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
2021-03-23 16:10:11 +08:00
parent d4690c2058
commit bf0f2990cf
198 changed files with 1668 additions and 1545 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
codes/A2C/.vscode/settings.json vendored
View File

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

11
codes/A2C/agent.py
View File

@@ -5,19 +5,18 @@ Author: John
Email: johnjim0816@gmail.com
Date: 2020-11-03 20:47:09
LastEditor: John
LastEditTime: 2020-11-08 22:16:29
LastEditTime: 2021-03-20 17:41:21
Discription:
Environment:
'''
from model import ActorCritic
from A2C.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
def __init__(self,n_states, n_actions, cfg):
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)
self.model = ActorCritic(n_states, n_actions, 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()

0
codes/A2C/common/__init__.py
View File

4
codes/A2C/env.py
View File

@@ -5,13 +5,13 @@ Author: John
Email: johnjim0816@gmail.com
Date: 2020-10-30 15:39:37
LastEditor: John
LastEditTime: 2020-11-03 20:52:07
LastEditTime: 2021-03-17 20:19:14
Discription:
Environment:
'''
import gym
from common.multiprocessing_env import SubprocVecEnv
from A2C.multiprocessing_env import SubprocVecEnv
# num_envs = 16
# env_name = "Pendulum-v0"

200
codes/A2C/main.py
View File

@@ -5,94 +5,73 @@
@Email: johnjim0816@gmail.com
@Date: 2020-06-11 20:58:21
@LastEditor: John
LastEditTime: 2020-11-08 22:19:56
LastEditTime: 2021-03-20 16:58:04
@Discription:
@Environment: python 3.7.9
'''
import sys,os
sys.path.append(os.getcwd()) # add current terminal path
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
import datetime
from A2C.agent import A2C
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):
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
for i_step in range(1, cfg.train_steps+1):
state = torch.FloatTensor(state).to(device)
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, _ = envs.step(action.cpu().numpy())
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(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)
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)
@@ -107,80 +86,17 @@ def train(cfg):
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)
cfg = A2CConfig()
env = gym.make('CartPole-v0')
env.seed(1) # set random seed for env
n_states = env.observation_space.shape[0]
n_actions = env.action_space.n
agent = A2C(n_states, n_actions, cfg)
train(cfg,env,agent)

5
codes/A2C/model.py
View File

@@ -5,7 +5,7 @@ Author: John
Email: johnjim0816@gmail.com
Date: 2020-11-03 20:45:25
LastEditor: John
LastEditTime: 2020-11-07 18:49:09
LastEditTime: 2021-03-20 17:41:33
Discription:
Environment:
'''
@@ -13,7 +13,7 @@ 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):
def __init__(self, n_states, n_actions, hidden_dim=256):
super(ActorCritic, self).__init__()
self.critic = nn.Sequential(
nn.Linear(n_states, hidden_dim),
@@ -30,6 +30,7 @@ class ActorCritic(nn.Module):
def forward(self, x):
value = self.critic(x)
print(x)
probs = self.actor(x)
dist = Categorical(probs)
return dist, value

0
codes/A2C/common/multiprocessing_env.py → codes/A2C/multiprocessing_env.py
View File

162
codes/A2C/test.py Normal file
View File

@@ -0,0 +1,162 @@
#!/usr/bin/env python
# coding=utf-8
'''
Author: John
Email: johnjim0816@gmail.com
Date: 2021-03-20 17:43:17
LastEditor: John
LastEditTime: 2021-03-20 19:36:24
Discription:
Environment:
'''
import sys
import torch
import gym
import numpy as np
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.autograd import Variable
import matplotlib.pyplot as plt
import pandas as pd
learning_rate = 3e-4
# Constants
GAMMA = 0.99
class A2CConfig:
''' hyperparameters
'''
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 = 3000
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")
class ActorCritic(nn.Module):
def __init__(self, n_states, n_actions, hidden_dim, learning_rate=3e-4):
super(ActorCritic, self).__init__()
self.n_actions = n_actions
self.critic_linear1 = nn.Linear(n_states, hidden_dim)
self.critic_linear2 = nn.Linear(hidden_dim, 1)
self.actor_linear1 = nn.Linear(n_states, hidden_dim)
self.actor_linear2 = nn.Linear(hidden_dim, n_actions)
def forward(self, state):
state = Variable(torch.from_numpy(state).float().unsqueeze(0))
value = F.relu(self.critic_linear1(state))
value = self.critic_linear2(value)
policy_dist = F.relu(self.actor_linear1(state))
policy_dist = F.softmax(self.actor_linear2(policy_dist), dim=1)
return value, policy_dist
class A2C:
def __init__(self,n_states,n_actions,cfg):
self.model = ActorCritic(n_states, n_actions, cfg.hidden_dim)
self.optimizer = optim.Adam(self.model.parameters(), lr=cfg.lr)
def choose_action(self,state):
pass
def update(self):
pass
def train(cfg,env,agent):
n_states = env.observation_space.shape[0]
n_actions = env.action_space.n
actor_critic = ActorCritic(n_states, n_actions, hidden_dim)
ac_optimizer = optim.Adam(actor_critic.parameters(), lr=learning_rate)
all_lengths = []
average_lengths = []
all_rewards = []
entropy_term = 0
for episode in range(cfg.train_eps):
log_probs = []
values = []
rewards = []
state = env.reset()
for steps in range(cfg.train_steps):
value, policy_dist = actor_critic.forward(state)
value = value.detach().numpy()[0,0]
dist = policy_dist.detach().numpy()
action = np.random.choice(n_actions, p=np.squeeze(dist))
log_prob = torch.log(policy_dist.squeeze(0)[action])
entropy = -np.sum(np.mean(dist) * np.log(dist))
new_state, reward, done, _ = env.step(action)
rewards.append(reward)
values.append(value)
log_probs.append(log_prob)
entropy_term += entropy
state = new_state
if done or steps == cfg.train_steps-1:
Qval, _ = actor_critic.forward(new_state)
Qval = Qval.detach().numpy()[0,0]
all_rewards.append(np.sum(rewards))
all_lengths.append(steps)
average_lengths.append(np.mean(all_lengths[-10:]))
if episode % 10 == 0:
sys.stdout.write("episode: {}, reward: {}, total length: {}, average length: {} \n".format(episode, np.sum(rewards), steps, average_lengths[-1]))
break
# compute Q values
Qvals = np.zeros_like(values)
for t in reversed(range(len(rewards))):
Qval = rewards[t] + GAMMA * Qval
Qvals[t] = Qval
#update actor critic
values = torch.FloatTensor(values)
Qvals = torch.FloatTensor(Qvals)
log_probs = torch.stack(log_probs)
advantage = Qvals - values
actor_loss = (-log_probs * advantage).mean()
critic_loss = 0.5 * advantage.pow(2).mean()
ac_loss = actor_loss + critic_loss + 0.001 * entropy_term
ac_optimizer.zero_grad()
ac_loss.backward()
ac_optimizer.step()
# Plot results
smoothed_rewards = pd.Series.rolling(pd.Series(all_rewards), 10).mean()
smoothed_rewards = [elem for elem in smoothed_rewards]
plt.plot(all_rewards)
plt.plot(smoothed_rewards)
plt.plot()
plt.xlabel('Episode')
plt.ylabel('Reward')
plt.show()
plt.plot(all_lengths)
plt.plot(average_lengths)
plt.xlabel('Episode')
plt.ylabel('Episode length')
plt.show()
if __name__ == "__main__":
cfg = A2CConfig
env = gym.make("CartPole-v0")
n_states = env.observation_space.shape[0]
n_actions = env.action_space.n
agent = A2C(n_states,n_actions,cfg)
train(cfg,env,agent)

2
codes/A2C/utils.py
View File

@@ -15,7 +15,7 @@ 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+'/'
RESULT_PATH = os.path.split(os.path.abspath(__file__))[0]+"/results/"+SEQUENCE+'/'
def save_results(rewards,moving_average_rewards,ep_steps,path=RESULT_PATH):