def demo1__discrete_action_space():
"""DEMO 1: Discrete action env: CartPole-v0 of gym"""
args = Arguments(agent_rl=None, env=None, gpu_id=None) # see Arguments() to see hyper-parameters
args.agent_rl = agent.AgentD3QN # choose an DRL algorithm
args.env = decorate_env(env=gym.make('CartPole-v0'))
args.net_dim = 2 ** 7 # change a default hyper-parameters
# args.env = decorate_env(env=gym.make('LunarLander-v2'))
# args.net_dim = 2 ** 8 # change a default hyper-parameters
train_and_evaluate(args)
def demo5():
args = Arguments(if_on_policy=False)
# args.agent_rl = agent.AgentModSAC
args.agent_rl = agent.AgentInterSAC
import pybullet_envs # for python-bullet-gym
dir(pybullet_envs)
args.env = decorate_env(gym.make('AntBulletEnv-v0'))
# args.env = decorate_env(gym.make('ReacherBulletEnv-v0'))
args.break_step = int(1e6 * 8) # (5e5) 1e6, UsedTime: (15,000s) 30,000s
args.reward_scale = 2 ** -2 # (-50) 0 ~ 2500 (3340)
args.max_memo = 2 ** 19
args.net_dim = 2 ** 7 # todo
args.eva_size = 2 ** 5 # for Recorder
args.show_gap = 2 ** 8 # for Recorder
train_and_evaluate(args)
def render_pybullet():
from eRL import agent
args = Arguments(if_on_policy=True)
args.agent_rl = agent.AgentGaePPO # agent.AgentPPO
import pybullet_envs # for python-bullet-gym
dir(pybullet_envs)
args.env = decorate_env(gym.make('ReacherBulletEnv-v0'))
args.gpu_id = 3
args.init_before_training()
net_dim = args.net_dim
state_dim = args.env.state_dim
action_dim = args.env.action_dim
# cwd = args.cwd
cwd = "ReacherBulletEnv-v0_2"
env = args.env
max_step = args.max_step
agent = args.agent_rl(net_dim, state_dim, action_dim)
agent.save_or_load_model(cwd, if_save=False) # if_save=False means load history
import cv2
import os
frame_save_dir = f'{cwd}/frame'
os.makedirs(frame_save_dir, exist_ok=True)
'''methods 1: Print Error in remote server: Invalid GLX version: major 1, minor 2'''
# import pybullet as pb
# phy = pb.connect(pb.GUI)
'''methods 2: Print Error in remote server: Invalid GLX version: major 1, minor 2'''
env.render() # https://github.com/benelot/pybullet-gym/issues/25
state = env.reset()
for i in range(max_step):
print(i)
actions, _noises = agent.select_actions((state,))
action = actions[0]
next_state, reward, done, _ = env.step(action)
frame = env.render()
if isinstance(frame, np.ndarray):
cv2.imshow(env.env_name, frame)
cv2.waitKey(20) # 1000ms = 1s
cv2.imwrite(f'{frame_save_dir}/{i:06}.png', frame)
if done:
break
state = next_state
print('end')
def demo41():
args = Arguments(if_on_policy=True)
args.agent_rl = agent.AgentGaePPO # agent.AgentPPO
import pybullet_envs # for python-bullet-gym
dir(pybullet_envs)
args.env = decorate_env(gym.make('ReacherBulletEnv-v0'))
args.break_step = int(5e4 * 8) # (5e4) 1e5, UsedTime: (400s) 800s
args.repeat_times = 2 ** 3
args.reward_scale = 2 ** 1 # (-15) 18 (30)
args.eval_times1 = 2 ** 2
args.eval_times1 = 2 ** 6
args.rollout_num = 4
train_and_evaluate__multiprocessing(args)
def demo2():
"""DEMO 2: Continuous action env, gym.Box2D"""
if_on_policy = False
args = Arguments(if_on_policy=if_on_policy) # on-policy has different hyper-parameters from off-policy
if if_on_policy:
args.agent_rl = agent.AgentGaePPO # on-policy: AgentPPO, AgentGaePPO
else:
args.agent_rl = agent.AgentModSAC # off-policy: AgentSAC, AgentModPPO, AgentTD3, AgentDDPG
env = gym.make('Pendulum-v0')
env.target_reward = -200 # set target_reward manually for env 'Pendulum-v0'
args.env = decorate_env(env=env)
args.net_dim = 2 ** 7 # change a default hyper-parameters
# args.env = decorate_env(env=gym.make('LunarLanderContinuous-v2'))
# args.env = decorate_env(env=gym.make('BipedalWalker-v3')) # recommend args.gamma = 0.95
train_and_evaluate(args)
def demo42():
args = Arguments(if_on_policy=True)
args.agent_rl = agent.AgentGaePPO # agent.AgentPPO
import pybullet_envs # for python-bullet-gym
dir(pybullet_envs)
args.env = decorate_env(gym.make('AntBulletEnv-v0'))
args.break_step = int(5e6 * 8) # (1e6) 5e6 UsedTime: 25697s
args.reward_scale = 2 ** -3 #
args.repeat_times = 2 ** 4
args.net_dim = 2 ** 9
args.batch_size = 2 ** 8
args.max_memo = 2 ** 12
args.show_gap = 2 ** 6
args.eval_times1 = 2 ** 2
args.rollout_num = 4
train_and_evaluate__multiprocessing(args)
def demo3():
"""DEMO 3: Custom Continuous action env: FinanceStock-v1"""
args = Arguments(if_on_policy=True)
args.agent_rl = agent.AgentGaePPO # PPO+GAE (on-policy)
from eRL.env import FinanceMultiStockEnv
args.env = FinanceMultiStockEnv(if_train=True) # a standard env for ElegantRL, not need decorate_env()
args.env_eval = FinanceMultiStockEnv(if_train=False)
args.break_step = int(5e6) # 5e6 (15e6) UsedTime 3,000s (9,000s)
args.net_dim = 2 ** 8
args.max_step = args.env.max_step
args.max_memo = (args.max_step - 1) * 8
args.batch_size = 2 ** 11
args.repeat_times = 2 ** 4
args.eval_times1 = 2 ** 3
# train_and_evaluate(args)
args.rollout_num = 8
args.if_break_early = False
train_and_evaluate__multiprocessing(args)