update projects

projects/codes/PolicyGradient/README.md

@@ -0,0 +1,27 @@
+# Policy Gradient
+
+
+Policy-based方法是强化学习中与Value-based(比如Q-learning)相对的方法，其目的是对策略本身进行梯度下降，相关基础知识参考[Datawhale-Policy Gradient](https://datawhalechina.github.io/leedeeprl-notes/#/chapter4/chapter4)。
+其中REINFORCE是一个最基本的Policy Gradient方法，主要解决策略梯度无法直接计算的问题，具体原理参考[CSDN-REINFORCE和Reparameterization Trick](https://blog.csdn.net/JohnJim0/article/details/110230703)
+
+## 伪代码
+
+结合REINFORCE原理，其伪代码如下：
+
+<img src="assets/image-20211016004808604.png" alt="image-20211016004808604" style="zoom:50%;" />
+
+https://pytorch.org/docs/stable/distributions.html
+
+加负号的原因是，在公式中应该是实现的梯度上升算法，而loss一般使用随机梯度下降的，所以加个负号保持一致性。
+
+![img](assets/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L0pvaG5KaW0w,size_16,color_FFFFFF,t_70-20210428001336032.png)
+
+## 实现
+
+## 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)

projects/codes/PolicyGradient/pg.py

@@ -0,0 +1,89 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2020-11-22 23:27:44
+LastEditor: John
+LastEditTime: 2022-02-10 01:25:27
+Discription: 
+Environment: 
+'''
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.distributions import Bernoulli
+from torch.autograd import Variable
+import numpy as np
+
+class MLP(nn.Module):
+    
+    ''' 多层感知机
+        输入：state维度
+        输出：概率
+    '''
+    def __init__(self,input_dim,hidden_dim = 36):
+        super(MLP, self).__init__()
+        # 24和36为hidden layer的层数，可根据input_dim, n_actions的情况来改变
+        self.fc1 = nn.Linear(input_dim, hidden_dim)
+        self.fc2 = nn.Linear(hidden_dim,hidden_dim)
+        self.fc3 = nn.Linear(hidden_dim, 1)  # Prob of Left
+
+    def forward(self, x):
+        x = F.relu(self.fc1(x))
+        x = F.relu(self.fc2(x))
+        x = F.sigmoid(self.fc3(x))
+        return x
+        
+class PolicyGradient:
+    
+    def __init__(self, n_states,cfg):
+        self.gamma = cfg.gamma
+        self.policy_net = MLP(n_states,hidden_dim=cfg.hidden_dim)
+        self.optimizer = torch.optim.RMSprop(self.policy_net.parameters(), lr=cfg.lr)
+        self.batch_size = cfg.batch_size
+
+    def choose_action(self,state):
+        
+        state = torch.from_numpy(state).float()
+        state = Variable(state)
+        probs = self.policy_net(state)
+        m = Bernoulli(probs) # 伯努利分布
+        action = m.sample()
+        action = action.data.numpy().astype(int)[0] # 转为标量
+        return action
+        
+    def update(self,reward_pool,state_pool,action_pool):
+        # Discount reward
+        running_add = 0
+        for i in reversed(range(len(reward_pool))):
+            if reward_pool[i] == 0:
+                running_add = 0
+            else:
+                running_add = running_add * self.gamma + reward_pool[i]
+                reward_pool[i] = running_add
+
+        # Normalize reward
+        reward_mean = np.mean(reward_pool)
+        reward_std = np.std(reward_pool)
+        for i in range(len(reward_pool)):
+            reward_pool[i] = (reward_pool[i] - reward_mean) / reward_std
+
+        # Gradient Desent
+        self.optimizer.zero_grad()
+
+        for i in range(len(reward_pool)):
+            state = state_pool[i]
+            action = Variable(torch.FloatTensor([action_pool[i]]))
+            reward = reward_pool[i]
+            state = Variable(torch.from_numpy(state).float())
+            probs = self.policy_net(state)
+            m = Bernoulli(probs)
+            loss = -m.log_prob(action) * reward  # Negtive score function x reward
+            # print(loss)
+            loss.backward()
+        self.optimizer.step()
+    def save(self,path):
+        torch.save(self.policy_net.state_dict(), path+'pg_checkpoint.pt')
+    def load(self,path):
+        self.policy_net.load_state_dict(torch.load(path+'pg_checkpoint.pt')) 

projects/codes/PolicyGradient/task0.py

@@ -0,0 +1,143 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2020-11-22 23:21:53
+LastEditor: John
+LastEditTime: 2022-07-21 21:44:00
+Discription: 
+Environment: 
+'''
+import sys,os
+curr_path = os.path.dirname(os.path.abspath(__file__))  # current path
+parent_path = os.path.dirname(curr_path)  # parent path
+sys.path.append(parent_path)  # add to system path
+
+import gym
+import torch
+import datetime
+import argparse
+from itertools import count
+
+from pg import PolicyGradient
+from common.utils import save_results, make_dir
+from common.utils import plot_rewards
+
+
+def get_args():
+    """ Hyperparameters
+    """
+    curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")  # Obtain current time
+    parser = argparse.ArgumentParser(description="hyperparameters")      
+    parser.add_argument('--algo_name',default='PolicyGradient',type=str,help="name of algorithm")
+    parser.add_argument('--env_name',default='CartPole-v0',type=str,help="name of environment")
+    parser.add_argument('--train_eps',default=300,type=int,help="episodes of training")
+    parser.add_argument('--test_eps',default=20,type=int,help="episodes of testing")
+    parser.add_argument('--gamma',default=0.99,type=float,help="discounted factor")
+    parser.add_argument('--lr',default=0.01,type=float,help="learning rate")
+    parser.add_argument('--batch_size',default=8,type=int)
+    parser.add_argument('--hidden_dim',default=36,type=int)
+    parser.add_argument('--device',default='cpu',type=str,help="cpu or cuda") 
+    parser.add_argument('--result_path',default=curr_path + "/outputs/" + parser.parse_args().env_name + \
+            '/' + curr_time + '/results/' )
+    parser.add_argument('--model_path',default=curr_path + "/outputs/" + parser.parse_args().env_name + \
+            '/' + curr_time + '/models/' ) # path to save models
+    parser.add_argument('--save_fig',default=True,type=bool,help="if save figure or not")           
+    args = parser.parse_args()                          
+    return args
+
+
+def env_agent_config(cfg,seed=1):
+    env = gym.make(cfg.env_name)  
+    env.seed(seed)
+    n_states = env.observation_space.shape[0]
+    agent = PolicyGradient(n_states,cfg)
+    return env,agent
+
+def train(cfg,env,agent):
+    print('Start training!')
+    print(f'Env:{cfg.env_name}, Algorithm:{cfg.algo_name}, Device:{cfg.device}')
+    state_pool = [] # temp states pool per several episodes
+    action_pool = []
+    reward_pool = [] 
+    rewards = []
+    ma_rewards = []
+    for i_ep in range(cfg.train_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
+            if done:
+                reward = 0
+            state_pool.append(state)
+            action_pool.append(float(action))
+            reward_pool.append(reward)
+            state = next_state
+            if done:
+                print(f'Episode：{i_ep+1}/{cfg.train_eps}, Reward:{ep_reward:.2f}')
+                break
+        if i_ep > 0 and i_ep % cfg.batch_size == 0:
+            agent.update(reward_pool,state_pool,action_pool)
+            state_pool = [] 
+            action_pool = []
+            reward_pool = []
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(
+                0.9*ma_rewards[-1]+0.1*ep_reward)
+        else:
+            ma_rewards.append(ep_reward)
+    print('Finish training!')
+    env.close() # close environment
+    return rewards, ma_rewards
+            
+
+def test(cfg,env,agent):
+    print('开始测试!')
+    print(f'环境：{cfg.env_name}, 算法：{cfg.algo_name}, 设备：{cfg.device}')
+    rewards = []
+    ma_rewards = []
+    for i_ep in range(cfg.test_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
+            if done:
+                reward = 0
+            state = next_state
+            if done:
+                print('回合：{}/{}, 奖励：{}'.format(i_ep + 1, cfg.train_eps, ep_reward))
+                break
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(
+                0.9*ma_rewards[-1]+0.1*ep_reward)
+        else:
+            ma_rewards.append(ep_reward)
+    print('完成测试！')
+    env.close()
+    return rewards, ma_rewards
+    
+if __name__ == "__main__":
+    cfg = Config()
+    # 训练
+    env, agent = env_agent_config(cfg)
+    rewards, ma_rewards = train(cfg, env, agent)
+    make_dir(cfg.result_path, cfg.model_path)  # 创建保存结果和模型路径的文件夹
+    agent.save(path=cfg.model_path)  # 保存模型
+    save_results(rewards, ma_rewards, tag='train',
+                 path=cfg.result_path)  # 保存结果
+    plot_rewards(rewards, ma_rewards, cfg, tag="train")  # 画出结果
+    # 测试
+    env, agent = env_agent_config(cfg)
+    agent.load(path=cfg.model_path)  # 导入模型
+    rewards, ma_rewards = test(cfg, env, agent)
+    save_results(rewards, ma_rewards, tag='test',
+                 path=cfg.result_path)  # 保存结果
+    plot_rewards(rewards, ma_rewards, cfg, tag="test")  # 画出结果
+