bacow/easy-rl
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

update

Browse Source
This commit is contained in:
johnjim0816
2021-05-03 23:00:01 +08:00
parent 895094a893
commit 8028f7145e
67 changed files with 738 additions and 1137 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
codes/A2C/README.md
View File

@@ -0,0 +1,5 @@
## A2C
https://towardsdatascience.com/understanding-actor-critic-methods-931b97b6df3f

29
codes/A2C/agent.py
View File

@@ -1,32 +1,27 @@
#!/usr/bin/env python
# coding=utf-8
'''
Author: John
Author: JiangJi
Email: johnjim0816@gmail.com
Date: 2020-11-03 20:47:09
LastEditor: John
LastEditTime: 2021-03-20 17:41:21
Date: 2021-05-03 22:16:08
LastEditor: JiangJi
LastEditTime: 2021-05-03 22:23:48
Discription:
Environment:
'''
from A2C.model import ActorCritic
import torch.optim as optim
from A2C.model import ActorCritic
class A2C:
def __init__(self,state_dim, action_dim, cfg):
self.gamma = 0.99
self.model = ActorCritic(state_dim, action_dim, hidden_dim=cfg.hidden_dim).to(cfg.device)
self.optimizer = optim.Adam(self.model.parameters(),lr=cfg.lr)
def choose_action(self, state):
dist, value = self.model(state)
action = dist.sample()
return action
def __init__(self,state_dim,action_dim,cfg) -> None:
self.gamma = cfg.gamma
self.device = cfg.device
self.model = ActorCritic(state_dim, action_dim, cfg.hidden_size).to(self.device)
self.optimizer = optim.Adam(self.model.parameters())
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
return returns

48
codes/A2C/env.py
View File

@@ -1,48 +0,0 @@
#!/usr/bin/env python
# coding=utf-8
'''
Author: John
Email: johnjim0816@gmail.com
Date: 2020-10-30 15:39:37
LastEditor: John
LastEditTime: 2021-03-17 20:19:14
Discription:
Environment:
'''
import gym
from A2C.multiprocessing_env import SubprocVecEnv
# num_envs = 16
# env = "Pendulum-v0"
def make_envs(num_envs=16,env="Pendulum-v0"):
''' 创建多个子环境
'''
num_envs = 16
env = "CartPole-v0"
def make_env():
def _thunk():
env = gym.make(env)
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 = "CartPole-v0"
# def make_env():
# def _thunk():
# env = gym.make(env)
# 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="CartPole-v0")

106
codes/A2C/main.py
View File

@@ -1,106 +0,0 @@
#!/usr/bin/env python
# coding=utf-8
'''
@Author: John
@Email: johnjim0816@gmail.com
@Date: 2020-06-11 20:58:21
@LastEditor: John
LastEditTime: 2021-04-05 11:14:39
@Discription:
@Environment: python 3.7.9
'''
import sys,os
curr_path = os.path.dirname(__file__)
parent_path=os.path.dirname(curr_path)
sys.path.append(parent_path) # add current terminal path to sys.path
import torch
import gym
import datetime
from A2C.agent import A2C
from common.utils import save_results,make_dir,del_empty_dir
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/"):
os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/saved_model/")
if not os.path.exists(SAVED_MODEL_PATH):
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/"):
os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/results/")
if not os.path.exists(RESULT_PATH):
os.mkdir(RESULT_PATH)
class A2CConfig:
def __init__(self):
self.gamma = 0.99
self.lr = 3e-4 # learnning rate
self.actor_lr = 1e-4 # learnning rate of actor network
self.memory_capacity = 10000 # capacity of replay memory
self.batch_size = 128
self.train_eps = 200
self.train_steps = 200
self.eval_eps = 200
self.eval_steps = 200
self.target_update = 4
self.hidden_dim=256
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def train(cfg,env,agent):
print('Start to train ! ')
for i_episode in range(cfg.train_eps):
state = env.reset()
log_probs = []
values = []
rewards = []
masks = []
entropy = 0
ep_reward = 0
for i_step in range(cfg.train_steps):
state = torch.FloatTensor(state).to(cfg.device)
dist, value = agent.model(state)
action = dist.sample()
next_state, reward, done, _ = env.step(action.cpu().numpy())
ep_reward+=reward
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(cfg.device))
masks.append(torch.FloatTensor(1 - done).unsqueeze(1).to(cfg.device))
if done:
break
print('Episode:{}/{}, Reward:{}, Steps:{}, Done:{}'.format(i_episode+1,cfg.train_eps,ep_reward,i_step+1,done))
next_state = torch.FloatTensor(next_state).to(cfg.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()
print('Complete training！')
if __name__ == "__main__":
cfg = A2CConfig()
env = gym.make('CartPole-v0')
env.seed(1) # set random seed for env
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.n
agent = A2C(state_dim, action_dim, cfg)
train(cfg,env,agent)

22
codes/A2C/model.py
View File

@@ -1,36 +1,36 @@
#!/usr/bin/env python
# coding=utf-8
'''
Author: John
Author: JiangJi
Email: johnjim0816@gmail.com
Date: 2020-11-03 20:45:25
LastEditor: John
LastEditTime: 2021-03-20 17:41:33
Date: 2021-05-03 21:38:54
LastEditor: JiangJi
LastEditTime: 2021-05-03 21:40:06
Discription:
Environment:
'''
import torch.nn as nn
import torch.nn.functional as F
from torch.distributions import Categorical
class ActorCritic(nn.Module):
def __init__(self, state_dim, action_dim, hidden_dim=256):
def __init__(self, num_inputs, num_outputs, hidden_size, std=0.0):
super(ActorCritic, self).__init__()
self.critic = nn.Sequential(
nn.Linear(state_dim, hidden_dim),
nn.Linear(num_inputs, hidden_size),
nn.ReLU(),
nn.Linear(hidden_dim, 1)
nn.Linear(hidden_size, 1)
)
self.actor = nn.Sequential(
nn.Linear(state_dim, hidden_dim),
nn.Linear(num_inputs, hidden_size),
nn.ReLU(),
nn.Linear(hidden_dim, action_dim),
nn.Linear(hidden_size, num_outputs),
nn.Softmax(dim=1),
)
def forward(self, x):
value = self.critic(x)
print(x)
probs = self.actor(x)
dist = Categorical(probs)
return dist, value

153
codes/A2C/multiprocessing_env.py
View File

@@ -1,153 +0,0 @@
#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

BIN
codes/A2C/outputs/CartPole-v0/20210503-224814/results/train_ma_rewards.npy Normal file
View File

Binary file not shown.

BIN
codes/A2C/outputs/CartPole-v0/20210503-224814/results/train_rewards.npy Normal file
View File

Binary file not shown.

BIN
codes/A2C/outputs/CartPole-v0/20210503-224814/results/train_rewards_curve.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 62 KiB

120
codes/A2C/task0_train.py Normal file
View File

@@ -0,0 +1,120 @@
import sys,os
curr_path = os.path.dirname(__file__)
parent_path = os.path.dirname(curr_path)
sys.path.append(parent_path) # add current terminal path to sys.path
import gym
import numpy as np
import torch
import torch.optim as optim
import datetime
from common.multiprocessing_env import SubprocVecEnv
from A2C.model import ActorCritic
from common.utils import save_results, make_dir
from common.plot import plot_rewards
curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # obtain current time
class A2CConfig:
def __init__(self) -> None:
self.algo='A2C'
self.env= 'CartPole-v0'
self.result_path = curr_path+"/outputs/" +self.env+'/'+curr_time+'/results/' # path to save results
self.model_path = curr_path+"/outputs/" +self.env+'/'+curr_time+'/models/' # path to save models
self.n_envs = 8
self.gamma = 0.99
self.hidden_size = 256
self.lr = 1e-3 # learning rate
self.max_frames = 30000
self.n_steps = 5
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def make_envs(env_name):
def _thunk():
env = gym.make(env_name)
env.seed(2)
return env
return _thunk
def test_env(env,model,vis=False):
state = env.reset()
if vis: env.render()
done = False
total_reward = 0
while not done:
state = torch.FloatTensor(state).unsqueeze(0).to(cfg.device)
dist, _ = model(state)
next_state, reward, done, _ = env.step(dist.sample().cpu().numpy()[0])
state = next_state
if vis: env.render()
total_reward += reward
return total_reward
def compute_returns(next_value, rewards, masks, gamma=0.99):
R = next_value
returns = []
for step in reversed(range(len(rewards))):
R = rewards[step] + gamma * R * masks[step]
returns.insert(0, R)
return returns
def train(cfg,envs):
env = gym.make(cfg.env) # a single env
env.seed(10)
state_dim = envs.observation_space.shape[0]
action_dim = envs.action_space.n
model = ActorCritic(state_dim, action_dim, cfg.hidden_size).to(cfg.device)
optimizer = optim.Adam(model.parameters())
frame_idx = 0
test_rewards = []
test_ma_rewards = []
state = envs.reset()
while frame_idx < cfg.max_frames:
log_probs = []
values = []
rewards = []
masks = []
entropy = 0
# rollout trajectory
for _ in range(cfg.n_steps):
state = torch.FloatTensor(state).to(cfg.device)
dist, value = model(state)
action = dist.sample()
next_state, reward, done, _ = envs.step(action.cpu().numpy())
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(cfg.device))
masks.append(torch.FloatTensor(1 - done).unsqueeze(1).to(cfg.device))
state = next_state
frame_idx += 1
if frame_idx % 100 == 0:
test_reward = np.mean([test_env(env,model) for _ in range(10)])
print(f"frame_idx:{frame_idx}, test_reward:{test_reward}")
test_rewards.append(test_reward)
if test_ma_rewards:
test_ma_rewards.append(0.9*test_ma_rewards[-1]+0.1*test_reward)
else:
test_ma_rewards.append(test_reward)
# plot(frame_idx, test_rewards)
next_state = torch.FloatTensor(next_state).to(cfg.device)
_, next_value = model(next_state)
returns = 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
optimizer.zero_grad()
loss.backward()
optimizer.step()
return test_rewards, test_ma_rewards
if __name__ == "__main__":
cfg = A2CConfig()
envs = [make_envs(cfg.env) for i in range(cfg.n_envs)]
envs = SubprocVecEnv(envs) # 8 env
rewards,ma_rewards = train(cfg,envs)
make_dir(cfg.result_path,cfg.model_path)
save_results(rewards,ma_rewards,tag='train',path=cfg.result_path)
plot_rewards(rewards,ma_rewards,tag="train",env=cfg.env,algo = cfg.algo,path=cfg.result_path)

32
codes/A2C/utils.py
View File

@@ -1,32 +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: 2020-11-03 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]+"/results/"+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！')