add dqn

codes/dqn/dqn.py

@@ -5,12 +5,16 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-12 00:50:49
 @LastEditor: John
-@LastEditTime: 2020-06-14 13:56:45
+LastEditTime: 2020-08-22 15:44:31
 @Discription: 
 @Environment: python 3.7.7
 '''
 '''off-policy
 '''
+
+
+
+
 import torch
 import torch.nn as nn
 import torch.optim as optim
@@ -20,79 +24,97 @@ 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"):
+    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
-        self.n_actions = n_actions
-        self.device = device
+        self.n_actions = n_actions  # 总的动作个数
+        self.device = device  # 设备，cpu或gpu等
         self.gamma = 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.policy_net = FCN(n_states, n_actions).to(self.device)
+        self.target_net = FCN(n_states, n_actions).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
-        self.optimizer = optim.Adam(self.policy_net.parameters(),lr=policy_lr)
+        # 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.loss = 0
         self.memory = ReplayBuffer(memory_capacity)
 
-    def select_action(self,state):
-        '''选择工作
+    def select_action(self, state):
+        '''选择动作
         Args:
-            state [array]: 状态
+            state [array]: [description]
         Returns:
-            [array]: 动作
+            action [array]: [description]
         '''
         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():
-                state   = torch.tensor([state],device=self.device,dtype=torch.float32) # 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
-                q_value = self.policy_net(state) # tensor([[-0.0798, -0.0079]], grad_fn=<AddmmBackward>)
-                action  = q_value.max(1)[1].item()
+                # 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:
             return
+        # 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]])
+        state_batch = torch.tensor(
+            state_batch, device=self.device, dtype=torch.float)
+        action_batch = torch.tensor(action_batch, device=self.device).unsqueeze(
+            1)  # 例如tensor([[1],...,[0]])
+        reward_batch = torch.tensor(
+            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
 
-        state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(self.batch_size)
-        
-        state_batch = torch.tensor(state_batch,device=self.device,dtype=torch.float) # 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]])
-        action_batch  = torch.tensor(action_batch,device=self.device).unsqueeze(1) # 例如tensor([[1],...,[0]])
-        reward_batch  = torch.tensor(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
-        # Compute Q(s_t, a) - the model computes Q(s_t), then we select the
-        # columns of actions taken. These are the actions which would've been taken
-        # for each batch state according to policy_net
-        q_values = self.policy_net(state_batch).gather(1, action_batch) # 等价于self.forward       
-        # Compute V(s_{t+1}) for all next states.
-        # Expected values of actions for non_final_next_states are computed based
-        # on the "older" target_net; selecting their best reward with max(1)[0].
-        # This is merged based on the mask, such that we'll have either the expected
-        # state value or 0 in case the state was final.
-     
+        # 计算当前(s_t,a)对应的Q(s_t, a)
+        # 关于torch.gather,对于a=torch.Tensor([[1,2],[3,4]])
+        # 那么a.gather(1,torch.Tensor([[0],[1]]))=torch.Tensor([[1],[3]])
+        q_values = self.policy_net(state_batch).gather(
+            dim=1, index=action_batch)  # 等价于self.forward
+        # 计算所有next states的V(s_{t+1})，即通过target_net中选取reward最大的对应states
         next_state_values = self.target_net(
-            next_state_batch).max(1)[0].detach() # tensor([ 0.0060, -0.0171,...,])
-        # Compute the expected Q values
-        expected_q_values = reward_batch + self.gamma * next_state_values * (1-done_batch[0])
-
-        # Compute Huber loss
-        # self.loss = nn.MSELoss(q_values, expected_q_values.unsqueeze(1)) 
-        self.loss = nn.MSELoss()(q_values,expected_q_values.unsqueeze(1))
-        # Optimize the model
-        self.optimizer.zero_grad() # zero_grad clears old gradients from the last step (otherwise you’d just accumulate the gradients from all loss.backward() calls).
-        self.loss.backward() # loss.backward() computes the derivative of the loss w.r.t. the parameters (or anything requiring gradients) using backpropagation.
-        for param in self.policy_net.parameters(): # clip防止梯度爆炸
+            next_state_batch).max(1)[0].detach()  # 比如tensor([ 0.0060, -0.0171,...,])
+        # 计算 expected_q_value
+        # 对于终止状态，此时done_batch[0]=1, 对应的expected_q_value等于reward
+        expected_q_values = reward_batch + self.gamma * \
+            next_state_values * (1-done_batch[0])
+        # self.loss = F.smooth_l1_loss(q_values,expected_q_values.unsqueeze(1)) # 计算 Huber loss
+        self.loss = nn.MSELoss()(q_values, expected_q_values.unsqueeze(1))  # 计算 均方误差loss
+        # 优化模型
+        self.optimizer.zero_grad()  # zero_grad清除上一步所有旧的gradients from the last step
+        # loss.backward()使用backpropagation计算loss相对于所有parameters(需要gradients)的微分
+        self.loss.backward()
+        for param in self.policy_net.parameters():  # clip防止梯度爆炸
             param.grad.data.clamp_(-1, 1)
-        self.optimizer.step() # causes the optimizer to take a step based on the gradients of the parameters.
+        self.optimizer.step()  # 更新模型
         
-
+    def save_model():
+        pass
+    def load_model():
+        pass

codes/dqn/main.py

@@ -5,7 +5,7 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-12 00:48:57
 @LastEditor: John
-@LastEditTime: 2020-07-20 23:02:16
+LastEditTime: 2020-08-22 18:02:56
 @Discription: 
 @Environment: python 3.7.7
 '''
@@ -14,26 +14,27 @@ import torch
 from dqn import DQN
 from plot import plot
 import argparse
+
 def get_args():
-    '''模型建立好之后只需要在这里调参
+    '''模型参数
     '''
     parser = argparse.ArgumentParser()
-
     parser.add_argument("--gamma", default=0.99,
                         type=float)  # q-learning中的gamma
     parser.add_argument("--epsilon_start", default=0.95,
                         type=float)  # 基于贪心选择action对应的参数epsilon
     parser.add_argument("--epsilon_end", default=0.01, type=float)
-    parser.add_argument("--epsilon_decay", default=200, type=float)
+    parser.add_argument("--epsilon_decay", default=500, type=float)
     parser.add_argument("--policy_lr", default=0.01, type=float)
     parser.add_argument("--memory_capacity", default=1000,
-                        type=int, help="capacity of Replay Memory")
+                        type=int, help="capacity of Replay Memory") 
 
     parser.add_argument("--batch_size", default=32, type=int,
                         help="batch size of memory sampling")
-    parser.add_argument("--max_episodes", default=200, type=int)
+    parser.add_argument("--max_episodes", default=200, type=int) # 训练的最大episode数目
     parser.add_argument("--max_steps", default=200, type=int)
-    parser.add_argument("--target_update", default=2, type=int,
+    # 将目标网络的更新频率改为1就是普通的dqn，大于1就是double dqn
+    parser.add_argument("--target_update", default=1, type=int,
                         help="when(every default 10 eisodes) to update target net ")
     config = parser.parse_args()
 
@@ -44,38 +45,34 @@ if __name__ == "__main__":
 
     cfg = get_args()
     # if gpu is to be used
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-    env = gym.make('CartPole-v0').unwrapped
-    env.seed(1)
+    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)
+                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 = []
+    ep_steps = []
     for i_episode in range(1, cfg.max_episodes+1):
-        # Initialize the environment and state
-        state = env.reset()
+        state = env.reset() # reset环境状态
         ep_reward = 0
-        for t in range(1, cfg.max_steps+1):
-            # Select and perform an action
-            action = agent.select_action(state)
-            next_state, reward, done, _ = env.step(action)
+        for i_step in range(1, cfg.max_steps+1):
+            action = agent.select_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
-            # Store the transition in memory
-            agent.memory.push(state,action,reward,next_state,done)
-            # Move to the next state
-            state = next_state
-            # Perform one step of the optimization (on the target network)
-            agent.update()
+            agent.memory.push(state, action, reward, next_state, done) # 将state等这些transition存入memory
+            state = next_state # 跳转到下一个状态
+            agent.update() # 每步更新网络
             if done:
                 break
-
-        # Update the target network, copying all weights and biases in DQN
+        # 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), 'Explore: %.2f' % agent.epsilon)
+              int(ep_reward), 'n_steps:', i_step, 'done: ', done,' Explore: %.2f' % agent.epsilon)
+        ep_steps.append(i_step)
         rewards.append(ep_reward)
         # 计算滑动窗口的reward
         if i_episode == 1:
@@ -83,14 +80,17 @@ if __name__ == "__main__":
         else:
             moving_average_rewards.append(
                 0.9*moving_average_rewards[-1]+0.1*ep_reward)
-
+    # 存储reward等相关结果
     import os
     import numpy as np
     output_path = os.path.dirname(__file__)+"/result/"
+    # 检测是否存在文件夹
     if not os.path.exists(output_path):
         os.mkdir(output_path)
     np.save(output_path+"rewards.npy", rewards)
     np.save(output_path+"moving_average_rewards.npy", moving_average_rewards)
+    np.save(output_path+"steps.npy", ep_steps)
     print('Complete！')
     plot(rewards)
     plot(moving_average_rewards, ylabel="moving_average_rewards")
+    plot(ep_steps, ylabel="steps_of_each_episode")

codes/dqn/model.py

@@ -5,7 +5,7 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-12 00:47:02
 @LastEditor: John
-@LastEditTime: 2020-06-14 11:23:04
+LastEditTime: 2020-08-19 16:55:54
 @Discription: 
 @Environment: python 3.7.7
 '''
@@ -14,17 +14,17 @@ import torch.nn.functional as F
 
 class FCN(nn.Module):
     def __init__(self, n_states=4, n_actions=18):
-        """
-        Initialize a deep Q-learning network for testing algorithm
-            n_states: number of features of input.
-            n_actions: number of action-value to output, one-to-one correspondence to action in game.
+        """ 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__()
-        self.fc1 = nn.Linear(n_states, 128)
-        self.fc2 = nn.Linear(128, 128)
-        self.fc3 = nn.Linear(128, n_actions)
+        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
         
     def forward(self, x):
-        x = F.relu(self.fc1(x))
+        # 各层对应的激活函数
+        x = F.relu(self.fc1(x)) 
         x = F.relu(self.fc2(x))
         return self.fc3(x)

codes/dqn/plot.py

@@ -5,19 +5,22 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-11 16:30:09
 @LastEditor: John
-@LastEditTime: 2020-06-14 11:38:42
+LastEditTime: 2020-08-20 16:34:34
 @Discription: 
 @Environment: python 3.7.7
 '''
 import matplotlib.pyplot as plt
+import pandas as pd
+import seaborn as sns
 import numpy as np
 import os 
 
 def plot(item,ylabel='rewards'):
+    sns.set()
     plt.figure()
     plt.plot(np.arange(len(item)), item)
     plt.title(ylabel+' of DQN') 
-    plt.ylabel('rewards')
+    plt.ylabel(ylabel)
     plt.xlabel('episodes')
     plt.savefig(os.path.dirname(__file__)+"/result/"+ylabel+".png")
     plt.show()