update

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)