def demo_custom_env_finance_rl():
from elegantrl2.agent import AgentPPO
'''choose an DRL algorithm'''
args = Arguments(if_on_policy=True)
args.agent = AgentPPO()
args.agent.lambda_entropy = 0.02 # todo ceta2
args.gpu_id = sys.argv[-1][-4]
args.random_seed = 1943210
"TotalStep: 10e4, TargetReturn: 3.0, UsedTime: 200s, FinanceStock-v1"
"TotalStep: 20e4, TargetReturn: 4.0, UsedTime: 400s, FinanceStock-v1"
"TotalStep: 30e4, TargetReturn: 4.2, UsedTime: 600s, FinanceStock-v1"
# from envs.FinRL.StockTrading import StockTradingEnv
args.gamma = 0.999
# args.env = StockTradingEnv(if_eval=False, gamma=gamma)
args.env = StockTradingVecEnv(if_eval=False, gamma=args.gamma, env_num=2)
args.env_eval = StockTradingEnv(if_eval=True, gamma=args.gamma)
args.net_dim = 2**9
args.batch_size = args.net_dim * 4
args.target_step = args.env.max_step * 2 # todo ceta0
args.repeat_times = 2**4
args.eval_gap = 2**8
args.eval_times1 = 2**0
args.eval_times2 = 2**1
args.break_step = int(16e6)
'''train and evaluate'''
# train_and_evaluate(args)
args.worker_num = 2
train_and_evaluate_mp(args)
def demo_continuous_action_on_policy():
args = Arguments(
if_on_policy=True
) # hyper-parameters of on-policy is different from off-policy
from elegantrl2.agent import AgentPPO
args.agent = AgentPPO()
args.agent.cri_target = True
'''choose environment'''
if_train_pendulum = 0
if if_train_pendulum:
"TotalStep: 4e5, TargetReward: -200, UsedTime: 400s"
env = gym.make('Pendulum-v0')
env.target_return = -200 # set target_reward manually for env 'Pendulum-v0'
args.env = PreprocessEnv(env=env)
args.reward_scale = 2**-3 # RewardRange: -1800 < -200 < -50 < 0
args.net_dim = 2**7
args.batch_size = args.net_dim * 2
args.target_step = args.env.max_step * 16
if_train_lunar_lander = 0
if if_train_lunar_lander:
"TotalStep: 4e5, TargetReward: 200, UsedTime: 900s"
args.env = PreprocessEnv(env=gym.make('LunarLanderContinuous-v2'))
args.reward_scale = 2**0 # RewardRange: -800 < -200 < 200 < 302
if_train_bipedal_walker = 1
if if_train_bipedal_walker:
"TotalStep: 8e5, TargetReward: 300, UsedTime: 1800s"
args.env = PreprocessEnv(env=gym.make('BipedalWalker-v3'))
args.reward_scale = 2**0 # RewardRange: -200 < -150 < 300 < 334
args.gamma = 0.97
args.if_per_or_gae = True
# args.agent.lambda_entropy = 0.05
'''train and evaluate'''
train_and_evaluate(args)
def check_agent():
from elegantrl2.agent import AgentPPO as Agent
agent = Agent()
net_dim = 2**7
state_dim = 8
action_dim = 2
agent.init(net_dim,
state_dim,
action_dim,
learning_rate=1e-4,
if_use_gae=False,
gpu_id=0)
# print(agent.act.state_dict())
for key, value in agent.act.state_dict().items():
print(key, value.shape)
def demo_continuous_action_on_policy_temp_mg():
args = Arguments(if_on_policy=True) # hyper-parameters of on-policy is different from off-policy
from elegantrl2.agent import AgentPPO
args.agent = AgentPPO()
args.agent.cri_target = False # todo ceta2
args.learning_rate = 2 ** -14
args.random_seed = 1943
# args.gpu_id = (0, 1, 2, 3)
args.gpu_id = (0, 1) # (2, 3)
# args.gpu_id = (2, 3)
# args.gpu_id = int(sys.argv[-1][-4])
args.random_seed = 1549
'''choose environment'''
if_train_pendulum = 0
if if_train_pendulum:
"TotalStep: 4e5, TargetReward: -200, UsedTime: 400s"
env = gym.make('Pendulum-v0')
env.target_return = -200 # set target_reward manually for env 'Pendulum-v0'
args.env = PreprocessEnv(env=env)
args.reward_scale = 2 ** -3 # RewardRange: -1800 < -200 < -50 < 0
args.net_dim = 2 ** 7
args.batch_size = args.net_dim * 2
args.target_step = args.env.max_step * 16
if_train_bipedal_walker = 1
if if_train_bipedal_walker:
"TotalStep: 8e5, TargetReward: 300, UsedTime: 1800s"
env_name = 'BipedalWalker-v3'
env = gym.make(env_name)
env.target_return = -50 # todo test
args.eval_gap = 2 ** 5 # todo test
args.env = PreprocessEnv(env=env)
# args.env = PreprocessVecEnv(env=env, env_num=2)
# args.env_eval = PreprocessEnv(env=env_name)
args.reward_scale = 2 ** 0 # RewardRange: -200 < -150 < 300 < 334
args.gamma = 0.97
args.target_step = args.env.max_step * 4
args.repeat_times = 2 ** 4
args.if_per_or_gae = True
args.agent.lambda_entropy = 0.05
args.break_step = int(8e6)
# if_train_finance_rl = 1
# if if_train_finance_rl:
'''train and evaluate'''
# train_and_evaluate(args)
args.worker_num = 2
if isinstance(args.gpu_id, int) or isinstance(args.gpu_id, str):
train_and_evaluate_mp(args)
elif isinstance(args.gpu_id, tuple) or isinstance(args.gpu_id, list):
train_and_evaluate_mg(args)
else:
print(f"Error in args.gpu_id {args.gpu_id}, type {type(args.gpu_id)}")
def demo_custom_env_finance_rl_nas89():
args = Arguments(
if_on_policy=True
) # hyper-parameters of on-policy is different from off-policy
args.random_seed = 1943
from elegantrl2.agent import AgentPPO
args.agent = AgentPPO()
args.agent.cri_target = True
args.agent.lambda_entropy = 0.04
from envs.FinRL.StockTrading import StockEnvDOW30, StockEnvNAS89, StockVecEnvNAS89
args.gamma = 0.999
if_dow30_daily = 1
if if_dow30_daily:
args.env = StockEnvDOW30(if_eval=False, gamma=args.gamma)
args.env_eval = StockEnvDOW30(if_eval=True, gamma=args.gamma)
else: # elif if_nas89_minute:
args.env = StockEnvNAS89(if_eval=False, gamma=args.gamma)
args.env_eval = StockEnvNAS89(if_eval=True, gamma=args.gamma)
args.repeat_times = 2**4
args.learning_rate = 2**-14
args.net_dim = int(2**8 * 1.5)
args.batch_size = args.net_dim * 4
args.target_step = args.env.max_step
args.eval_gap = 2**8
args.eval_times1 = 2**0
args.eval_times2 = 2**1
args.break_step = int(16e6)
args.if_allow_break = False
if_single_env = 1
if if_single_env:
args.gpu_id = 0
args.worker_num = 4
train_and_evaluate_mp(args)
if_batch_env = 0
if if_batch_env:
args.env = StockVecEnvNAS89(if_eval=False, gamma=args.gamma, env_num=2)
args.gpu_id = 3
args.random_seed += args.gpu_id
args.worker_num = 2
train_and_evaluate_mp(args)
if_multi_learner = 0
if if_multi_learner:
args.env = StockVecEnvNAS89(if_eval=False, gamma=args.gamma, env_num=2)
args.gpu_id = (0, 1)
args.worker_num = 2
train_and_evaluate_mg(args)
def demo_continuous_action_on_policy():
args = Arguments(
if_on_policy=True
) # hyper-parameters of on-policy is different from off-policy
from elegantrl2.agent import AgentPPO
args.agent = AgentPPO()
args.gpu_id = sys.argv[-1][-4]
args.agent.cri_target = True
args.learning_rate = 2**-14
args.random_seed = 1943
'''choose environment'''
if_train_pendulum = 0
if if_train_pendulum:
"TotalStep: 4e5, TargetReward: -200, UsedTime: 400s"
env = gym.make('Pendulum-v0')
env.target_return = -200 # set target_reward manually for env 'Pendulum-v0'
args.env = PreprocessEnv(env=env)
args.reward_scale = 2**-3 # RewardRange: -1800 < -200 < -50 < 0
args.net_dim = 2**7
args.batch_size = args.net_dim * 2
args.target_step = args.env.max_step * 16
if_train_lunar_lander = 0
if if_train_lunar_lander:
"TotalStep: 6e5, TargetReward: 200, UsedTime: 800s"
env_name = 'LunarLanderContinuous-v2'
# args.env = PreprocessEnv(env=env_name)
args.env = PreprocessVecEnv(env=env_name, env_num=2)
args.env_eval = PreprocessEnv(env=env_name)
args.reward_scale = 2**0 # RewardRange: -800 < -200 < 200 < 302
args.break_step = int(8e6)
args.if_per_or_gae = True
args.target_step = args.env.max_step * 8
args.repeat_times = 2**4
if_train_bipedal_walker = 1
if if_train_bipedal_walker:
"TotalStep: 8e5, TargetReward: 300, UsedTime: 1800s"
env_name = 'BipedalWalker-v3'
# args.env = PreprocessEnv(env=env_name)
args.env = PreprocessVecEnv(env=env_name, env_num=2)
args.env_eval = PreprocessEnv(env=env_name)
args.reward_scale = 2**0 # RewardRange: -200 < -150 < 300 < 334
args.gamma = 0.97
args.target_step = args.env.max_step * 8
args.repeat_times = 2**4
args.if_per_or_gae = True
args.agent.lambda_entropy = 0.05
args.break_step = int(8e6)
'''train and evaluate'''
# train_and_evaluate(args)
args.worker_num = 2
train_and_evaluate_mp(args)
def demo_custom_env_finance_rl_dow30(): # 1.7+ 2.0+
args = Arguments(
if_on_policy=True
) # hyper-parameters of on-policy is different from off-policy
args.random_seed = 19430
from elegantrl2.agent import AgentPPO
args.agent = AgentPPO()
args.agent.cri_target = True
args.agent.lambda_entropy = 0.02
args.gamma = 0.995
from envs.FinRL.StockTrading import StockEnvDOW30, StockVecEnvDOW30
args.env = StockEnvDOW30(if_eval=False, gamma=args.gamma)
args.env_eval = StockEnvDOW30(if_eval=True, gamma=args.gamma)
args.repeat_times = 2**4
args.learning_rate = 2**-14
args.net_dim = 2**8
args.batch_size = args.net_dim
args.eval_gap = 2**7
args.eval_times1 = 2**0
args.eval_times2 = 2**1
args.break_step = int(10e6)
args.if_allow_break = False
if_single_env = 0
if if_single_env:
args.gpu_id = int(sys.argv[-1][-4])
args.random_seed += int(args.gpu_id)
args.target_step = args.env.max_step * 4
args.worker_num = 4
train_and_evaluate_mp(args)
if_batch_env = 1
if if_batch_env:
args.env = StockVecEnvDOW30(if_eval=False, gamma=args.gamma, env_num=4)
args.gpu_id = int(sys.argv[-1][-4])
args.random_seed += args.gpu_id
args.target_step = args.env.max_step
args.worker_num = 4
train_and_evaluate_mp(args)
if_multi_learner = 0
if if_multi_learner:
args.env = StockVecEnvDOW30(if_eval=False, gamma=args.gamma, env_num=2)
args.gpu_id = (0, 1)
args.worker_num = 2
train_and_evaluate_mg(args)
def demo_custom_env_finance_rl():
args = Arguments(
if_on_policy=True
) # hyper-parameters of on-policy is different from off-policy
args.random_seed = 0
from elegantrl2.agent import AgentPPO
args.agent = AgentPPO()
args.agent.cri_target = True
args.agent.lambda_entropy = 0.04
from envs.FinRL.StockTrading import StockEnvNAS89, StockVecEnvNAS89
args.gamma = 0.999
args.env = StockEnvNAS89(if_eval=False, gamma=args.gamma)
args.env_eval = StockEnvNAS89(if_eval=True, gamma=args.gamma)
args.repeat_times = 2**4
args.learning_rate = 2**-14
args.net_dim = int(2**8 * 1.5)
args.batch_size = args.net_dim * 4
if_single_env = 0
if if_single_env:
args.gpu_id = 0
args.worker_num = 4
train_and_evaluate_mp(args)
if_batch_env = 1
if if_batch_env:
args.env = StockVecEnvNAS89(if_eval=False, gamma=args.gamma, env_num=2)
args.gpu_id = 0
args.worker_num = 2
train_and_evaluate_mp(args)
if_multi_learner = 0
if if_multi_learner:
args.env = StockVecEnvNAS89(if_eval=False, gamma=args.gamma, env_num=2)
args.gpu_id = (0, 1)
args.worker_num = 2
train_and_evaluate_mg(args)
"TotalStep: 52e5, TargetReturn: 2.35, UsedTime: 3934s, FinanceStock-v2"
"TotalStep: 81e5, TargetReturn: 2.47, UsedTime: 6129s, FinanceStock-v2"
"TotalStep: 19e5, TargetReturn: 2.50, UsedTime: 1654s, FinanceStock-v2 GPU 2, 3"
"TotalStep: 65e5, TargetReturn: 4.61, UsedTime: 5659s, FinanceStock-v2 GPU 2, 3"
"TotalStep: 18e5, TargetReturn: 2.50, UsedTime: 1452s, FinanceStock-v2 GPU 0, 1"
"TotalStep: 61e5, TargetReturn: 3.92, UsedTime: 4921s, FinanceStock-v2 GPU 0, 1"
"TotalStep: 4e5, TargetReturn: 2.20, UsedTime: 583s, FinanceStock-v2 GPU 0, 1, 2, 3"
"TotalStep: 11e6, TargetReturn: 4.39, UsedTime: 9648s, FinanceStock-v2 GPU 0, 1, 2, 3"
def demo_custom_env_finance_rl():
args = Arguments(
if_on_policy=True
) # hyper-parameters of on-policy is different from off-policy
args.random_seed = 0
from elegantrl2.agent import AgentPPO
args.agent = AgentPPO()
args.agent.cri_target = True
args.agent.lambda_entropy = 0.04
from envs.FinRL.StockTrading import StockEnvNAS89, StockVecEnvNAS89
args.gamma = 0.999
args.env = StockEnvNAS89(if_eval=False, gamma=args.gamma)
args.eval_env = StockEnvNAS89(if_eval=True, gamma=args.gamma)
args.repeat_times = 2**4
args.learning_rate = 2**-14
args.net_dim = int(2**8 * 1.5)
args.batch_size = args.net_dim * 4
if_single_env = 0
if if_single_env:
args.gpu_id = 0
args.worker_num = 4
train_and_evaluate_mp(args)
if_batch_env = 1
if if_batch_env:
args.env = StockVecEnvNAS89(if_eval=False, gamma=args.gamma, env_num=2)
args.gpu_id = 0
args.worker_num = 2
train_and_evaluate_mp(args)
if_multi_learner = 0
if if_multi_learner:
args.env = StockVecEnvNAS89(if_eval=False, gamma=args.gamma, env_num=2)
args.gpu_id = (0, 1)
args.worker_num = 2
train_and_evaluate_mg(args)
def demo_get_video_to_watch_gym_render():
import cv2 # pip3 install opencv-python
import gym # pip3 install gym==0.17 pyglet==1.5.0 # env.render() bug in gym==0.18, pyglet==1.6
import torch
"""parameters"""
env_name = 'LunarLanderContinuous-v2'
env = PreprocessEnv(env=gym.make(env_name))
'''initialize agent'''
agent = None # means use random action
if agent is None: # use random action
device = None
else:
from elegantrl2.agent import AgentPPO
agent = AgentPPO() # means use the policy network which saved in cwd
cwd = f'./{env_name}_{agent.__class__.__name__}/' # current working directory path
net_dim = 2 ** 9 # 2 ** 7
state_dim = env.state_dim
action_dim = env.action_dim
agent.init(net_dim, state_dim, action_dim)
agent.save_load_model(cwd=cwd, if_save=False)
device = agent.device
'''initialize evaluete and env.render()'''
save_frame_dir = '' # means don't save video, just open the env.render()
# save_frame_dir = 'frames' # means save video in this directory
if save_frame_dir:
os.makedirs(save_frame_dir, exist_ok=True)
state = env.reset()
episode_return = 0
step = 0
for i in range(2 ** 10):
print(i) if i % 128 == 0 else None
for j in range(1):
if agent is None:
action = env.action_space.sample()
else:
s_tensor = torch.as_tensor((state,), dtype=torch.float32, device=device)
a_tensor = agent.act(s_tensor)
action = a_tensor.detach().cpu().numpy()[0] # if use 'with torch.no_grad()', then '.detach()' not need.
next_state, reward, done, _ = env.step(action)
episode_return += reward
step += 1
if done:
print(f'{i:>6}, {step:6.0f}, {episode_return:8.3f}, {reward:8.3f}')
state = env.reset()
episode_return = 0
step = 0
else:
state = next_state
if save_frame_dir:
frame = env.render('rgb_array')
cv2.imwrite(f'{save_frame_dir}/{i:06}.png', frame)
cv2.imshow('OpenCV Window', frame)
cv2.waitKey(1)
else:
env.render()
env.close()
'''convert frames png/jpg to video mp4/avi using ffmpeg'''
if save_frame_dir:
frame_shape = cv2.imread(f'{save_frame_dir}/{3:06}.png').shape
print(f"frame_shape: {frame_shape}")
save_video = 'gym_render.mp4'
os.system(f"| Convert frames to video using ffmpeg. Save in {save_video}")
os.system(f'ffmpeg -r 60 -f image2 -s {frame_shape[0]}x{frame_shape[1]} '
f'-i ./{save_frame_dir}/%06d.png '
f'-crf 25 -vb 20M -pix_fmt yuv420p {save_video}')
def get_video_to_watch_gym_render():
import cv2 # pip3 install opencv-python
import gym # pip3 install gym==0.17 pyglet==1.5.0 # env.render() bug in gym==0.18, pyglet==1.6
import torch
'''choose env'''
import pybullet_envs # for python-bullet-gym
dir(pybullet_envs)
# from elegantrl2.env import PreprocessEnv
env_name = [
'BipedalWalker-v3', 'AntBulletEnv-v0', 'KukaBulletEnv-v0',
'ReacherBulletEnv-v0', 'PusherBulletEnv-v0', "ThrowerBulletEnv-v0",
"StrikerBulletEnv-v0"
][1]
env = PreprocessEnv(env=gym.make(env_name))
'''initialize agent'''
agent = None
from elegantrl2.agent import AgentPPO
agent = AgentPPO()
agent.if_use_dn = True
net_dim = 2**8
cwd = f'./{env_name}_4/'
# from elegantrl2.agent import AgentModSAC
# agent = AgentModSAC()
# agent.if_use_dn = True
# net_dim = 2 ** 8
# cwd = f'./{env_name}_2/'
device = None
if agent is not None:
state_dim = env.state_dim
action_dim = env.action_dim
agent.init(net_dim, state_dim, action_dim)
agent.save_load_model(cwd=cwd, if_save=False)
device = agent.device
rd.seed(194686)
torch.manual_seed(1942876)
'''initialize evaluete and env.render()'''
save_frame_dir = 'frames'
if save_frame_dir:
os.makedirs(save_frame_dir, exist_ok=True)
state = env.reset()
episode_return = 0
step = 0
for i in range(2**9):
print(i) if i % 128 == 0 else None
for j in range(1):
if agent is not None:
s_tensor = torch.as_tensor((state, ),
dtype=torch.float32,
device=device)
a_tensor = agent.act(s_tensor)
action = a_tensor.detach().cpu().numpy(
)[0] # if use 'with torch.no_grad()', then '.detach()' not need.
else:
action = env.action_space.sample()
next_state, reward, done, _ = env.step(action)
episode_return += reward
step += 1
if done:
print(
f'{i:>6}, {step:6.0f}, {episode_return:8.3f}, {reward:8.3f}'
)
state = env.reset()
episode_return = 0
step = 0
else:
state = next_state
frame = env.render('rgb_array')
frame = frame[50:210, 50:270] # (240, 320) AntPyBulletEnv-v0
# frame = cv2.resize(frame[:, :500], (500//2, 720//2))
cv2.imwrite(f'{save_frame_dir}/{i:06}.png', frame)
cv2.imshow('', frame)
cv2.waitKey(1)
env.close()
# exit()
'''convert frames png/jpg to video mp4/avi using ffmpeg'''
if save_frame_dir:
frame_shape = cv2.imread(f'{save_frame_dir}/{3:06}.png').shape
print(f"frame_shape: {frame_shape}")
save_video = 'gym_render.mp4'
os.system(
f"| Convert frames to video using ffmpeg. Save in {save_video}")
os.system(
f'ffmpeg -r 60 -f image2 -s {frame_shape[0]}x{frame_shape[1]} '
f'-i ./{save_frame_dir}/%06d.png '
f'-crf 25 -vb 20M -pix_fmt yuv420p {save_video}')
def check_stock_trading_env():
if_eval = True # False
env = StockTradingEnv(if_eval=if_eval)
action_dim = env.action_dim
state = env.reset()
print('| check_stock_trading_env, state_dim', len(state))
from time import time
timer = time()
# ============================================================
policy_name = 'Random Action 1e-2'
step = 1
done = False
episode_return = 0
# state = env.reset()
while not done:
action = rd.uniform(-1, 1, size=action_dim) * 1e-2
next_state, reward, done, _ = env.step(action)
# print(';', len(next_state), env.day, reward)
episode_return += reward
step += 1
print()
print(f"| {policy_name}:")
print(f"| step {step}, UsedTime {time() - timer:.3e}")
print(f"| gamma_reward \t\t\t{env.gamma_reward:.3e}")
print(f"| episode return \t\t{episode_return:.3e}")
print(
f"| discount return \t\t{episode_return / step / (1 - env.gamma):.3e}")
print(f"| env episode return \t{env.episode_return:.3e}")
# ============================================================
policy_name = 'Buy 4 Action'
step = 1
done = False
episode_return = 0
# state = env.reset()
while not done:
action = np.zeros(action_dim)
action[:3] = 1
next_state, reward, done, _ = env.step(action)
# print(';', len(next_state), env.day, reward)
episode_return += reward
step += 1
print()
print(f"| {policy_name}:")
print(f"| step {step}, UsedTime {time() - timer:.3e}")
print(f"| gamma_reward \t\t\t{env.gamma_reward:.3e}")
print(f"| episode return \t\t{episode_return:.3e}")
print(
f"| discount return \t\t{episode_return / step / (1 - env.gamma):.3e}")
print(f"| env episode return \t{env.episode_return:.3e}")
# ============================================================
'''draw_cumulative_return'''
from elegantrl2.agent import AgentPPO
from elegantrl2.run import Arguments
args = Arguments(if_on_policy=True)
args.agent = AgentPPO()
args.env = StockTradingEnv(if_eval=True)
args.if_remove = False
args.cwd = './StockTradingEnv-v1_AgentPPO'
args.init_before_training()
env.draw_cumulative_return(args, torch)
def demo_custom_env_finance_rl_nas89(): # 1.7+ 2.0+
args = Arguments(
if_on_policy=True
) # hyper-parameters of on-policy is different from off-policy
args.random_seed = 19430
from elegantrl2.agent import AgentPPO
args.agent = AgentPPO()
args.agent.lambda_entropy = 0.02
from envs.FinRL.StockTrading import StockEnvNAS89
args.gamma = 0.999
args.env = StockEnvNAS89(if_eval=False,
gamma=args.gamma,
turbulence_thresh=30)
args.eval_env = StockEnvNAS89(if_eval=True,
gamma=args.gamma,
turbulence_thresh=15)
args.net_dim = 2**9
args.repeat_times = 2**4
args.learning_rate = 2**-14
args.batch_size = args.net_dim * 4
args.eval_gap = 2**8
args.eval_times1 = 2**0
args.eval_times2 = 2**1
args.break_step = int(8e6)
args.if_allow_break = False
if_single_proc = 0
if if_single_proc:
args.gpu_id = int(sys.argv[-1][-4])
args.random_seed += int(args.gpu_id)
args.target_step = args.env.max_step * 4
train_and_evaluate(args)
if_single_env = 1
if if_single_env:
args.gpu_id = int(sys.argv[-1][-4])
args.random_seed += int(args.gpu_id)
args.target_step = args.env.max_step * 1
args.worker_num = 4
train_and_evaluate_mp(args)
if_multi_learner = 0
if if_multi_learner:
args.gpu_id = (2, 3) if len(sys.argv) == 1 else eval(
sys.argv[-1]) # python main.py -GPU 0,1
args.repeat_times = 2**4
args.target_step = args.env.max_step
args.worker_num = 4
train_and_evaluate_mg(args)
if_batch_env = 0
if if_batch_env:
from envs.FinRL.StockTrading import StockVecEnvNAS89
args.env = StockVecEnvNAS89(if_eval=False, gamma=args.gamma, env_num=2)
args.gpu_id = int(sys.argv[-1][-4])
args.random_seed += args.gpu_id
args.target_step = args.env.max_step
args.worker_num = 4
train_and_evaluate_mp(args)