add A2C

codes/A2C/.vscode/settings.json

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

codes/A2C/agent.py

@@ -0,0 +1,33 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2020-11-03 20:47:09
+LastEditor: John
+LastEditTime: 2020-11-08 22:16:29
+Discription: 
+Environment: 
+'''
+from model import ActorCritic
+import torch.optim as optim
+
+class A2C:
+    def __init__(self,n_states, n_actions, hidden_dim=256,device="cpu",lr = 3e-4):
+        self.device = device
+        self.gamma = 0.99
+        self.model = ActorCritic(n_states, n_actions, hidden_dim=hidden_dim).to(device)
+        self.optimizer = optim.Adam(self.model.parameters(),lr=lr)
+    def choose_action(self, state):
+        dist, value = self.model(state)
+        action = dist.sample()
+        return action
+    def compute_returns(self,next_value, rewards, masks):
+        R = next_value
+        returns = []
+        for step in reversed(range(len(rewards))):
+            R = rewards[step] + self.gamma * R * masks[step]
+            returns.insert(0, R)
+        return returns
+    def update(self):
+        pass

codes/A2C/common/multiprocessing_env.py

@@ -0,0 +1,153 @@
+#This code is from openai baseline
+#https://github.com/openai/baselines/tree/master/baselines/common/vec_env
+
+import numpy as np
+from multiprocessing import Process, Pipe
+
+def worker(remote, parent_remote, env_fn_wrapper):
+    parent_remote.close()
+    env = env_fn_wrapper.x()
+    while True:
+        cmd, data = remote.recv()
+        if cmd == 'step':
+            ob, reward, done, info = env.step(data)
+            if done:
+                ob = env.reset()
+            remote.send((ob, reward, done, info))
+        elif cmd == 'reset':
+            ob = env.reset()
+            remote.send(ob)
+        elif cmd == 'reset_task':
+            ob = env.reset_task()
+            remote.send(ob)
+        elif cmd == 'close':
+            remote.close()
+            break
+        elif cmd == 'get_spaces':
+            remote.send((env.observation_space, env.action_space))
+        else:
+            raise NotImplementedError
+
+class VecEnv(object):
+    """
+    An abstract asynchronous, vectorized environment.
+    """
+    def __init__(self, num_envs, observation_space, action_space):
+        self.num_envs = num_envs
+        self.observation_space = observation_space
+        self.action_space = action_space
+
+    def reset(self):
+        """
+        Reset all the environments and return an array of
+        observations, or a tuple of observation arrays.
+        If step_async is still doing work, that work will
+        be cancelled and step_wait() should not be called
+        until step_async() is invoked again.
+        """
+        pass
+
+    def step_async(self, actions):
+        """
+        Tell all the environments to start taking a step
+        with the given actions.
+        Call step_wait() to get the results of the step.
+        You should not call this if a step_async run is
+        already pending.
+        """
+        pass
+
+    def step_wait(self):
+        """
+        Wait for the step taken with step_async().
+        Returns (obs, rews, dones, infos):
+         - obs: an array of observations, or a tuple of
+                arrays of observations.
+         - rews: an array of rewards
+         - dones: an array of "episode done" booleans
+         - infos: a sequence of info objects
+        """
+        pass
+
+    def close(self):
+        """
+        Clean up the environments' resources.
+        """
+        pass
+
+    def step(self, actions):
+        self.step_async(actions)
+        return self.step_wait()
+
+    
+class CloudpickleWrapper(object):
+    """
+    Uses cloudpickle to serialize contents (otherwise multiprocessing tries to use pickle)
+    """
+    def __init__(self, x):
+        self.x = x
+    def __getstate__(self):
+        import cloudpickle
+        return cloudpickle.dumps(self.x)
+    def __setstate__(self, ob):
+        import pickle
+        self.x = pickle.loads(ob)
+
+        
+class SubprocVecEnv(VecEnv):
+    def __init__(self, env_fns, spaces=None):
+        """
+        envs: list of gym environments to run in subprocesses
+        """
+        self.waiting = False
+        self.closed = False
+        nenvs = len(env_fns)
+        self.nenvs = nenvs
+        self.remotes, self.work_remotes = zip(*[Pipe() for _ in range(nenvs)])
+        self.ps = [Process(target=worker, args=(work_remote, remote, CloudpickleWrapper(env_fn)))
+            for (work_remote, remote, env_fn) in zip(self.work_remotes, self.remotes, env_fns)]
+        for p in self.ps:
+            p.daemon = True # if the main process crashes, we should not cause things to hang
+            p.start()
+        for remote in self.work_remotes:
+            remote.close()
+
+        self.remotes[0].send(('get_spaces', None))
+        observation_space, action_space = self.remotes[0].recv()
+        VecEnv.__init__(self, len(env_fns), observation_space, action_space)
+
+    def step_async(self, actions):
+        for remote, action in zip(self.remotes, actions):
+            remote.send(('step', action))
+        self.waiting = True
+
+    def step_wait(self):
+        results = [remote.recv() for remote in self.remotes]
+        self.waiting = False
+        obs, rews, dones, infos = zip(*results)
+        return np.stack(obs), np.stack(rews), np.stack(dones), infos
+
+    def reset(self):
+        for remote in self.remotes:
+            remote.send(('reset', None))
+        return np.stack([remote.recv() for remote in self.remotes])
+
+    def reset_task(self):
+        for remote in self.remotes:
+            remote.send(('reset_task', None))
+        return np.stack([remote.recv() for remote in self.remotes])
+
+    def close(self):
+        if self.closed:
+            return
+        if self.waiting:
+            for remote in self.remotes:            
+                remote.recv()
+        for remote in self.remotes:
+            remote.send(('close', None))
+        for p in self.ps:
+            p.join()
+            self.closed = True
+            
+    def __len__(self):
+        return self.nenvs

codes/A2C/env.py

@@ -0,0 +1,48 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2020-10-30 15:39:37
+LastEditor: John
+LastEditTime: 2020-11-03 20:52:07
+Discription: 
+Environment: 
+'''
+
+import gym
+from common.multiprocessing_env import SubprocVecEnv
+
+# num_envs = 16
+# env_name = "Pendulum-v0"
+
+def make_envs(num_envs=16,env_name="Pendulum-v0"):
+    ''' 创建多个子环境
+    '''
+    num_envs = 16
+    env_name = "CartPole-v0"
+    def make_env():
+        def _thunk():
+            env = gym.make(env_name)
+            return env
+
+        return _thunk
+
+    envs = [make_env() for i in range(num_envs)]
+    envs = SubprocVecEnv(envs)
+    return envs
+# if __name__ == "__main__":
+
+#     num_envs = 16
+#     env_name = "CartPole-v0"
+#     def make_env():
+#         def _thunk():
+#             env = gym.make(env_name)
+#             return env
+
+#         return _thunk
+
+#     envs = [make_env() for i in range(num_envs)]
+#     envs = SubprocVecEnv(envs)
+if __name__ == "__main__":
+    envs = make_envs(num_envs=16,env_name="CartPole-v0") 

codes/A2C/main.py

@@ -0,0 +1,186 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+@Author: John
+@Email: johnjim0816@gmail.com
+@Date: 2020-06-11 20:58:21
+@LastEditor: John
+LastEditTime: 2020-11-08 22:19:56
+@Discription: 
+@Environment: python 3.7.9
+'''
+import torch
+import gym
+import os
+import numpy as np
+import argparse
+from torch.utils.tensorboard import SummaryWriter
+
+from agent import A2C
+from env import make_envs
+from utils import SEQUENCE, SAVED_MODEL_PATH, RESULT_PATH
+from utils import save_model,save_results
+
+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)  # reward 折扣因子
+    parser.add_argument("--lr", default=3e-4, 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=4000, type=int)
+    parser.add_argument("--train_steps", default=5, 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 test_env(agent,device='cpu'):
+    env = gym.make("CartPole-v0")
+    state = env.reset()
+    ep_reward=0
+    for _ in range(200):
+        state = torch.FloatTensor(state).unsqueeze(0).to(device)
+        dist, value = agent.model(state)
+        action = dist.sample()
+        next_state, reward, done, _ = env.step(action.cpu().numpy()[0])
+        state = next_state
+        ep_reward += reward
+        if done:
+            break
+    return ep_reward
+
+
+def train(cfg):
+    print('Start to train ! \n')
+    envs = make_envs(num_envs=16,env_name="CartPole-v0")
+    n_states = envs.observation_space.shape[0]
+    n_actions = envs.action_space.n
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    agent = A2C(n_states, n_actions, hidden_dim=256)
+    # moving_average_rewards = []
+    # ep_steps = []
+    log_dir=os.path.split(os.path.abspath(__file__))[0]+"/logs/train/" + SEQUENCE
+    writer = SummaryWriter(log_dir)
+    state = envs.reset()
+    for i_episode in range(1, cfg.train_eps+1):
+        log_probs = []
+        values    = []
+        rewards = []
+        masks     = []
+        entropy = 0
+        for i_step in range(1, cfg.train_steps+1):
+            state = torch.FloatTensor(state).to(device)
+            dist, value = agent.model(state)
+            action = dist.sample()
+            next_state, reward, done, _ = envs.step(action.cpu().numpy())
+            state = next_state                                                  
+            log_prob = dist.log_prob(action)
+            entropy += dist.entropy().mean()
+            log_probs.append(log_prob)
+            values.append(value)
+            rewards.append(torch.FloatTensor(reward).unsqueeze(1).to(device))
+            masks.append(torch.FloatTensor(1 - done).unsqueeze(1).to(device))
+        if i_episode%20 == 0:
+            print("reward",test_env(agent,device='cpu'))
+        next_state = torch.FloatTensor(next_state).to(device)
+        _, next_value =agent.model(next_state)
+        returns = agent.compute_returns(next_value, rewards, masks)
+
+        log_probs = torch.cat(log_probs)
+        returns   = torch.cat(returns).detach()
+        values    = torch.cat(values)
+        advantage = returns - values
+        actor_loss  = -(log_probs * advantage.detach()).mean()
+        critic_loss = advantage.pow(2).mean()
+        loss = actor_loss + 0.5 * critic_loss - 0.001 * entropy
+        
+        agent.optimizer.zero_grad()
+        loss.backward()
+        agent.optimizer.step()
+    for _ in range(100):
+        print("test_reward",test_env(agent,device='cpu'))
+        
+        # print('Episode:', i_episode, ' Reward: %i' %
+        #       int(ep_reward[0]), 'n_steps:', i_step)
+        # ep_steps.append(i_step)
+        # rewards.append(ep_reward)
+        # if i_episode == 1:
+        #     moving_average_rewards.append(ep_reward[0])
+        # else:
+        #     moving_average_rewards.append(
+        #         0.9*moving_average_rewards[-1]+0.1*ep_reward[0])
+        # 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()
+    print('Complete training！')
+    ''' 保存模型 '''
+    # save_model(agent,model_path=SAVED_MODEL_PATH)
+    # '''存储reward等相关结果'''
+    # save_results(rewards,moving_average_rewards,ep_steps,tag='train',result_path=RESULT_PATH)
+
+# 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.choose_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)
+
+if __name__ == "__main__":
+    cfg = get_args()
+    train(cfg)
+    
+    # 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)

codes/A2C/model.py

@@ -0,0 +1,35 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2020-11-03 20:45:25
+LastEditor: John
+LastEditTime: 2020-11-07 18:49:09
+Discription: 
+Environment: 
+'''
+import torch.nn as nn
+from torch.distributions import Categorical
+
+class ActorCritic(nn.Module):
+    def __init__(self, n_states, n_actions, hidden_dim=256, std=0.0):
+        super(ActorCritic, self).__init__()
+        self.critic = nn.Sequential(
+            nn.Linear(n_states, hidden_dim),
+            nn.ReLU(),
+            nn.Linear(hidden_dim, 1)
+        )
+        
+        self.actor = nn.Sequential(
+            nn.Linear(n_states, hidden_dim),
+            nn.ReLU(),
+            nn.Linear(hidden_dim, n_actions),
+            nn.Softmax(dim=1),
+        )
+        
+    def forward(self, x):
+        value = self.critic(x)
+        probs = self.actor(x)
+        dist  = Categorical(probs)
+        return dist, value

codes/A2C/utils.py

@@ -0,0 +1,32 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2020-10-15 21:31:19
+LastEditor: John
+LastEditTime: 2020-11-03 17:05:48
+Discription: 
+Environment: 
+'''
+import os
+import numpy as np
+import datetime
+
+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 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 )
+    
+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！')