update
This commit is contained in:
JohnJim0816
@@ -5,37 +5,58 @@
|
||||
@Email: johnjim0816@gmail.com
|
||||
@Date: 2020-06-12 00:48:57
|
||||
@LastEditor: John
|
||||
LastEditTime: 2020-12-22 15:39:46
|
||||
LastEditTime: 2021-03-17 20:11:19
|
||||
@Discription:
|
||||
@Environment: python 3.7.7
|
||||
'''
|
||||
import sys,os
|
||||
sys.path.append(os.getcwd()) # add current terminal path
|
||||
import gym
|
||||
import torch
|
||||
from torch.utils.tensorboard import SummaryWriter
|
||||
import os
|
||||
from agent import DQN
|
||||
from params import SEQUENCE,SAVED_MODEL_PATH,RESULT_PATH
|
||||
from params import get_args
|
||||
from utils import save_results
|
||||
import datetime
|
||||
from DoubleDQN.agent import DoubleDQN
|
||||
from common.plot import plot_rewards
|
||||
from common.utils import save_results
|
||||
|
||||
def train(cfg):
|
||||
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 DoubleDQNConfig:
|
||||
def __init__(self):
|
||||
self.algo = "Double DQN" # 算法名称
|
||||
self.gamma = 0.99
|
||||
self.epsilon_start = 0.9 # e-greedy策略的初始epsilon
|
||||
self.epsilon_end = 0.01
|
||||
self.epsilon_decay = 200
|
||||
self.lr = 0.01 # 学习率
|
||||
self.memory_capacity = 10000 # Replay Memory容量
|
||||
self.batch_size = 128
|
||||
self.train_eps = 250 # 训练的episode数目
|
||||
self.train_steps = 200 # 训练每个episode的最大长度
|
||||
self.target_update = 2 # target net的更新频率
|
||||
self.eval_eps = 20 # 测试的episode数目
|
||||
self.eval_steps = 200 # 测试每个episode的最大长度
|
||||
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测gpu
|
||||
self.hidden_dim = 128 # 神经网络隐藏层维度
|
||||
|
||||
|
||||
def train(cfg,env,agent):
|
||||
print('Start to train !')
|
||||
device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测gpu
|
||||
env = gym.make('CartPole-v0').unwrapped # 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]
|
||||
n_actions = env.action_space.n
|
||||
agent = DQN(n_states=n_states, n_actions=n_actions, device=device, gamma=cfg.gamma, epsilon_start=cfg.epsilon_start,
|
||||
epsilon_end=cfg.epsilon_end, epsilon_decay=cfg.epsilon_decay, policy_lr=cfg.policy_lr, memory_capacity=cfg.memory_capacity, batch_size=cfg.batch_size)
|
||||
rewards = []
|
||||
moving_average_rewards = []
|
||||
rewards,ma_rewards = [],[]
|
||||
ep_steps = []
|
||||
log_dir=os.path.split(os.path.abspath(__file__))[0]+"/logs/train/" + SEQUENCE
|
||||
writer = SummaryWriter(log_dir)
|
||||
for i_episode in range(1, cfg.train_eps+1):
|
||||
for i_episode in range(cfg.train_eps):
|
||||
state = env.reset() # reset环境状态
|
||||
ep_reward = 0
|
||||
for i_step in range(1, cfg.train_steps+1):
|
||||
for i_step in range(cfg.train_steps):
|
||||
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
|
||||
@@ -47,80 +68,26 @@ def train(cfg):
|
||||
# 更新target network，复制DQN中的所有weights and biases
|
||||
if i_episode % cfg.target_update == 0:
|
||||
agent.target_net.load_state_dict(agent.policy_net.state_dict())
|
||||
print('Episode:', i_episode, ' Reward: %i' %
|
||||
int(ep_reward), 'n_steps:', i_step, 'done: ', done,' Explore: %.2f' % agent.epsilon)
|
||||
print('Episode:{}/{}, Reward:{}, Steps:{}, Done:{}'.format(i_episode+1,cfg.train_eps,ep_reward,i_step,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)
|
||||
if ma_rewards:
|
||||
ma_rewards.append(
|
||||
0.9*ma_rewards[-1]+0.1*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()
|
||||
ma_rewards.append(ep_reward)
|
||||
print('Complete training！')
|
||||
''' 保存模型 '''
|
||||
if not os.path.exists(SAVED_MODEL_PATH): # 检测是否存在文件夹
|
||||
os.mkdir(SAVED_MODEL_PATH)
|
||||
agent.save_model(SAVED_MODEL_PATH+'checkpoint.pth')
|
||||
print('model saved！')
|
||||
'''存储reward等相关结果'''
|
||||
save_results(rewards,moving_average_rewards,ep_steps,tag='train',result_path=RESULT_PATH)
|
||||
return rewards,ma_rewards
|
||||
|
||||
|
||||
def eval(cfg, saved_model_path = SAVED_MODEL_PATH):
|
||||
print('start to eval !')
|
||||
device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测gpu
|
||||
if __name__ == "__main__":
|
||||
cfg = DoubleDQNConfig()
|
||||
env = gym.make('CartPole-v0').unwrapped # 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]
|
||||
n_actions = env.action_space.n
|
||||
agent = DQN(n_states=n_states, n_actions=n_actions, device=device, gamma=cfg.gamma, epsilon_start=cfg.epsilon_start,
|
||||
epsilon_end=cfg.epsilon_end, epsilon_decay=cfg.epsilon_decay, policy_lr=cfg.policy_lr, memory_capacity=cfg.memory_capacity, batch_size=cfg.batch_size)
|
||||
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,train=False) # 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'''
|
||||
save_results(rewards,moving_average_rewards,ep_steps,tag='eval',result_path=RESULT_PATH)
|
||||
print('Complete evaling！')
|
||||
|
||||
if __name__ == "__main__":
|
||||
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)
|
||||
agent = DoubleDQN(n_states,n_actions,cfg)
|
||||
rewards,ma_rewards = train(cfg,env,agent)
|
||||
agent.save(path=SAVED_MODEL_PATH)
|
||||
save_results(rewards,ma_rewards,tag='train',path=RESULT_PATH)
|
||||
plot_rewards(rewards,ma_rewards,tag="train",algo = cfg.algo,path=RESULT_PATH)
|
||||
|
||||
Reference in New Issue
Block a user