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"] == 'field':
import gym_fieldedmove
env = gym.make('FieldedMove-v0')
# env = FrameStack(env, 4)
elif args["env_kind"] == "ple":
import gym_ple
env = gym.make(args['env'])
env._max_episode_steps = args['max_episode_steps']
# env = MaxAndSkipEnv(env, skip=4)
env = ProcessFrame84(env, crop=False)
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
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:
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 atari_wrap(env, max_episode_steps=None):
env = NoopResetEnv(env, noop_max=30)
env = MaxAndSkipEnv(env, skip=4)
if max_episode_steps is not None:
env = TimeLimit(env, max_episode_steps=max_episode_steps)
env = TimeLimitMask(env)
return env
def custom_make_atari(env_id, skip_frames=1):
from baselines.common.atari_wrappers import NoopResetEnv
env = gym.make(env_id)
assert 'NoFrameskip' in env.spec.id
env = NoopResetEnv(env, noop_max=30)
env = SkipFramesEnv(env, skip=skip_frames)
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(env_name='PongNoFrameskip-v4', size=42, skip=4, is_train=True):
env = gym.make(env_name)
env = NoopResetEnv(env, noop_max=300)
if is_train:
env = MaxAndSkipEnv(env, skip=skip)
env = WarpFrame(env, width=size, height=size, grayscale=True) # obs_space is now (84,84,1)
env = ScaledFloatFrame(env)
env = ChannelFirstFrameStack(env, 4)
return env
def _wrap_deepmind_ram(env):
"""Applies various Atari-specific wrappers to make learning easier."""
env = EpisodicLifeEnv(env)
env = NoopResetEnv(env, noop_max=30)
env = MaxAndSkipEnv(env, skip=4)
if 'FIRE' in env.unwrapped.get_action_meanings():
env = FireResetEnv(env)
env = ClipRewardEnv(env)
return env
def make_atari(env_id):
spec = gym_registry.spec(env_id)
# not actually needed, but we feel safer
spec.max_episode_steps = None
spec.max_episode_time = None
env = spec.make()
assert 'NoFrameskip' in env.spec.id
env = NoopResetEnv(env, noop_max=30)
env = MaxAndSkipEnv(env, skip=frame_skip)
return env
def wrap_env(env, episode_life=False):
if episode_life:
env = EpisodicLifeEnv(env)
env = NoopResetEnv(env, 30)
env = MaxAndSkipEnv(env, 4)
if env.unwrapped.get_action_meanings()[1] == 'FIRE':
env = FireResetEnv(env)
env = WarpFrame(env) # , width=84, height=84)
env = FrameStack(env, 4)
env = ScaledFloatFrame(env)
return env
def _thunk():
env = gym.make(**env_base)
env = NoopResetEnv(env, noop_max=30)
env = MaxAndSkipEnv(env, skip=4)
env = RewardCollector(env)
env = EpisodicLifeEnv(env)
env = ClipRewardEnv(env)
env = WarpFrame(env)
env = ScaledFloatFrame(env)
env = TransposeImage(env)
env = UnrealEnvBaseWrapper(env)
return env
def create_env(self, env):
env = gym.make(env)
env = NoopResetEnv(env, noop_max=30)
env = MaxAndSkipEnv(env, skip=4)
env = RewardCollector(env)
env = EpisodicLifeEnv(env)
env = ClipRewardEnv(env)
env = WarpFrame(env)
env = FrameStack(env, 4)
env = ConvertToNumpy(env)
env = TransposeImage(env)
env = ScaledFloatFrame(env)
return env
def make_env_all_params(rank, args):
"""Initialize the environment and apply wrappers.
Parameters
----------
rank :
Rank of the environment.
args :
Hyperparameters for this run.
Returns
-------
env
Environment with its individual wrappers.
"""
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"] == "roboarm":
from real_robots.envs import REALRobotEnv
env = REALRobotEnv(objects=3, action_type="cartesian")
env = CartesianControlDiscrete(
env,
crop_obs=args["crop_obs"],
repeat=args["act_repeat"],
touch_reward=args["touch_reward"],
random_force=args["random_force"],
)
print("adding monitor")
env = Monitor(env, filename=None)
return env
def make_env_all_params(rank, args):
env = gym.make(GAME_NAME)
env = NoopResetEnv(env, noop_max=NOOP_MAX)
env = MaxAndSkipEnv(env, skip=4)
env = ProcessFrame84(env, crop=False)
env = FrameStack(env, 4)
# env = ExtraTimeLimit(env,10000)
env = AddRandomStateToInfo(env)
env = Monitor(
env,
os.path.join(
'C:/Users/Elias/OneDrive/Winfo Studium/SS19/Masterarbeit/logs',
'%.2i' % rank))
return env
def make_atari_env(args, frame_stack=True):
import gym
from baselines.common.atari_wrappers import FrameStack, NoopResetEnv, FrameStack
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)
return env
def make_env(env_name='PongNoFrameskip-v4',
size=84,
skip=4,
scale=True,
is_train=True):
env = gym.make(env_name)
env = NoopResetEnv(env, noop_max=30)
if is_train:
env = MaxAndSkipEnv(env, skip=skip)
if env.unwrapped.ale.lives() > 0:
env = EpisodicLifeEnv(env)
if 'FIRE' in env.unwrapped.get_action_meanings():
env = FireResetEnv(env)
env = WarpFrame(env, width=size, height=size,
grayscale=True) # obs_space is now (84,84,1)
if scale:
env = ScaledFloatFrame(env)
env = ChannelFirstFrameStack(env, 4)
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 _thunk():
episodic_life = True
env = gym.make(env_id)
env.seed(seed + rank)
env.frameskip = 1
if log_dir is not None:
env = bench.Monitor(env, os.path.join(log_dir, str(rank)))
if episodic_life:
env = EpisodicLifeEnv(env)
env = NoopResetEnv(env, noop_max=30)
env = MaxAndSkipEnv(env, skip=1)
if 'FIRE' in env.unwrapped.get_action_meanings():
env = FireResetEnv(env)
if grey_scale:
env = WarpMiniPacmanFrameGreyScale(env)
else:
env = WarpMiniPacmanFrameRGB(env)
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
def _make_env(env_name,
render=False,
max_episode_steps=18000,
frame_shape=(1, 84, 84),
inter_area=False,
noop_max=30):
env = gym.make(f'{env_name}NoFrameskip-v4')
assert 'NoFrameskip' in env.spec.id
env = NoopResetEnv(env, noop_max=noop_max)
env = SkipEnv(env, skip=4)
env = TimeLimit(env, max_episode_steps=max_episode_steps)
if 'FIRE' in env.unwrapped.get_action_meanings():
env = atari_wrappers.FireResetEnv(env)
grayscale = frame_shape[0] == 1
height, width = frame_shape[1:]
env = WarpFrame(env,
width=width,
height=height,
grayscale=grayscale,
inter_area=inter_area)
env = ClipRewardEnv(env)
if render:
env = RenderingEnv(env)
return env
def atari_setup(env):
# from baselines.common.atari_wrappers
assert 'NoFrameskip' in env.spec.id
env = NoopResetEnv(env, noop_max=30)
env = MaxAndSkipEnv(env, skip=4)
return env
def _thunk(env):
assert 'NoFrameskip' in env.spec.id
return NoopResetEnv(env, noop_max=noop_max)
def _thunk():
env = gym.make(env_id)
env.seed(seed + rank)
env = NoopResetEnv(env, noop_max=30)
env = MaxAndSkipEnv(env, skip=2)
env = WarpFrame(env)
# Janky Fix to Resize Environments to be 50x50
env.width = 50
env.height = 50
env = ScaledFloatFrame(env)
if not eval:
env = ClipRewardEnv(env)
env = EpisodicLifeEnv(env)
env = FrameStack(env, 3)
env = TransposeOb(env)
return env
def make_atari(env_id):
env = gym.make(env_id)
# assert 'NoFrameskip' in env.spec.id
env = NoopResetEnv(env, noop_max=30)
# env = MaxAndSkipEnv(env, skip=4)
return env
