update DQN

codes/dqn/.vscode/settings.json

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

codes/dqn/README.md

@@ -1,3 +1,8 @@
+## 思路
+
+见[我的博客](https://blog.csdn.net/JohnJim0/article/details/109557173)
+## 环境
+
 python 3.7.9
 
 pytorch 1.6.0
@@ -6,6 +11,7 @@ tensorboard 2.3.0
 
 torchvision 0.7.0 
 
+## 使用
 
 train: 
 
@@ -18,7 +24,12 @@ eval:
 ```python
 python main.py --train 0 
 ```
-
+可视化：
 ```python
 tensorboard --logdir logs 
-```
+```
+
+## Torch知识
+
+[with torch.no_grad()](https://www.jianshu.com/p/1cea017f5d11)
+

codes/dqn/agent.py

@@ -5,7 +5,7 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-12 00:50:49
 @LastEditor: John
-LastEditTime: 2020-10-15 21:56:21
+LastEditTime: 2020-11-22 11:12:30
 @Discription: 
 @Environment: python 3.7.7
 '''
@@ -24,11 +24,12 @@ 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
+        
         self.n_actions = n_actions  # 总的动作个数
         self.device = device  # 设备，cpu或gpu等
-        self.gamma = gamma
+        self.gamma = gamma # 奖励的折扣因子
         # e-greedy策略相关参数
+        self.actions_count = 0 # 用于epsilon的衰减计数
         self.epsilon = 0
         self.epsilon_start = epsilon_start
         self.epsilon_end = epsilon_end
@@ -67,12 +68,11 @@ class DQN:
                 action = random.randrange(self.n_actions)
             return action
         else: 
-            with torch.no_grad():
+            with torch.no_grad(): # 取消保存梯度
                     # 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)
+                        [state], device='cpu', dtype=torch.float32) # 如tensor([[-0.0798, -0.0079]], grad_fn=<AddmmBackward>)
-                    # 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]))
@@ -86,8 +86,8 @@ 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
+        '''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]])
+        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(
@@ -99,9 +99,8 @@ class DQN:
         done_batch = torch.tensor(np.float32(
             done_batch), device=self.device).unsqueeze(1)  # 将bool转为float然后转为张量
 
-        # 计算当前(s_t,a)对应的Q(s_t, a)
+        '''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]])
+        '''torch.gather:对于a=torch.Tensor([[1,2],[3,4]]),那么a.gather(1,torch.Tensor([[0],[1]]))=torch.Tensor([[1],[3]])'''
-        # 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
@@ -119,6 +118,7 @@ class DQN:
         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()  # 更新模型
 
     def save_model(self,path):

codes/dqn/main.py

@@ -5,7 +5,7 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-12 00:48:57
 @LastEditor: John
-LastEditTime: 2020-10-15 22:00:28
+LastEditTime: 2020-11-23 11:58:17
 @Discription: 
 @Environment: python 3.7.7
 '''
@@ -16,7 +16,7 @@ import argparse
 from torch.utils.tensorboard import SummaryWriter
 import datetime
 import os
-from utils import save_results
+from utils import save_results,save_model
 
 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+'/'
@@ -53,7 +53,7 @@ def get_args():
 def train(cfg):
     print('Start to train ! \n')
     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 = gym.make('CartPole-v0')
     env.seed(1) # 设置env随机种子
     n_states = env.observation_space.shape[0]
     n_actions = env.action_space.n
@@ -95,10 +95,7 @@ def train(cfg):
     writer.close()
     print('Complete training！')
     ''' 保存模型 '''
-    if not os.path.exists(SAVED_MODEL_PATH): # 检测是否存在文件夹
+    save_model(agent,model_path=SAVED_MODEL_PATH)
-        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)
     
@@ -110,7 +107,7 @@ def eval(cfg, saved_model_path = SAVED_MODEL_PATH):
     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,
+    agent = DQN(n_states=n_states, n_actions=n_actions, device="cpu", 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 = []

codes/dqn/plot.py

@@ -5,7 +5,7 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-11 16:30:09
 @LastEditor: John
-LastEditTime: 2020-10-15 22:01:50
+LastEditTime: 2020-11-23 13:48:31
 @Discription: 
 @Environment: python 3.7.7
 '''
@@ -27,18 +27,6 @@ def plot(item,ylabel='rewards_train', save_fig = True):
         plt.savefig(os.path.dirname(__file__)+"/result/"+ylabel+".png")
     plt.show()
 
-# def plot(item,ylabel='rewards'):
-#     
-#     df = pd.DataFrame(dict(time=np.arange(len(item)),value=item))
-#     g = sns.relplot(x="time", y="value", kind="line", data=df)
-#     # g.fig.autofmt_xdate()
-#     # sns.lineplot(time=time, data=item, color="r", condition="behavior_cloning")
-#     # # sns.tsplot(time=time, data=x2, color="b", condition="dagger")
-#     # plt.ylabel("Reward")
-#     # plt.xlabel("Iteration Number")
-#     # plt.title("Imitation Learning")
-
-    # plt.show()
 if __name__ == "__main__":
 
     output_path = os.path.split(os.path.abspath(__file__))[0]+"/result/"

codes/dqn/utils.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2020-10-15 21:28:00
 LastEditor: John
-LastEditTime: 2020-10-15 21:50:30
+LastEditTime: 2020-10-30 16:56:55
 Discription: 
 Environment: 
 '''
@@ -14,8 +14,17 @@ import numpy as np
 
 
 def save_results(rewards,moving_average_rewards,ep_steps,tag='train',result_path='./result'):
+    '''保存reward等结果
+    '''
     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_'+tag+'.npy', rewards)
     np.save(result_path+'moving_average_rewards_'+tag+'.npy', moving_average_rewards)
-    np.save(result_path+'steps_'+tag+'.npy',ep_steps )
+    np.save(result_path+'steps_'+tag+'.npy',ep_steps )
+    print('results saved!')
+
+def save_model(agent,model_path='./saved_model'):
+    if not os.path.exists(model_path): # 检测是否存在文件夹
+        os.mkdir(model_path)
+    agent.save_model(model_path+'checkpoint.pth')
+    print('model saved！')