update

codes/ddpg/.vscode/settings.json

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

codes/ddpg/README.md

@@ -1,26 +1,5 @@
+# DDPG
 
-python 3.7.9
+## 伪代码
 
-pytorch 1.6.0
-
-tensorboard 2.3.0 
-
-torchvision 0.7.0 
-
-train: 
-
-```python
-python main.py 
-```
-
-eval: 
-
-```python
-python main.py --train 0 
-```
-
-open tensorboard:
-
-```python
-tensorboard --logdir logs 
-```
+![image-20210320151900695](assets/image-20210320151900695.png)

codes/ddpg/agent.py

@@ -5,7 +5,7 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-09 20:25:52
 @LastEditor: John
-LastEditTime: 2020-09-02 01:19:13
+LastEditTime: 2021-03-17 20:43:25
 @Discription: 
 @Environment: python 3.7.7
 '''
@@ -14,18 +14,17 @@ import torch
 import torch.nn as nn
 import torch.optim as optim
 
-from model import Actor, Critic
-from memory import ReplayBuffer
+from common.model import Actor, Critic
+from common.memory import ReplayBuffer
 
 
 class DDPG:
-    def __init__(self, n_states, n_actions, hidden_dim=30, device="cpu", critic_lr=1e-3,
-                 actor_lr=1e-4, gamma=0.99, soft_tau=1e-2, memory_capacity=100000, batch_size=128):
-        self.device = device
-        self.critic = Critic(n_states, n_actions, hidden_dim).to(device)
-        self.actor = Actor(n_states, n_actions, hidden_dim).to(device)
-        self.target_critic = Critic(n_states, n_actions, hidden_dim).to(device)
-        self.target_actor = Actor(n_states, n_actions, hidden_dim).to(device)
+    def __init__(self, n_states, n_actions, cfg):
+        self.device = cfg.device
+        self.critic = Critic(n_states, n_actions, cfg.hidden_dim).to(cfg.device)
+        self.actor = Actor(n_states, n_actions, cfg.hidden_dim).to(cfg.device)
+        self.target_critic = Critic(n_states, n_actions, cfg.hidden_dim).to(cfg.device)
+        self.target_actor = Actor(n_states, n_actions, cfg.hidden_dim).to(cfg.device)
 
         for target_param, param in zip(self.target_critic.parameters(), self.critic.parameters()):
             target_param.data.copy_(param.data)
@@ -33,14 +32,14 @@ class DDPG:
             target_param.data.copy_(param.data)
 
         self.critic_optimizer = optim.Adam(
-            self.critic.parameters(),  lr=critic_lr)
-        self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=actor_lr)
-        self.memory = ReplayBuffer(memory_capacity)
-        self.batch_size = batch_size
-        self.soft_tau = soft_tau
-        self.gamma = gamma
+            self.critic.parameters(),  lr=cfg.critic_lr)
+        self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=cfg.actor_lr)
+        self.memory = ReplayBuffer(cfg.memory_capacity)
+        self.batch_size = cfg.batch_size
+        self.soft_tau = cfg.soft_tau
+        self.gamma = cfg.gamma
 
-    def select_action(self, state):
+    def choose_action(self, state):
         state = torch.FloatTensor(state).unsqueeze(0).to(self.device)
         action = self.actor(state)
         # torch.detach()用于切断反向传播
@@ -87,8 +86,8 @@ class DDPG:
                 target_param.data * (1.0 - self.soft_tau) +
                 param.data * self.soft_tau
             )
-    def save_model(self,path):
-        torch.save(self.target_actor.state_dict(), path)
+    def save(self,path):
+        torch.save(self.target_net.state_dict(), path+'DDPG_checkpoint.pth')
 
-    def load_model(self,path):
-        self.actor.load_state_dict(torch.load(path))
+    def load(self,path):
+        self.actor.load_state_dict(torch.load(path+'DDPG_checkpoint.pth')) 

codes/ddpg/env.py

@@ -5,7 +5,7 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-10 15:28:30
 @LastEditor: John
-LastEditTime: 2020-09-01 10:57:36
+LastEditTime: 2021-03-19 19:56:46
 @Discription: 
 @Environment: python 3.7.7
 '''
@@ -29,4 +29,33 @@ class NormalizedActions(gym.ActionWrapper):
         upper_bound = self.action_space.high
         action = 2 * (action - low_bound) / (upper_bound - low_bound) - 1
         action = np.clip(action, low_bound, upper_bound)
-        return action
+        return action
+
+class OUNoise(object):
+    '''Ornstein–Uhlenbeck
+    '''
+    def __init__(self, action_space, mu=0.0, theta=0.15, max_sigma=0.3, min_sigma=0.3, decay_period=100000):
+        self.mu           = mu
+        self.theta        = theta
+        self.sigma        = max_sigma
+        self.max_sigma    = max_sigma
+        self.min_sigma    = min_sigma
+        self.decay_period = decay_period
+        self.n_actions   = action_space.shape[0]
+        self.low          = action_space.low
+        self.high         = action_space.high
+        self.reset()
+        
+    def reset(self):
+        self.obs = np.ones(self.n_actions) * self.mu
+        
+    def evolve_obs(self):
+        x  = self.obs
+        dx = self.theta * (self.mu - x) + self.sigma * np.random.randn(self.n_actions)
+        self.obs = x + dx
+        return self.obs
+    
+    def get_action(self, action, t=0):
+        ou_obs = self.evolve_obs()
+        self.sigma = self.max_sigma - (self.max_sigma - self.min_sigma) * min(1.0, t / self.decay_period)
+        return np.clip(action + ou_obs, self.low, self.high)

codes/ddpg/main.py

@@ -5,74 +5,60 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-11 20:58:21
 @LastEditor: John
-LastEditTime: 2020-10-15 21:23:39
+LastEditTime: 2021-03-19 19:57:00
 @Discription: 
 @Environment: python 3.7.7
 '''
-from token import NUMBER
-from typing import Sequence
+import sys,os
+sys.path.append(os.getcwd()) # 添加当前终端路径
 import torch
 import gym
-from agent import DDPG
-from env import NormalizedActions
-from noise import OUNoise
-import os
 import numpy as np
-import argparse
-from torch.utils.tensorboard import SummaryWriter
 import datetime
+from DDPG.agent import DDPG
+from DDPG.env import NormalizedActions,OUNoise
+from common.plot import plot_rewards
+from common.utils import save_results
 
-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+'/'
+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)
 
-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)  # q-learning中的gamma
-    parser.add_argument("--critic_lr", default=1e-3, 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=200, type=int)
-    parser.add_argument("--train_steps", default=200, 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 train(cfg):
-    print('Start to train ! \n')
-    env = NormalizedActions(gym.make("Pendulum-v0"))
-
-    # 增加action噪声
-    ou_noise = OUNoise(env.action_space)
-
-    n_states = env.observation_space.shape[0]
-    n_actions = env.action_space.shape[0]
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-    agent = DDPG(n_states, n_actions, device="cpu", critic_lr=1e-3,
-                 actor_lr=1e-4, gamma=0.99, soft_tau=1e-2, memory_capacity=100000, batch_size=128)
+class DDPGConfig:
+    def __init__(self):
+        self.gamma = 0.99
+        self.critic_lr = 1e-3  
+        self.actor_lr = 1e-4 
+        self.memory_capacity = 10000
+        self.batch_size = 128
+        self.train_eps =300
+        self.train_steps = 200
+        self.eval_eps = 200
+        self.eval_steps = 200
+        self.target_update = 4
+        self.hidden_dim = 30
+        self.soft_tau=1e-2
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+def train(cfg,env,agent):
+    print('Start to train ! ')
+    ou_noise = OUNoise(env.action_space) # action noise
     rewards = []
-    moving_average_rewards = []
+    ma_rewards = [] # moving average 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()
         ou_noise.reset()
         ep_reward = 0
-        for i_step in range(1, cfg.train_steps+1):
-            action = agent.select_action(state)
+        for i_step in range(cfg.train_steps):
+            action = agent.choose_action(state)
             action = ou_noise.get_action(
                 action, i_step)  # 即paper中的random process
             next_state, reward, done, _ = env.step(action)
@@ -82,80 +68,25 @@ def train(cfg):
             state = next_state
             if done:
                 break
-        print('Episode:', i_episode, ' Reward: %i' %
-              int(ep_reward), 'n_steps:', i_step)
+        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)
-        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')
-    '''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_train.npy', rewards)
-    np.save(RESULT_PATH+'moving_average_rewards_train.npy', moving_average_rewards)
-    np.save(RESULT_PATH+'steps_train.npy', ep_steps)
-
-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.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
-            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)
+    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 = DDPGConfig()
+    env = NormalizedActions(gym.make("Pendulum-v0"))
+    env.seed(1) # 设置env随机种子
+    n_states = env.observation_space.shape[0]
+    n_actions = env.action_space.shape[0]
+    agent = DDPG(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/ddpg/memory.py

@@ -1,34 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-@Author: John
-@Email: johnjim0816@gmail.com
-@Date: 2020-06-10 15:27:16
-@LastEditor: John
-@LastEditTime: 2020-06-13 00:29:45
-@Discription: 
-@Environment: python 3.7.7
-'''
-import random
-import numpy as np
-
-class ReplayBuffer:
-    
-    def __init__(self, capacity):
-        self.capacity = capacity
-        self.buffer = []
-        self.position = 0
-    
-    def push(self, state, action, reward, next_state, done):
-        if len(self.buffer) < self.capacity:
-            self.buffer.append(None)
-        self.buffer[self.position] = (state, action, reward, next_state, done)
-        self.position = (self.position + 1) % self.capacity
-    
-    def sample(self, batch_size):
-        batch = random.sample(self.buffer, batch_size)
-        state_batch, action_batch, reward_batch, next_state_batch, done_batch = map(np.stack, zip(*batch))
-        return state_batch, action_batch, reward_batch, next_state_batch, done_batch
-    
-    def __len__(self):
-        return len(self.buffer)

codes/ddpg/model.py

@@ -1,50 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-@Author: John
-@Email: johnjim0816@gmail.com
-@Date: 2020-06-10 15:03:59
-@LastEditor: John
-LastEditTime: 2020-08-22 19:09:54
-@Discription: 
-@Environment: python 3.7.7
-'''
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-
-class Critic(nn.Module):
-    def __init__(self, n_obs, n_actions, hidden_size, init_w=3e-3):
-        super(Critic, self).__init__()
-        
-        self.linear1 = nn.Linear(n_obs + n_actions, hidden_size)
-        self.linear2 = nn.Linear(hidden_size, hidden_size)
-        self.linear3 = nn.Linear(hidden_size, 1)
-        # 随机初始化为较小的值
-        self.linear3.weight.data.uniform_(-init_w, init_w)
-        self.linear3.bias.data.uniform_(-init_w, init_w)
-        
-    def forward(self, state, action):
-        # 按维数1拼接
-        x = torch.cat([state, action], 1)
-        x = F.relu(self.linear1(x))
-        x = F.relu(self.linear2(x))
-        x = self.linear3(x)
-        return x
-
-class Actor(nn.Module):
-    def __init__(self, n_obs, n_actions, hidden_size, init_w=3e-3):
-        super(Actor, self).__init__()  
-        self.linear1 = nn.Linear(n_obs, hidden_size)
-        self.linear2 = nn.Linear(hidden_size, hidden_size)
-        self.linear3 = nn.Linear(hidden_size, n_actions)
-        
-        self.linear3.weight.data.uniform_(-init_w, init_w)
-        self.linear3.bias.data.uniform_(-init_w, init_w)
-        
-    def forward(self, x):
-        x = F.relu(self.linear1(x))
-        x = F.relu(self.linear2(x))
-        x = F.tanh(self.linear3(x))
-        return x
-    

codes/ddpg/noise.py

@@ -1,39 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-@Author: John
-@Email: johnjim0816@gmail.com
-@Date: 2020-06-11 20:58:59
-@LastEditor: John
-@LastEditTime: 2020-06-11 20:59:20
-@Discription: 
-@Environment: python 3.7.7
-'''
-import numpy as np
-
-class OUNoise(object):
-    def __init__(self, action_space, mu=0.0, theta=0.15, max_sigma=0.3, min_sigma=0.3, decay_period=100000):
-        self.mu           = mu
-        self.theta        = theta
-        self.sigma        = max_sigma
-        self.max_sigma    = max_sigma
-        self.min_sigma    = min_sigma
-        self.decay_period = decay_period
-        self.n_actions   = action_space.shape[0]
-        self.low          = action_space.low
-        self.high         = action_space.high
-        self.reset()
-        
-    def reset(self):
-        self.obs = np.ones(self.n_actions) * self.mu
-        
-    def evolve_obs(self):
-        x  = self.obs
-        dx = self.theta * (self.mu - x) + self.sigma * np.random.randn(self.n_actions)
-        self.obs = x + dx
-        return self.obs
-    
-    def get_action(self, action, t=0):
-        ou_obs = self.evolve_obs()
-        self.sigma = self.max_sigma - (self.max_sigma - self.min_sigma) * min(1.0, t / self.decay_period)
-        return np.clip(action + ou_obs, self.low, self.high)

codes/ddpg/plot.py

@@ -1,46 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-@Author: John
-@Email: johnjim0816@gmail.com
-@Date: 2020-06-11 16:30:09
-@LastEditor: John
-LastEditTime: 2020-10-15 21:32:05
-@Discription: 
-@Environment: python 3.7.7
-'''
-import matplotlib.pyplot as plt
-import seaborn as sns
-import numpy as np
-import os 
-
-def plot_results(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 DDPG') 
-    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_results(rewards)
-    plot_results(moving_average_rewards,ylabel='moving_average_rewards_'+tag)
-    plot_results(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_results(rewards,ylabel='rewards_'+tag)
-    plot_results(moving_average_rewards,ylabel='moving_average_rewards_'+tag)
-    plot_results(steps,ylabel='steps_'+tag)

codes/ddpg/utils.py

@@ -1,21 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-Author: John
-Email: johnjim0816@gmail.com
-Date: 2020-10-15 21:31:19
-LastEditor: John
-LastEditTime: 2020-10-15 21:31:25
-Discription: 
-Environment: 
-'''
-import os
-import numpy as np
-RESULT_PATH = os.path.split(os.path.abspath(__file__))[0]+"/result/"+SEQUENCE+'/'
-
-def save_results(rewards,moving_average_rewards,ep_steps,path=RESULT_PATH):
-    if not os.path.exists(path): # 检测是否存在文件夹
-            os.mkdir(path)
-    np.save(RESULT_PATH+'rewards_train.npy', rewards)
-    np.save(RESULT_PATH+'moving_average_rewards_train.npy', moving_average_rewards)
-    np.save(RESULT_PATH+'steps_train.npy',ep_steps )