def make_env_all_params(rank, add_monitor, args):
if args["env_kind"] == 'atari':
env = gym.make(args['env'])
assert 'NoFrameskip' in env.spec.id
env = NoopResetEnv(env, noop_max=args['noop_max'])
env = MaxAndSkipEnv(env, skip=4)
env = ProcessFrame84(env, crop=False)
env = FrameStack(env, 4)
env = ExtraTimeLimit(env, args['max_episode_steps'])
if 'Montezuma' in args['env']:
env = MontezumaInfoWrapper(env)
env = AddRandomStateToInfo(env)
elif args["env_kind"] == 'mario':
env = make_mario_env()
elif args["env_kind"] == "retro_multi":
env = make_multi_pong()
elif args["env_kind"] == 'robopong':
if args["env"] == "pong":
env = make_robo_pong()
elif args["env"] == "hockey":
env = make_robo_hockey()
elif args["env_kind"] == "my_games":
env = gym.make(args['env'])
env = MaxAndSkipEnv(env, skip=4)
env = WarpFrame(env)
env = FrameStack(env, 4)
if add_monitor:
env = Monitor(env, osp.join(logger.get_dir(), '%.2i' % rank))
return env
def make_env_all_params(rank, add_monitor, args):
if args["env_kind"] == 'atari':
env = gym.make(args['env'])
assert 'NoFrameskip' in env.spec.id
if args["stickyAtari"]: # 在智能体执行动作时增加随机性
env._max_episode_steps = args['max_episode_steps'] * 4
env = StickyActionEnv(env)
else:
env = NoopResetEnv(env, noop_max=args['noop_max'])
env = MaxAndSkipEnv(env, skip=4) # 每个动作连续执行4步
env = ProcessFrame84(env, crop=False) # 处理观测
env = FrameStack(env, 4) # 将连续4帧叠加起来作为输入
env = ExtraTimeLimit(env, args['max_episode_steps'])
if not args["stickyAtari"]:
env = ExtraTimeLimit(env,
args['max_episode_steps']) # 限制了一个周期的最大时间步
if 'Montezuma' in args['env']: # 记录智能体的位置, 所在的房间, 已经访问的房间
env = MontezumaInfoWrapper(env)
env = AddRandomStateToInfo(env)
elif args["env_kind"] == 'mario': # 超级马里奥
env = make_mario_env()
elif args["env_kind"] == "retro_multi": # 多智能体游戏, Multi-Pong
env = make_multi_pong()
elif args["env_kind"] == 'robopong':
if args["env"] == "pong":
env = make_robo_pong()
elif args["env"] == "hockey":
env = make_robo_hockey()
if add_monitor:
env = Monitor(env, osp.join(logger.get_dir(), '%.2i' % rank))
return env
def make_env_all_params(rank, add_monitor, args):
if args["env_kind"] == "atari":
env = gym.make(args["env"])
assert "NoFrameskip" in env.spec.id
# from self-supervised exploration via disagreement
if args["stickyAtari"] == "true":
env = StickyActionEnv(env)
env._max_episode_steps = args["max_episode_steps"] * 4
env = MaxAndSkipEnv(env, skip=4)
env = ProcessFrame84(env, crop=False)
env = FrameStack(env, 4)
env = ExtraTimeLimit(env, args["max_episode_steps"])
if "Montezuma" in args["env"]:
env = MontezumaInfoWrapper(env)
env = AddRandomStateToInfo(env)
if args["noisy_tv"] == "true":
env = NoisyTVEnvWrapper(env)
# assert env.action_space == spaces.Discrete(7)
elif args["env_kind"] == "mario":
env = make_mario_env()
if args["noisy_tv"] == "true":
env = NoisyTVEnvWrapperMario(env)
elif args["env_kind"] == "retro_multi":
env = make_multi_pong()
elif args["env_kind"] == "robopong":
if args["env"] == "pong":
env = make_robo_pong()
elif args["env"] == "hockey":
env = make_robo_hockey()
if add_monitor:
env = Monitor(env, osp.join(logger.get_dir(), "%.2i" % rank))
return env
def make_env_all_params(rank, add_monitor, args, logdir):
if args["env_kind"] == 'atari':
env = gym.make(args['env'])
assert 'NoFrameskip' in env.spec.id
env = NoopResetEnv(env, noop_max=args['noop_max'])
env = MaxAndSkipEnv(env, skip=4)
env = ProcessFrame84(env, crop=False)
env = FrameStack(env, 4)
env = ExtraTimeLimit(env, args['max_episode_steps'])
if 'Montezuma' in args['env']:
env = MontezumaInfoWrapper(env)
env = AddRandomStateToInfo(env)
elif args["env_kind"] == 'mario':
env = make_mario_env()
elif args["env_kind"] == "retro_multi":
env = make_multi_pong()
elif args["env_kind"] == 'robopong':
if args["env"] == "pong":
env = make_robo_pong()
elif args["env"] == "hockey":
env = make_robo_hockey()
elif args["env_kind"] == "dm_suite":
env = make_dm_suite(task=args["env"],
logdir=logdir,
to_record=args["to_record"])
if add_monitor:
env = TempMonitor(env)
return env
def make_env_all_params(rank, add_monitor, args, sleep_multiple=2):
if args["env_kind"] == 'ObstacleTowerEnv':
env = _make_obs_env(rank, add_monitor, args, sleep_multiple)
elif args["env_kind"] == 'atari':
env = gym.make(args['env'])
assert 'NoFrameskip' in env.spec.id
env = NoopResetEnv(env, noop_max=args['noop_max'])
env = MaxAndSkipEnv(env, skip=4)
env = ProcessFrame84(env, crop=False)
env = FrameStack(env, 4)
env = ExtraTimeLimit(env, args['max_episode_steps'])
if 'Montezuma' in args['env']:
env = MontezumaInfoWrapper(env)
env = AddRandomStateToInfo(env)
if rank == 2:
env = RenderWrapper(env)
elif args["env_kind"] == 'mario':
env = make_mario_env()
elif args["env_kind"] == "retro_multi":
env = make_multi_pong()
elif args["env_kind"] == 'robopong':
if args["env"] == "pong":
env = make_robo_pong()
elif args["env"] == "hockey":
env = make_robo_hockey()
if add_monitor:
logdir = osp.join('summaries', args["exp_name"])
logger.configure(logdir)
env = Monitor(env, osp.join(logger.get_dir(), '%.2i' % rank))
return env
def make_env_all_params(rank, add_monitor, args):
if args["env_kind"] == 'atari':
env = gym.make(args['env'])
assert 'NoFrameskip' in env.spec.id
env = NoopResetEnv(env, noop_max=args['noop_max'])
env = MaxAndSkipEnv(env, skip=4)
env = ProcessFrame84(env, crop=False)
env = FrameStack(env, 4)
env = ExtraTimeLimit(env, args['max_episode_steps'])
if 'Montezuma' in args['env']:
env = MontezumaInfoWrapper(env)
env = AddRandomStateToInfo(env)
elif args["env_kind"] == 'mario':
env = make_mario_env()
elif args["env_kind"] == "retro_multi":
env = make_multi_pong()
elif args["env_kind"] == 'robopong':
if args["env"] == "pong":
env = make_robo_pong()
elif args["env"] == "hockey":
env = make_robo_hockey()
if args["env_kind"] == 'atari' and add_monitor:
#env = Monitor(env, osp.join(logger.get_dir(), '%.2i' % rank))
env = Monitor(env,
os.path.join(os.getcwd(), 'test_video'),
force=True,
video_callable=lambda episode_id: episode_id % 20 == 0)
#env = Monitor(env, os.path.join(os.getcwd(), 'test_video'),video_callable=lambda episode_id: True )#,force=True)
return env
def make_env_all_params(rank, add_monitor, args):
if args["env_kind"] == 'atari':
env = make_atari_env(args)
elif args["env_kind"] == 'mario':
env = make_mario_env()
elif args["env_kind"] == "retro_multi":
env = make_multi_pong()
elif args["env_kind"] == 'robopong':
if args["env"] == "pong":
env = make_robo_pong()
elif args["env"] == "hockey":
env = make_robo_hockey()
if add_monitor:
env = Monitor(env, osp.join(logger.get_dir(), '%.2i' % rank))
return env
def make_env_all_params(rank, add_monitor, args):
if args["env_kind"] == 'atari':
env = gym.make(args['env'])
assert 'NoFrameskip' in env.spec.id
env = NoopResetEnv(env, noop_max=args['noop_max'])
env = MaxAndSkipEnv(env, skip=4)
env = ProcessFrame84(env, crop=False)
env = FrameStack(env, 4)
env = ExtraTimeLimit(env, args['max_episode_steps'])
if 'Montezuma' in args['env']:
env = MontezumaInfoWrapper(env)
env = AddRandomStateToInfo(env)
elif args["env_kind"] == 'mario':
env = make_mario_env()
elif args["env_kind"] == "retro_multi":
env = make_multi_pong()
elif args["env_kind"] == 'robopong':
if args["env"] == "pong":
env = make_robo_pong()
elif args["env"] == "hockey":
env = make_robo_hockey()
if add_monitor:
#print(osp.join(logger.get_dir(), '%.2i' % rank + '.monitor.csv'))
env = Monitor(env, osp.join(logger.get_dir(), '%.2i' % rank))
"""
env = DummyVecEnv([lambda: env])
env = VecVideoRecorder(env, directory = './vid',
record_video_trigger=lambda step: step == 0,
video_length= 100,)
env.reset()
"""
#env = wrappers.Monitor(env,'./vid/',force = True,write_upon_reset = True, video_callable=lambda episode: True)
#print(osp.join(logger.get_dir()))
#env = Monitor(env, osp.join(logger.get_dir()))
#env = Monitor(env, "./vid", video_callable=lambda episode_id: True,force=True)
return env
def make_env_all_params(rank, add_monitor, args):
if args["env_kind"] == 'atari':
env = gym.make(args['env'])
assert 'NoFrameskip' in env.spec.id
if args["stickyAtari"]:
env._max_episode_steps = args['max_episode_steps'] * 4
env = StickyActionEnv(env)
else:
env = NoopResetEnv(env, noop_max=args['noop_max'])
env = MaxAndSkipEnv(env, skip=4)
env = ProcessFrame84(env, crop=False)
env = FrameStack(env, 4)
if not args["stickyAtari"]:
env = ExtraTimeLimit(env, args['max_episode_steps'])
if 'Montezuma' in args['env']:
env = MontezumaInfoWrapper(env)
env = AddRandomStateToInfo(env)
elif args["env_kind"] == 'mario':
env = make_mario_env()
elif args["env_kind"] == "retro_multi":
env = make_multi_pong()
elif args["env_kind"] == 'unity':
env = make_unity_maze(args["env"],
seed=args["seed"],
rank=rank,
ext_coeff=args["ext_coeff"],
recordUnityVid=args['recordUnityVid'],
expID=args["unityExpID"],
startLoc=args["startLoc"],
door=args["door"],
tv=args["tv"],
testenv=args["testenv"],
logdir=logger.get_dir())
if add_monitor:
env = Monitor(env, osp.join(logger.get_dir(), '%.2i' % rank))
return env
from wrappers import make_mario_env
import numpy as np
import warnings
import tensorflow as tf
from stable_baselines.common.vec_env import SubprocVecEnv
warnings.filterwarnings('ignore')
if __name__ == '__main__':
env_id = 'SuperMarioBros-1-1-v0'
nenvs = 8
frameskip = 4
env = SubprocVecEnv(
env_fns=[make_mario_env(env_id, frameskip, i) for i in range(nenvs)],
start_method='spawn')
action_n = env.action_space.n
print('observation_shape: {}'.format(env.observation_space.shape))
ph_obs = tf.to_float(
tf.placeholder(dtype=env.observation_space.dtype,
shape=(None, ) + env.observation_space.shape,
name='ph_obs'))
out = tf.layers.conv2d(ph_obs,
filters=5,
kernel_size=10,
strides=10,
activation=tf.nn.relu,
kernel_initializer=tf.initializers.orthogonal(1.0),
