,ma_rewards = train(cfg,env,agent)
+    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)

codes/DoubleDQN/memory.py

@@ -5,12 +5,11 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-10 15:27:16
 @LastEditor: John
-LastEditTime: 2020-12-22 12:56:27
+LastEditTime: 2021-01-20 18:58:37
 @Discription: 
 @Environment: python 3.7.7
 '''
 import random
-import numpy as np
 
 class ReplayBuffer:
     

codes/DoubleDQN/model.py

@@ -12,13 +12,13 @@ LastEditTime: 2020-08-19 16:55:54
 import torch.nn as nn
 import torch.nn.functional as F
 
-class FCN(nn.Module):
+class MLP(nn.Module):
     def __init__(self, n_states=4, n_actions=18):
         """ 初始化q网络，为全连接网络
             n_states: 输入的feature即环境的state数目
             n_actions: 输出的action总个数
         """
-        super(FCN, self).__init__()
+        super(MLP, self).__init__()
         self.fc1 = nn.Linear(n_states, 128) # 输入层
         self.fc2 = nn.Linear(128, 128) # 隐藏层
         self.fc3 = nn.Linear(128, n_actions) # 输出层

codes/DoubleDQN/params.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2020-12-22 15:22:17
 LastEditor: John
-LastEditTime: 2020-12-22 15:26:09
+LastEditTime: 2021-01-21 14:30:38
 Discription: 
 Environment: 
 '''
@@ -16,7 +16,10 @@ import argparse
 ALGO_NAME = 'Double DQN'
 SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
 SAVED_MODEL_PATH = os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"+SEQUENCE+'/'
-RESULT_PATH = os.path.split(os.path.abspath(__file__))[0]+"/result/"+SEQUENCE+'/'
+RESULT_PATH = os.path.split(os.path.abspath(__file__))[0]+"/results/"+SEQUENCE+'/'
+
+TRAIN_LOG_DIR=os.path.split(os.path.abspath(__file__))[0]+"/logs/train/" + SEQUENCE
+EVAL_LOG_DIR=os.path.split(os.path.abspath(__file__))[0]+"/logs/eval/" + SEQUENCE
 
 def get_args():
     '''模型参数

codes/DoubleDQN/plot.py

@@ -24,14 +24,14 @@ def plot(item,ylabel='rewards_train', save_fig = True):
     plt.ylabel(ylabel)
     plt.xlabel('episodes')
     if save_fig:
-        plt.savefig(os.path.dirname(__file__)+"/result/"+ylabel+".png")
+        plt.savefig(os.path.dirname(__file__)+"/results/"+ylabel+".png")
     plt.show()
 
 
     # plt.show()
 if __name__ == "__main__":
 
-    output_path = os.path.split(os.path.abspath(__file__))[0]+"/result/"
+    output_path = os.path.split(os.path.abspath(__file__))[0]+"/results/"
     tag = 'train'
     rewards=np.load(output_path+"rewards_"+tag+".npy", )
     moving_average_rewards=np.load(output_path+"moving_average_rewards_"+tag+".npy",)

codes/DoubleDQN/utils.py

@@ -13,7 +13,7 @@ import os
 import numpy as np
 
 
-def save_results(rewards,moving_average_rewards,ep_steps,tag='train',result_path='./result'):
+def save_results(rewards,moving_average_rewards,ep_steps,tag='train',result_path='./results'):
     if not os.path.exists(result_path): # 检测是否存在文件夹
         os.mkdir(result_path)
     np.save(result_path+'rewards_'+tag+'.npy', rewards)

codes/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020 John Jim
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

codes/MonteCarlo/README.md

@@ -2,10 +2,10 @@
 
 ## 环境说明
 
-见[环境说明](https://github.com/datawhalechina/leedeeprl-notes/blob/master/codes/env_info.md)中的The Racetrack
+见[环境说明](https://github.com/JohnJim0816/reinforcement-learning-tutorials/blob/master/env_info.md)中的The Racetrack
 
 ## First-Visit MC 介绍
 
-伪代码：
+### 伪代码
 
 ![mc_control_algo](assets/mc_control_algo.png)

codes/MonteCarlo/agent.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2021-03-12 16:14:34
 LastEditor: John
-LastEditTime: 2021-03-12 16:15:12
+LastEditTime: 2021-03-17 12:35:06
 Discription: 
 Environment: 
 '''
@@ -26,11 +26,13 @@ class FisrtVisitMC:
         
     def choose_action(self,state):
         ''' e-greed policy '''
-        best_action = np.argmax(self.Q[state])
-        # action = best_action
-        action_probs = np.ones(self.n_actions, dtype=float) * self.epsilon / self.n_actions
-        action_probs[best_action] += (1.0 - self.epsilon)
-        action = np.random.choice(np.arange(len(action_probs)), p=action_probs)
+        if state in self.Q.keys():
+            best_action = np.argmax(self.Q[state])
+            action_probs = np.ones(self.n_actions, dtype=float) * self.epsilon / self.n_actions
+            action_probs[best_action] += (1.0 - self.epsilon)
+            action = np.random.choice(np.arange(len(action_probs)), p=action_probs)
+        else:
+            action = np.random.randint(0,self.n_actions)
         return action
     def update(self,one_ep_transition):
         # Find all (state, action) pairs we've visited in this one_ep_transition

codes/MonteCarlo/main.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2021-03-11 14:26:44
 LastEditor: John
-LastEditTime: 2021-03-12 16:15:46
+LastEditTime: 2021-03-17 12:35:36
 Discription: 
 Environment: 
 '''
@@ -35,7 +35,7 @@ class MCConfig:
     def __init__(self): 
         self.epsilon = 0.15 # epsilon: The probability to select a random action . 
         self.gamma = 0.9 # gamma: Gamma discount factor.
-        self.n_episodes = 300
+        self.n_episodes = 150
         self.n_steps = 2000
 
 def get_mc_args():
@@ -58,8 +58,8 @@ def mc_train(cfg,env,agent):
         one_ep_transition = []
         state = env.reset()
         ep_reward = 0
-        # while True:
-        for t in range(cfg.n_steps):
+        while True:
+        # for t in range(cfg.n_steps):
             action = agent.choose_action(state)
             next_state, reward, done = env.step(action)
             ep_reward+=reward

codes/PolicyGradient/README.md

@@ -1,38 +1,15 @@
 # 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
-
-python 3.7.9
-
-pytorch 1.6.0
-
-tensorboard 2.3.0 
-
-torchvision 0.7.0 
-
+python 3.7.9、pytorch 1.6.0
 ## 程序运行方法
 
-train: 
-
-```python
-python main.py 
-```
-
-eval: 
-
-```python
-python main.py --train 0 
-```
-tensorboard：
-```python
-tensorboard --logdir logs 
-```
-
-
 ## 参考
 
 [REINFORCE和Reparameterization Trick](https://blog.csdn.net/JohnJim0/article/details/110230703)

codes/PolicyGradient/agent.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2020-11-22 23:27:44
 LastEditor: John
-LastEditTime: 2020-11-23 17:04:37
+LastEditTime: 2021-03-13 11:50:16
 Discription: 
 Environment: 
 '''
@@ -14,24 +14,23 @@ from torch.distributions import Bernoulli
 from torch.autograd import Variable
 import numpy as np
 
-from model import FCN
+from common.model import MLP1
 
 class PolicyGradient:
     
-    def __init__(self, state_dim,device='cpu',gamma = 0.99,lr = 0.01,batch_size=5):
-        self.gamma = gamma
-        self.policy_net = FCN(state_dim)
-        self.optimizer = torch.optim.RMSprop(self.policy_net.parameters(), lr=lr)
-        self.batch_size = batch_size
+    def __init__(self, n_states,cfg):
+        self.gamma = cfg.gamma
+        self.policy_net = MLP1(n_states,hidden_dim=cfg.hidden_dim)
+        self.optimizer = torch.optim.RMSprop(self.policy_net.parameters(), lr=cfg.lr)
+        self.batch_size = cfg.batch_size
 
     def choose_action(self,state):
         
         state = torch.from_numpy(state).float()
         state = Variable(state)
         probs = self.policy_net(state)
-        m = Bernoulli(probs)
+        m = Bernoulli(probs) # 伯努利分布
         action = m.sample()
-
         action = action.data.numpy().astype(int)[0] # 转为标量
         return action
         
@@ -67,6 +66,6 @@ class PolicyGradient:
             loss.backward()
         self.optimizer.step()
     def save_model(self,path):
-        torch.save(self.policy_net.state_dict(), path)
+        torch.save(self.policy_net.state_dict(), path+'pg_checkpoint.pth')
     def load_model(self,path):
-        self.policy_net.load_state_dict(torch.load(path)) 
+        self.policy_net.load_state_dict(torch.load(path+'pg_checkpoint.pth')) 

codes/PolicyGradient/env.py

@@ -1,19 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-Author: John
-Email: johnjim0816@gmail.com
-Date: 2020-11-22 23:23:10
-LastEditor: John
-LastEditTime: 2020-11-23 11:55:24
-Discription: 
-Environment: 
-'''
-import gym
-
-def env_init():
-    env = gym.make('CartPole-v0') # 可google为什么unwrapped gym，此处一般不需要
-    env.seed(1) # 设置env随机种子
-    state_dim = env.observation_space.shape[0]
-    n_actions = env.action_space.n
-    return env,state_dim,n_actions

codes/PolicyGradient/main.py

@@ -5,34 +5,47 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2020-11-22 23:21:53
 LastEditor: John
-LastEditTime: 2020-11-24 19:52:40
+LastEditTime: 2021-03-13 11:50:32
 Discription: 
 Environment: 
 '''
+import sys,os
+sys.path.append(os.getcwd()) # 添加当前终端路径
 from itertools import count
-import torch
-import os
-from torch.utils.tensorboard import SummaryWriter
+import datetime
+import gym
+from PolicyGradient.agent import PolicyGradient
+from common.plot import plot_rewards
+from common.utils import save_results
 
-from env import env_init
-from params import get_args
-from agent import PolicyGradient
-from params import SEQUENCE, SAVED_MODEL_PATH, RESULT_PATH
-from utils import save_results,save_model
-from plot import plot
-def train(cfg):
-    env,state_dim,n_actions = env_init()
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测gpu
-    agent  = PolicyGradient(state_dim,device = device,lr = cfg.policy_lr)
+SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # 获取当前时间
+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 PGConfig:
+    def __init__(self):
+        self.train_eps = 300 # 训练的episode数目
+        self.batch_size = 8
+        self.lr = 0.01 # 学习率
+        self.gamma = 0.99
+        self.hidden_dim = 36 # 隐藏层维度
+        
+def train(cfg,env,agent):
     '''下面带pool都是存放的transition序列用于gradient'''
     state_pool = [] # 存放每batch_size个episode的state序列
     action_pool = []
     reward_pool = [] 
     ''' 存储每个episode的reward用于绘图'''
     rewards = []
-    moving_average_rewards = []
-    log_dir=os.path.split(os.path.abspath(__file__))[0]+"/logs/train/" + SEQUENCE
-    writer = SummaryWriter(log_dir) # 使用tensorboard的writer
+    ma_rewards = []
     for i_episode in range(cfg.train_eps):
         state = env.reset()
         ep_reward = 0
@@ -55,55 +68,22 @@ def train(cfg):
             action_pool = []
             reward_pool = []
         rewards.append(ep_reward)
-        if i_episode == 0:
-            moving_average_rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(
+                0.9*ma_rewards[-1]+0.1*ep_reward)
         else:
-            moving_average_rewards.append(
-                0.9*moving_average_rewards[-1]+0.1*ep_reward)
-        writer.add_scalars('rewards',{'raw':rewards[-1], 'moving_average': moving_average_rewards[-1]}, i_episode+1)
-    writer.close()
-    print('Complete training！')
-    save_model(agent,model_path=SAVED_MODEL_PATH)
-    '''存储reward等相关结果'''
-    save_results(rewards,moving_average_rewards,tag='train',result_path=RESULT_PATH)
-    plot(rewards)
-    plot(moving_average_rewards,ylabel='moving_average_rewards_train')
-        
-def eval(cfg,saved_model_path = SAVED_MODEL_PATH):
-    env,state_dim,n_actions = env_init()
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测gpu
-    agent  = PolicyGradient(state_dim,device = device,lr = cfg.policy_lr)
-    agent.load_model(saved_model_path+'checkpoint.pth')
-    rewards = []
-    moving_average_rewards = []
-    log_dir=os.path.split(os.path.abspath(__file__))[0]+"/logs/eval/" + SEQUENCE
-    writer = SummaryWriter(log_dir) # 使用tensorboard的writer
-    for i_episode in range(cfg.eval_eps):
-        state = env.reset()
-        ep_reward = 0
-        for _ in count():
-            action = agent.choose_action(state) # 根据当前环境state选择action
-            next_state, reward, done, _ = env.step(action)
-            ep_reward += reward     
-            state = next_state
-            if done:
-                print('Episode:', i_episode, ' Reward:',  ep_reward)
-                break
-        rewards.append(ep_reward)
-        if i_episode == 0:
-            moving_average_rewards.append(ep_reward)
-        else:
-            moving_average_rewards.append(
-                0.9*moving_average_rewards[-1]+0.1*ep_reward)
-        writer.add_scalars('rewards',{'raw':rewards[-1], 'moving_average': moving_average_rewards[-1]}, i_episode+1)
-    writer.close()
-    print('Complete evaling！')
-    
+            ma_rewards.append(ep_reward)
+    print('complete training！')
+    return rewards, ma_rewards
+            
 if __name__ == "__main__":
-    cfg = get_args()
-    if cfg.train:
-        train(cfg)
-        eval(cfg)
-    else:
-        model_path = os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"
-        eval(cfg,saved_model_path=model_path)
+    cfg = PGConfig()
+    env = gym.make('CartPole-v0') # 可google为什么unwrapped gym，此处一般不需要
+    env.seed(1) # 设置env随机种子
+    n_states = env.observation_space.shape[0]
+    n_actions = env.action_space.n
+    agent  = PolicyGradient(n_states,cfg)
+    rewards, ma_rewards = train(cfg,env,agent)
+    agent.save_model(SAVED_MODEL_PATH)
+    save_results(rewards,ma_rewards,tag='train',path=RESULT_PATH)
+    plot_rewards(rewards,ma_rewards,tag="train",algo = "Policy Gradient",path=RESULT_PATH)

codes/PolicyGradient/model.py

@@ -1,27 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-Author: John
-Email: johnjim0816@gmail.com
-Date: 2020-11-22 23:18:46
-LastEditor: John
-LastEditTime: 2020-11-27 16:55:25
-Discription: 
-Environment: 
-'''
-import torch.nn as nn
-import torch.nn.functional as F
-class FCN(nn.Module):
-    ''' 全连接网络'''
-    def __init__(self,state_dim):
-        super(FCN, self).__init__()
-        # 24和36为hidden layer的层数，可根据state_dim, n_actions的情况来改变
-        self.fc1 = nn.Linear(state_dim, 36)
-        self.fc2 = nn.Linear(36, 36)
-        self.fc3 = nn.Linear(36, 1)  # Prob of Left
-
-    def forward(self, x):
-        x = F.relu(self.fc1(x))
-        x = F.relu(self.fc2(x))
-        x = F.sigmoid(self.fc3(x))
-        return x

codes/PolicyGradient/params.py

@@ -1,29 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-Author: John
-Email: johnjim0816@gmail.com
-Date: 2020-11-22 23:25:37
-LastEditor: John
-LastEditTime: 2020-11-26 19:11:21
-Discription: 存储参数
-Environment: 
-'''
-import argparse
-import datetime
-import os
-
-SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
-SAVED_MODEL_PATH = os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"+SEQUENCE+'/'
-RESULT_PATH = os.path.split(os.path.abspath(__file__))[0]+"/result/"+SEQUENCE+'/'
-
-def get_args():
-    '''训练参数'''
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--train", default=1, type=int)  # 1 表示训练，0表示只进行eval
-    parser.add_argument("--train_eps", default=300, type=int) # 训练的最大episode数目
-    parser.add_argument("--eval_eps", default=100, type=int)  # 训练的最大episode数目
-    parser.add_argument("--batch_size", default=4, type=int) # 用于gradient的episode数目
-    parser.add_argument("--policy_lr", default=0.01, type=float) # 学习率
-    config = parser.parse_args()
-    return config

codes/PolicyGradient/plot.py

@@ -1,46 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-Author: John
-Email: johnjim0816@gmail.com
-Date: 2020-11-23 13:48:46
-LastEditor: John
-LastEditTime: 2020-11-23 13:48:48
-Discription: 
-Environment: 
-'''
-import matplotlib.pyplot as plt
-import seaborn as sns
-import numpy as np
-import os 
-
-def plot(item,ylabel='rewards_train', save_fig = True):
-    '''plot using searborn to plot 
-    '''
-    sns.set()
-    plt.figure()
-    plt.plot(np.arange(len(item)), item)
-    plt.title(ylabel+' of DQN') 
-    plt.ylabel(ylabel)
-    plt.xlabel('episodes')
-    if save_fig:
-        plt.savefig(os.path.dirname(__file__)+"/result/"+ylabel+".png")
-    plt.show()
-
-if __name__ == "__main__":
-
-    output_path = os.path.split(os.path.abspath(__file__))[0]+"/result/"
-    tag = 'train'
-    rewards=np.load(output_path+"rewards_"+tag+".npy", )
-    moving_average_rewards=np.load(output_path+"moving_average_rewards_"+tag+".npy",)
-    steps=np.load(output_path+"steps_"+tag+".npy")
-    plot(rewards)
-    plot(moving_average_rewards,ylabel='moving_average_rewards_'+tag)
-    plot(steps,ylabel='steps_'+tag)
-    tag = 'eval'
-    rewards=np.load(output_path+"rewards_"+tag+".npy", )
-    moving_average_rewards=np.load(output_path+"moving_average_rewards_"+tag+".npy",)
-    steps=np.load(output_path+"steps_"+tag+".npy")
-    plot(rewards,ylabel='rewards_'+tag)
-    plot(moving_average_rewards,ylabel='moving_average_rewards_'+tag)
-    plot(steps,ylabel='steps_'+tag)