def run(lock, shared_eps_num, shared_eps_reward, hyper_param):
model = SAC(
env_name='Pendulum-v0',
load_dir='./ckpt/ckpt_' + str(hyper_param[0]) + '_' +
str(hyper_param[1]),
log_dir='./log/log_' + str(hyper_param[0]) + '_' + str(hyper_param[1]),
buffer_size=1e6,
seed=hyper_param[1],
max_episode_steps=500, # manual set
batch_size=hyper_param[0],
discount=0.99,
learning_starts=500,
tau=0.005,
save_eps_num=100)
timesteps = 0
total_timesteps = 1e5
max_eps_steps = 100
# train
while timesteps < total_timesteps:
episode_reward = 0
done = False
eps_steps = 0
obs = model.env.reset()
while not done and eps_steps < max_eps_steps:
action = model.predict(obs)
new_obs, reward, done, info = model.env.step(action)
model.replay_buffer.push(obs, action, reward, new_obs, done)
obs = new_obs
episode_reward += reward
eps_steps += 1
timesteps += 1
if timesteps > model.learning_starts:
model.train_step()
model.episode_num += 1
model.writer.add_scalar('episode_reward', episode_reward,
model.episode_num)
lock.acquire()
shared_eps_num.value = model.episode_num
shared_eps_reward.value = episode_reward
lock.release()
from env import Env
import numpy as np
# from rllite.common import choose_gpu, GymDelay
# choose your GPU if you have more than one
# choose_gpu(0)
env = Env(10666)
# set
model = SAC(
external_env=env, # import your env
env_name="ssl_env", # your env name
load_dir='./ckpt',
log_dir="./log",
buffer_size=1e6,
seed=2,
max_episode_steps=500, # manual set
batch_size=64,
discount=0.99,
learning_starts=1000,
tau=0.005,
save_eps_num=100)
# model.learn(1e6)
timesteps = 0
total_timesteps = 1e6
max_eps_steps = 500
# tricky limit
xy_acc = 4.0 / 75.0 / 2
from rllite import SAC
model = SAC('Pendulum-v0').learn(1e6)
from rllite import SAC
# set
model = SAC(env_name='Pendulum-v0',
load_dir='./ckpt',
log_dir="./log",
buffer_size=1e6,
seed=1,
max_episode_steps=None,
batch_size=64,
discount=0.99,
learning_starts=500,
tau=0.005,
save_eps_num=100)
# train
model.learn(1e6)
# eval
for _ in range(10):
done = False
obs = model.env.reset()
while not done:
action = model.predict(obs)
obs, reward, done, info = model.env.step(action)
model.env.render()
model.env.close()