def _thunk():
if callable(env_id):
env = env_id()
elif env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
obs_shape = env.observation_space.shape
if add_timestep and len(
obs_shape) == 1 and str(env).find('TimeLimit') > -1:
env = AddTimestep(env)
if log_dir is not None:
env = bench.Monitor(env,
os.path.join(log_dir, str(rank)),
allow_early_resets=allow_early_resets)
if is_atari:
if len(env.observation_space.shape) == 3:
env = wrap_deepmind(env)
return env
def _thunk():
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
raise 'Atari not supported'
env.seed(seed + rank)
obs_shape = env.observation_space.shape
if str(env.__class__.__name__).find('TimeLimit') >= 0:
env = TimeLimitMask(env)
if log_dir is not None:
env = bench.Monitor(
env,
os.path.join(log_dir, str(rank)),
allow_early_resets=allow_early_resets)
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env, op=[2, 0, 1])
return env
def _thunk():
random_seed(seed)
if 'diabcombolockhallway' == env_id:
env = DiabolicalLockMaze(horizon=horizon, seed=seed, noise=noise)
elif 'maze' == env_id:
env = build_env_maze(horizon=100, size=20)
elif env_id.startswith("dm"):
import dm_control2gym
_, domain, task = env_id.split('-')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
if not 'combolock' in env_id:
env.seed(seed + rank)
env = OriginalReturnWrapper(env) #aggregate episode return
if is_atari:
env = wrap_deepmind(env,
episode_life=episode_life,
clip_rewards=False,
frame_stack=False,
scale=False)
obs_shape = env.observation_space.shape
if len(obs_shape) == 3:
env = TransposeImage(env)
env = FrameStack(env, 4)
return env
def _thunk():
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
elif env_id in ['duckietown']:
from a2c_ppo_acktr.duckietown.env import launch_env
from a2c_ppo_acktr.duckietown.wrappers import NormalizeWrapper, ImgWrapper,\
DtRewardWrapper, ActionWrapper, ResizeWrapper
from a2c_ppo_acktr.duckietown.teacher import PurePursuitExpert
env = launch_env()
env = ResizeWrapper(env)
env = NormalizeWrapper(env)
env = ImgWrapper(env)
env = ActionWrapper(env)
env = DtRewardWrapper(env)
elif env_id in retro_envs:
env = make_retro(game=env_id)
#env = SuperMarioKartDiscretizer(env)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id, max_episode_steps=max_steps)
env.seed(seed + rank)
#TODO: Figure out what todo here
if is_atari:
env = TimeLimitMask(env)
if log_dir is not None:
env = bench.Monitor(env,
os.path.join(log_dir, str(rank)),
allow_early_resets=allow_early_resets)
if is_atari:
if len(env.observation_space.shape) == 3:
env = wrap_deepmind(env)
elif env_id in retro_envs:
if len(env.observation_space.shape) == 3:
env = wrap_deepmind_retro(env, frame_stack=0)
elif len(env.observation_space.shape) == 3:
if env_id not in ['duckietown'] and env_id not in retro_envs:
raise NotImplementedError(
"CNN models work only for atari,\n"
"please use a custom wrapper for a custom pixel input env.\n"
"See wrap_deepmind for an example.")
# If the input has shape (W,H,3), wrap for PyTorch convolutions
if env_id not in ['duckietown']:
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env, op=[2, 0, 1])
if time:
env = TimeFeatureWrapper(env)
return env
def _thunk():
if "mario" in env_id:
env = gym_super_mario_bros.make('SuperMarioBros-v0')
env = JoypadSpace(env, SIMPLE_MOVEMENT)
elif env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
if log_dir is not None:
env = bench.Monitor(env, os.path.join(log_dir, str(rank)))
if is_atari:
env = wrap_deepmind(env, clip_rewards=False)
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if "mario" in env_id:
env = wrap_deepmind(env, clip_rewards=False, episode_life=False)
env = WrapPyTorchMario(env)
elif len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = WrapPyTorch(env)
return env
def _thunk():
if env_id.startswith("Pomme"):
env = envs.pommerman.make_env(env_id)
elif env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
obs_shape = env.observation_space.shape
if add_timestep and len(
obs_shape) == 1 and str(env).find('TimeLimit') > -1:
env = AddTimestep(env)
if log_dir is not None:
env = Monitor(env,
os.path.join(log_dir, str(rank)),
allow_early_resets=allow_early_resets)
if is_atari:
env = wrap_deepmind(env)
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env)
return env
def _thunk():
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
if str(env.__class__.__name__).find('TimeLimit') >= 0:
env = TimeLimitMask(env)
if log_dir is not None:
env = env
if is_atari:
if len(env.observation_space.shape) == 3:
env = wrap_deepmind(env)
elif len(env.observation_space.shape) == 3:
raise NotImplementedError(
"CNN models work only for atari,\n"
"please use a custom wrapper for a custom pixel input env.\n"
"See wrap_deepmind for an example.")
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env, op=[2, 0, 1])
return env
def _thunk():
# I think this is not needed
# random_seed(seed)
if env_id.startswith("dm"):
import dm_control2gym
_, domain, task = env_id.split('-')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
# My code for reacher env:
if env_id == 'reacher':
# from gym_unity.envs.unity_env import UnityEnv
# from p2_continuous_control.unity_env import UnityEnv
# env = UnityEnv('p2_continuous_control/Reacher_Linux/Reacher.x86_64')
env = make_reacher()
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
env = OriginalReturnWrapper(env)
if is_atari:
env = wrap_deepmind(env,
episode_life=episode_life,
clip_rewards=False,
frame_stack=False,
scale=False)
obs_shape = env.observation_space.shape
if len(obs_shape) == 3:
env = TransposeImage(env)
env = FrameStack(env, 4)
return env
def _thunk():
if env_id == 'StockGame':
env = StockGame('603399', '2017-01-01', '2017-07-10', 100000.0)
elif env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
obs_shape = env.observation_space.shape
if add_timestep and len(obs_shape) == 1 \
and str(env).find('TimeLimit') > -1:
env = AddTimestep(env)
if log_dir is not None:
env = bench.Monitor(env, os.path.join(log_dir, str(rank)))
if is_atari:
env = wrap_deepmind(env)
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = WrapPyTorch(env)
return env
def _thunk():
random_seed(seed)
if env_id.startswith("dm"):
import dm_control2gym
_, domain, task = env_id.split('-')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
if log_dir is not None:
# env = Monitor(env=env, filename=os.path.join(log_dir, str(rank)), allow_early_resets=True)
env = bench.Monitor(env=env,
filename=os.path.join(log_dir, str(rank)),
allow_early_resets=True)
if is_atari:
env = wrap_deepmind(env,
episode_life=episode_life,
clip_rewards=False,
frame_stack=False,
scale=False)
if not episode_life and 'FIRE' in env.unwrapped.get_action_meanings(
):
env = LifeFireEnv(env)
obs_shape = env.observation_space.shape
if len(obs_shape) == 3:
env = TransposeImage(env)
env = FrameStack(env, stack_size)
return env
def _thunk():
if callable(env_id):
env = env_id()
elif env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = gym.make(env_id)
env = gym.wrappers.AtariPreprocessing(env)
env.seed(seed + rank)
obs_shape = env.observation_space.shape
if add_timestep and len(
obs_shape) == 1 and str(env).find('TimeLimit') > -1:
env = AddTimestep(env)
if log_dir is not None:
# env = gym.wrappers.Monitor(env, os.path.join(log_dir, str(rank)), force=True)
pass
if is_atari:
if len(env.observation_space.shape) == 3:
# TODO: implement same preprocessing as in deepmind
# env = wrap_deepmind(env)
pass
return env
def _thunk():
random_seed(seed)
if env_id.startswith("dm"):
import dm_control2gym
_, domain, task = env_id.split('-')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
env = OriginalReturnWrapper(env)
if is_atari:
env = wrap_deepmind(env,
episode_life=episode_life,
clip_rewards=False,
frame_stack=False,
scale=False)
obs_shape = env.observation_space.shape
if len(obs_shape) == 3:
env = TransposeImage(env)
env = FrameStack(env, 4)
return env
def _thunk():
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
elif 'Mini' in env_id:
import gym_minigrid
env = gym_minigrid.envs.dynamicobstacles.DynamicObstaclesEnv(
size=5, n_obstacles=1)
# env = gym_minigrid.envs.multiroom.MultiRoomEnv(
# minNumRooms=2, maxNumRooms=2, maxRoomSize=4
# )
# import pdb; pdb.set_trace()
# env = gym.make(env_id)
# env = gym_minigrid.wrappers.NoOpAsync(env, costs=[2, 1])
env = gym_minigrid.wrappers.NoOpAsync(env,
costs=[2, 1],
which_obs='first')
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
is_minigrid = 'minigrid' in env_id.lower()
env.seed(seed + rank)
if is_minigrid:
from gym_minigrid.wrappers import ImgObsWrapper, RGBImgObsWrapper, RGBImgPartialObsWrapper, FlatObsWrapper
# env = RGBImgPartialObsWrapper(
# env, tile_size=2)
# env = ImgObsWrapper(env)
env = FlatObsWrapper(env)
# env.observation_space = env.observation_space['image']
if str(env.__class__.__name__).find('TimeLimit') >= 0:
env = TimeLimitMask(env)
if log_dir is not None:
env = bench.Monitor(env,
os.path.join(log_dir, str(rank)),
allow_early_resets=allow_early_resets)
if is_atari:
if len(env.observation_space.shape) == 3:
env = wrap_deepmind(env)
elif is_minigrid:
pass
elif len(env.observation_space.shape) == 3:
raise NotImplementedError(
"CNN models work only for atari,\n"
"please use a custom wrapper for a custom pixel input env.\n"
"See wrap_deepmind for an example.")
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env, op=[2, 0, 1])
return env
def make_dm2gym_env_obs(env_name="cheetah_run", num_repeat_action=1):
""" Invoke POMDP tasks """
env_name_tuple = DM_TASKS[env_name.lower()]
domain_name, task_name = env_name_tuple[0], env_name_tuple[1]
env = dm_control2gym.make(domain_name=domain_name, task_name=task_name)
env = PixelObservationWrapper(env=env)
env = wrap_deepmind(env=env, num_repeat_action=num_repeat_action)
return env
def _thunk():
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
# ss('sss')
env = env.env
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
# print(hasattr(gym.envs, 'atari'))
# print(gym.envs)
# print(getattr(gym.envs, 'atari'))
# print('- - '*20)
# print(env.unwrapped)
# print(gym.envs.atari.atari_env.AtariEnv)
# # print(gym.envs.atari.atari_env.AtariEnv)
# print(isinstance(
# env.unwrapped, gym.envs.atari.atari_env.AtariEnv))
# print(is_atari)
# ss('sss')
# if is_atari:
#
# env = make_atari(env_id)
# ss('make atari')
env.seed(seed + rank)
# obs_shape = env.observation_space.shape
# print(str(env.__class__.__name__))
# ss('name')
if str(env.__class__.__name__).find('TimeLimit') >= 0:
# print('hiosdfoi '* 30)
env = TimeLimitMask(env)
# if log_dir is not None:
# env = bench.Monitor(
# env,
# os.path.join(log_dir, str(rank)),
# allow_early_resets=allow_early_resets)
# if is_atari:
# if len(env.observation_space.shape) == 3:
# env = wrap_deepmind(env)
# elif len(env.observation_space.shape) == 3:
# raise NotImplementedError(
# "CNN models work only for atari,\n"
# "please use a custom wrapper for a custom pixel input env.\n"
# "See wrap_deepmind for an example.")
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
# print(obs_shape)
# ss('is atari')
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env, op=[2, 0, 1])
return env
def _thunk():
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
elif 'academy' in env_id or '11' in env_id: # gfootball environments
#env = football_env.create_environment(
# env_name=env_id, stacked=('stacked' in state),
# with_checkpoints=('with_checkpoints' in reward_experiment),
# logdir=log_dir,
# enable_goal_videos=dump_scores and (seed == 0),
# enable_full_episode_videos=dump_full_episodes and (seed == 0),
# render=render and (seed == 0),
# dump_frequency=50 if render and seed == 0 else 0,
# representation='extracted')
env = football_env.create_environment(
env_name=env_id,
stacked=('stacked' in state),
rewards=reward_experiment,
logdir=log_dir,
render=render and (seed == 0),
dump_frequency=50 if render and seed == 0 else 0)
env = EpisodeRewardScoreWrapper(
env,
number_of_left_players_agent_controls=1,
number_of_right_players_agent_controls=0)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
obs_shape = env.observation_space.shape
if str(env.__class__.__name__).find('TimeLimit') >= 0:
env = TimeLimitMask(env)
if is_atari:
if len(env.observation_space.shape) == 3:
env = wrap_deepmind(env, frame_stack=True)
elif len(env.observation_space.shape
) == 3 and 'academy' not in env_id and '11' not in env_id:
raise NotImplementedError(
"CNN models work only for atari,\n"
"please use a custom wrapper for a custom pixel input env.\n"
"See wrap_deepmind for an example.")
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env, op=[2, 0, 1])
return env
def _thunk():
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
if obs_keys is not None:
env = gym.wrappers.FlattenDictWrapper(env, dict_keys=obs_keys)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
obs_shape = env.observation_space.shape
# if str(env.__class__.__name__).find('TimeLimit') >= 0:
# env = TimeLimitMask(env)
if log_dir is not None:
if save_video:
env = bench.Monitor(env,
os.path.join(log_dir + '/eval/monitor',
str(rank)),
allow_early_resets=allow_early_resets)
env = gym.wrappers.Monitor(env,
os.path.join(
log_dir + '/eval/video',
str(rank)),
force=True)
else:
env = bench.Monitor(env,
os.path.join(log_dir + '/monitor',
str(rank)),
allow_early_resets=allow_early_resets)
if is_atari:
if len(env.observation_space.shape) == 3:
env = wrap_deepmind(env)
elif len(env.observation_space.shape) == 3:
raise NotImplementedError(
"CNN models work only for atari,\n"
"please use a custom wrapper for a custom pixel input env.\n"
"See wrap_deepmind for an example.")
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env, op=[2, 0, 1])
return env
def __init__(self, domain_name, task_name, log_dir=None):
from dm_control import suite
import dm_control2gym
BaseTask.__init__(self)
self.name = domain_name + '_' + task_name
self.env = dm_control2gym.make(domain_name, task_name)
self.action_dim = self.env.action_space.shape[0]
self.state_dim = self.env.observation_space.shape[0]
self.env = self.set_monitor(self.env, log_dir)
def _thunk():
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.configure(config)
envSeed = seed + rank if seed is not None else None
# environment.render_axis = ax
env.thisSeed = envSeed
env.nenv = envNum
if envNum > 1:
env.phase = 'train'
else:
env.phase = 'test'
if ax:
env.render_axis = ax
if test_case >= 0:
env.test_case = test_case
env.seed(seed + rank)
if str(env.__class__.__name__).find('TimeLimit') >= 0:
env = TimeLimitMask(env)
# if log_dir is not None:
env = bench.Monitor(env, None, allow_early_resets=allow_early_resets)
print(env)
if isinstance(env.observation_space, Box):
if is_atari:
if len(env.observation_space.shape) == 3:
env = wrap_deepmind(env)
elif len(env.observation_space.shape) == 3:
raise NotImplementedError(
"CNN models work only for atari,\n"
"please use a custom wrapper for a custom pixel input env.\n"
"See wrap_deepmind for an example.")
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env, op=[2, 0, 1])
return env
def _thunk():
if env_id.startswith("Sawyer"):
env = robosuite.make(
env_id,
has_renderer=False,
has_offscreen_renderer=False,
ignore_done=False,
horizon=1000, #200,
use_camera_obs=False,
gripper_visualization=False,
reward_shaping=True,
control_freq=100 #10, # 100
)
env = GymGymWrapper(env)
elif env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
obs_shape = env.observation_space.shape
if str(env.__class__.__name__).find('TimeLimit') >= 0:
env = TimeLimitMask(env)
if log_dir is not None:
env = bench.Monitor(env,
os.path.join(log_dir, str(rank)),
allow_early_resets=allow_early_resets)
if is_atari:
if len(env.observation_space.shape) == 3:
env = wrap_deepmind(env)
elif len(env.observation_space.shape) == 3:
raise NotImplementedError(
"CNN models work only for atari,\n"
"please use a custom wrapper for a custom pixel input env.\n"
"See wrap_deepmind for an example.")
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env, op=[2, 0, 1])
return env
def _thunk():
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
if str(env.__class__.__name__).find('TimeLimit') >= 0:
env = TimeLimitMask(env)
# minigrid
keep_classes = ['agent', 'goal', 'wall', 'empty']
if 'key' in env_id.lower():
keep_classes.extend(['door', 'key'])
if env_id.startswith('MiniGrid'):
env = mgwr.FullyObsWrapper(env)
env = mgwr.ImgObsWrapper(env)
env = mgwr.FullyObsOneHotWrapper(env,
drop_color=1,
keep_classes=keep_classes,
flatten=False)
if log_dir is not None:
env = bench.Monitor(env,
os.path.join(log_dir, str(rank)),
allow_early_resets=allow_early_resets)
if is_atari:
if len(env.observation_space.shape) == 3:
env = wrap_deepmind(env)
elif len(env.observation_space.shape) == 3:
if env_id.startswith('CarRacing'):
env = WarpFrame(env, width=96, height=96, grayscale=True)
env = ScaledFloatFrame(env)
else:
raise NotImplementedError
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3:
env = TransposeImage(env, op=[2, 0, 1])
return env
def _thunk():
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
elif env_id == "tetris_single":
env = TetrisSingleEnv(obs_type=obs_type, mode=mode)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
obs_shape = env.observation_space.shape
if str(env.__class__.__name__).find('TimeLimit') >= 0:
env = TimeLimitMask(env)
# if log_dir is not None:
# env = bench.Monitor(
# env,
# os.path.join(log_dir, str(rank)),
# allow_early_resets=allow_early_resets)
if is_atari:
if len(env.observation_space.shape) == 3:
env = wrap_deepmind(env)
elif env_id.startswith("tetris"):
# env = wrap_deepmind(env, episode_life=False, clip_rewards=False, frame_stack=False, scale=False)
# env = NoopResetEnv(env, noop_max=30)
env = MaxAndSkipEnv(env, skip=skip_frames)
# pass
if obs_type == "image":
env = WarpFrame(env, 224, 224)
elif len(env.observation_space.shape) == 3:
raise NotImplementedError(
"CNN models work only for atari,\n"
"please use a custom wrapper for a custom pixel input env.\n"
"See wrap_deepmind for an example.")
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env, op=[2, 0, 1])
return env
def _thunk():
if env_id.find("doorenv") > -1:
env = gym.make(env_id, **env_kwargs) #
env._max_episode_steps = 512
if env_kwargs['unity']:
env.env.init(rank)
elif env_id.find('Fetch') > -1:
env = gym.make(env_id, reward_type="sparse")
env = gym.wrappers.FlattenDictWrapper(
env, dict_keys=['observation', 'desired_goal'])
else:
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
env._max_episode_steps = 512
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
obs_shape = env.observation_space.shape
if str(env.__class__.__name__).find('TimeLimit') >= 0:
env = TimeLimitMask(env)
if log_dir is not None:
env = bench.Monitor(env,
os.path.join(log_dir, str(rank)),
allow_early_resets=allow_early_resets)
if is_atari:
if len(env.observation_space.shape) == 3:
env = wrap_deepmind(env)
elif len(env.observation_space.shape) == 3:
raise NotImplementedError(
"CNN models work only for atari,\n"
"please use a custom wrapper for a custom pixel input env.\n"
"See wrap_deepmind for an example.")
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env, op=[2, 0, 1])
return env
def _thunk():
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
if 'gameoflife' in env_id.lower():
#print("ENV RANK: ", rank)
power_puzzle = False
#if args.power_puzzle:
# power_puzzle = True
if rank == 0:
render = render_gui
else:render = False
env.configure(map_width=map_width, render=render)
#env.setMapSize(map_width, print_map=print_map, render_gui=render, empty_start=not args.random_terrain, max_step=max_step, rank=rank, simple_reward=simple_reward, power_puzzle=power_puzzle)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
obs_shape = env.observation_space.shape
if add_timestep and len(
obs_shape) == 1 and str(env).find('TimeLimit') > -1:
env = AddTimestep(env)
if log_dir is not None:
env = MicropolisMonitor(env, os.path.join(log_dir, str(rank)),
allow_early_resets=allow_early_resets)
#print(log_dir)
#
#print(type(env))
#print(dir(env))
#raise Exception
if is_atari:
env = wrap_deepmind(env)
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env)
return env
def _thunk():
#print("Make envs", params)
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id, **TspParams.current().ENVIRONMENT_KWARGS)
is_atari = (hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv))
if is_atari:
env = make_atari(env_id)
is_minigrid = "MiniGrid" in env_id
if set_time_limit is not None:
env = TimeLimit(env, set_time_limit)
env.seed(seed + rank)
obs_shape = env.observation_space.shape
if str(env.__class__.__name__).find(
'TimeLimit') >= 0 or set_time_limit is not None:
env = TimeLimitMask(env)
if log_dir is not None:
env = bench.Monitor(env,
os.path.join(log_dir, str(rank)),
allow_early_resets=allow_early_resets)
if is_atari:
if len(env.observation_space.shape) == 3:
env = wrap_deepmind(env)
elif is_minigrid:
pass
elif len(env.observation_space.shape) == 3:
raise NotImplementedError(
"CNN models work only for atari,\n"
"please use a custom wrapper for a custom pixel input env.\n"
"See wrap_deepmind for an example.")
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env, op=[2, 0, 1])
return env
def _thunk():
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = NoopResetEnv(env, noop_max=30)
env = MaxAndSkipEnv(env, skip=4)
env.seed(seed + rank)
if str(env.__class__.__name__).find('TimeLimit') >= 0:
env = TimeLimitMask(env)
if log_dir is not None:
env = Monitor(
env,
os.path.join(log_dir, str(rank)),
allow_early_resets=allow_early_resets)
if is_atari:
if len(env.observation_space.shape) == 3:
env = EpisodicLifeEnv(env)
if "FIRE" in env.unwrapped.get_action_meanings():
env = FireResetEnv(env)
env = WarpFrame(env, width=84, height=84)
env = ClipRewardEnv(env)
elif len(env.observation_space.shape) == 3:
# env = EpisodicLifeEnv(env)
# if 'FIRE' in env.unwrapped.get_action_meanings():
# env = FireResetEnv(env)
env = WarpFrame(env, width=64, height=64)
# raise NotImplementedError(
# "CNN models work only for atari,\n"
# "please use a custom wrapper for a custom pixel input env.\n"
# "See wrap_deepmind for an example.")
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env, op=[2, 0, 1])
return env
def make_env(args):
if args.env_type == 'dmc':
import dm_control2gym
env = dm_control2gym.make(
domain_name=args.domain_name,
task_name=args.task_name,
task_kwargs={'random': args.seed},
visualize_reward=True)
elif args.env_type == 'mjc':
from pointmass import point_mass
env = gym.make(args.env_name)
elif args.env_type == 'ant':
import ant_env_mujoco
env = gym.make(args.env_name)
else:
assert 'unknown env type: %s' % args.env_type
return env
def __init__(self,
domain_name,
task_name,
max_steps=sys.maxsize,
log_dir=None):
from dm_control import suite
import dm_control2gym
BasicTask.__init__(self, max_steps)
self.name = domain_name + '_' + task_name
self.env = dm_control2gym.make(domain_name, task_name)
self.action_dim = self.env.action_space.shape[0]
self.state_dim = self.env.observation_space.shape[0]
if log_dir is not None:
mkdir(log_dir)
self.env = Monitor(self.env, '%s/%s' % (log_dir, uuid.uuid1()))
def _thunk():
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
if 'micropolis' in env_id.lower():
print("ENV RANK: ", rank)
if rank == 0:
env.setMapSize(map_width,
print_map=print_map,
parallel_gui=parallel_py2gui,
render_gui=render_gui,
empty_start=True,
noreward=noreward,
max_step=max_step,
rank=rank)
else:
env.setMapSize(map_width, rank=rank)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
env.seed(seed + rank)
obs_shape = env.observation_space.shape
if add_timestep and len(
obs_shape) == 1 and str(env).find('TimeLimit') > -1:
env = AddTimestep(env)
if log_dir is not None:
env = bench.Monitor(env,
os.path.join(log_dir, str(rank)),
allow_early_resets=allow_early_resets)
if is_atari:
env = wrap_deepmind(env)
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env)
return env
def _thunk():
print("CUSTOM GYM:", custom_gym)
if custom_gym is not None and custom_gym != "":
module = importlib.import_module(custom_gym, package=None)
print("imported env '{}'".format((custom_gym)))
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = make_atari(env_id)
if not is_atari and scale_img:
env = WarpFrame(env, color_img)
if duckietown:
env = DuckietownRewardWrapper(env)
if dt_discrete:
env = DuckietownDiscreteWrapper(env)
env = Normalize(env)
env.seed(seed + rank)
obs_shape = env.observation_space.shape
if add_timestep and len(
obs_shape) == 1 and str(env).find('TimeLimit') > -1:
env = AddTimestep(env)
if log_dir is not None:
env = bench.Monitor(env,
os.path.join(log_dir, str(rank)),
allow_early_resets=True)
if is_atari:
env = wrap_deepmind(env)
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = WrapPyTorch(env)
return env