update PolicyGradient

codes/PolicyGradient/README.md

@@ -0,0 +1,42 @@
+# Policy Gradient
+实现的是Policy Gradient最基本的REINFORCE方法
+## 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 
+
+## 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)
+
+[Policy Gradient paper](https://papers.nips.cc/paper/1713-policy-gradient-methods-for-reinforcement-learning-with-function-approximation.pdf)
+
+[REINFORCE](https://towardsdatascience.com/policy-gradient-methods-104c783251e0)

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 12:05:03
+LastEditTime: 2020-11-23 17:04:37
 Discription: 
 Environment: 
 '''
@@ -18,9 +18,9 @@ from model import FCN
 
 class PolicyGradient:
     
-    def __init__(self, n_states,device='cpu',gamma = 0.99,lr = 0.01,batch_size=5):
+    def __init__(self, state_dim,device='cpu',gamma = 0.99,lr = 0.01,batch_size=5):
         self.gamma = gamma
-        self.policy_net = FCN(n_states)
+        self.policy_net = FCN(state_dim)
         self.optimizer = torch.optim.RMSprop(self.policy_net.parameters(), lr=lr)
         self.batch_size = batch_size
 
@@ -65,4 +65,8 @@ class PolicyGradient:
             loss = -m.log_prob(action) * reward  # Negtive score function x reward
             # print(loss)
             loss.backward()
-        self.optimizer.step()
+        self.optimizer.step()
+    def save_model(self,path):
+        torch.save(self.policy_net.state_dict(), path)
+    def load_model(self,path):
+        self.policy_net.load_state_dict(torch.load(path)) 

codes/PolicyGradient/env.py

@@ -14,6 +14,6 @@ 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
-    n_states = env.observation_space.shape[0]
+    state_dim = env.observation_space.shape[0]
     n_actions = env.action_space.n
-    return env,n_states,n_actions
+    return env,state_dim,n_actions

codes/PolicyGradient/main.py

@@ -5,28 +5,38 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2020-11-22 23:21:53
 LastEditor: John
-LastEditTime: 2020-11-23 12:06:15
+LastEditTime: 2020-11-24 19:52:40
 Discription: 
 Environment: 
 '''
 from itertools import count
 import torch
+import os
+from torch.utils.tensorboard import SummaryWriter
+
 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,n_states,n_actions = env_init()
+    env,state_dim,n_actions = env_init()
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测gpu
-    agent  = PolicyGradient(n_states,device = device,lr = cfg.policy_lr)
+    agent  = PolicyGradient(state_dim,device = device,lr = cfg.policy_lr)
     '''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
     for i_episode in range(cfg.train_eps):
         state = env.reset()
         ep_reward = 0
-        for t in count():
+        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
@@ -39,14 +49,61 @@ def train(cfg):
             if done:
                 print('Episode:', i_episode, ' Reward:',  ep_reward)
                 break
-        # if i_episode % cfg.batch_size == 0:
-        if i_episode > 0 and i_episode % 5 == 0:
+        if i_episode > 0 and i_episode % cfg.batch_size == 0:
             agent.update(reward_pool,state_pool,action_pool)
             state_pool = [] # 每个episode的state
             action_pool = []
             reward_pool = []
+        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 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！')
+    
 if __name__ == "__main__":
     cfg = get_args()
-    train(cfg)
+    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)

codes/PolicyGradient/model.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2020-11-22 23:18:46
 LastEditor: John
-LastEditTime: 2020-11-23 01:58:22
+LastEditTime: 2020-11-27 16:55:25
 Discription: 
 Environment: 
 '''
@@ -13,11 +13,11 @@ import torch.nn as nn
 import torch.nn.functional as F
 class FCN(nn.Module):
     ''' 全连接网络'''
-    def __init__(self,n_states):
+    def __init__(self,state_dim):
         super(FCN, self).__init__()
-        # 24和36为hidden layer的层数，可根据n_states, n_actions的情况来改变
-        self.fc1 = nn.Linear(n_states, 24)
-        self.fc2 = nn.Linear(24, 36)
+        # 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):

codes/PolicyGradient/params.py

@@ -5,15 +5,25 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2020-11-22 23:25:37
 LastEditor: John
-LastEditTime: 2020-11-22 23:32:44
+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_eps", default=1200, type=int) # 训练的最大episode数目
+    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

@@ -0,0 +1,46 @@
+#!/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) 

codes/PolicyGradient/utils.py

@@ -0,0 +1,29 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2020-11-23 13:44:52
+LastEditor: John
+LastEditTime: 2020-11-23 13:45:42
+Discription: 
+Environment: 
+'''
+import os
+import numpy as np
+
+
+def save_results(rewards,moving_average_rewards,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)
+    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！')