def wrap_modified_rr(env, episode_life=True, episode_reward=False, episode_frame=False,
norm_rewards=True,
frame_stack=False, scale=False):
"""Configure environment for DeepMind-style Atari modified as described in RUDDER paper;
"""
if episode_life:
print("Episode Life")
env = EpisodicLifeEnv(env)
if episode_reward:
print("Episode Reward")
env = EpisodicRewardEnv(env)
if episode_frame:
print("Episode Frame")
env = EpisodicFrameEnv(env)
_ori_r_games = ['DoubleDunk', 'Boxing', 'Freeway', 'Pong',
'Bowling', 'Skiing', 'IceHockey', 'Enduro']
original_reward = any([game in env.spec.id for game in _ori_r_games])
if 'FIRE' in env.unwrapped.get_action_meanings():
env = FireResetEnv(env)
env = WarpFrame(env)
if scale:
env = ScaledFloatFrame(env)
if norm_rewards and not original_reward:
print("Normalizing reward....")
env = NormRewardEnv(env, 100.)
else:
print("Normal reward")
if frame_stack:
env = FrameStack(env, 4)
return env
def wrap_deepmind(env,
downsample=True,
episode_life=True,
clip_rewards=True,
frame_stack=False,
scale=False,
color=False):
"""Configure environment for DeepMind-style Atari.
"""
if ("videopinball" in str(env.spec.id).lower()) or ('tennis' in str(
env.spec.id).lower()):
env = WarpFrame(env, width=160, height=210, grayscale=False)
if episode_life:
env = EpisodicLifeEnv(env)
if 'FIRE' in env.unwrapped.get_action_meanings():
env = FireResetEnv(env)
if downsample:
env = WarpFrame(env, grayscale=False)
if not color:
env = GrayscaleWrapper(env)
if scale:
env = ScaledFloatFrame(env)
if clip_rewards:
env = ClipRewardEnv(env)
if frame_stack:
env = FrameStack(env, 4)
return env
def atari_play_env(env):
env = EpisodicLifeEnv(env)
if 'FIRE' in env.unwrapped.get_action_meanings():
env = FireResetEnv(env)
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 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 wrap_deepmind(env, episode_life=True, clip_rewards=True, frame_stack=False, scale=False):
"""Configure environment for DeepMind-style Atari.
"""
if episode_life:
env = EpisodicLifeEnv(env)
if 'FIRE' in env.unwrapped.get_action_meanings():
env = FireResetEnv(env)
env = WarpFrame(env)
if scale:
env = ScaledFloatFrame(env)
if clip_rewards:
env = ClipRewardEnv(env)
if frame_stack:
env = FrameStack(env, 4)
return env
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 wrap_atari(env, stack_frames=4, training=True, crop_top=True):
if training:
env = EpisodicLifeEnv(env)
env = NoopResetEnv(env)
if 'FIRE' in env.unwrapped.get_action_meanings():
if (training):
env = FireResetEnv(env)
else:
env = FireAfterLifeLost(env)
env = ProcessFrame(env, 18 if crop_top else 0)
env = ImageToPyTorch(env)
env = FrameStack(env, stack_frames)
return env
def wrap_deepmind_custom(env,
episode_life=True,
clip_rewards=True,
frame_stack=frame_stack,
scale=False):
if episode_life:
env = EpisodicLifeEnv(env)
if 'FIRE' in env.unwrapped.get_action_meanings():
env = FireResetEnv(env)
env = WarpFrame(env, size=size)
if augment:
env = AugmentColor(env)
if scale:
env = ScaledFloatFrame(env)
if clip_rewards:
env = ClipRewardEnv(env)
if frame_stack:
env = FrameStack(env, frame_stack)
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 deepmind_wrap(atari_env,
episode_life=True,
clip_rewards=True,
frame_stack=False,
scale=False):
""" matching deepmind papers
"""
if episode_life:
env = EpisodicLifeEnv(atari_env)
if 'FIRE' in env.unwrapped.get_action_meanings():
env = FireResetEnv(env)
env = WarpFrame(env)
if scale:
env = ScaledFloatFrame(env)
if clip_rewards:
env = ClipRewardEnv(env)
if frame_stack:
env = FrameStack(env, 4)
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 __init__(self,
env,
stack=4,
episodic_life=False,
episodic_score=False,
clip_reward=False):
"""
- grayscale
- resize
- crop
- stack
See Also
--------
baselines.common.atari_wrappers.LazyFrames
"""
# env = EpisodicLifeEnv(env) # no effect on Pong
if episodic_life:
env = EpisodicLifeEnv(env)
gym.Wrapper.__init__(self, env)
self.stack = stack
self.frames = deque([], maxlen=stack)
shp = env.observation_space.shape
# self.observation_space = gym.spaces.Box(low=0, high=1.0, shape=(
# shp[2] * stack, shp[0], shp[1]), dtype=np.float)
self.observation_space = gym.spaces.Box(low=0,
high=255,
shape=(shp[2] * stack, shp[0],
shp[1]),
dtype=np.uint8)
self.episodic_score = episodic_score
self.clip_reward = clip_reward
def custom_wrapper(env,
deterministic_reset=True,
episode_life=True,
warp=True,
grayscale=False,
frame_stack=True,
frames=4,
ATARI_labels=True):
if deterministic_reset:
env = DeterminsticNoopResetEnv(env, 66)
if episode_life:
env = EpisodicLifeEnv(env)
if warp:
env = WarpFrame(env, height=100, width=100, grayscale=grayscale)
if not frame_stack:
assert frames == 1, "Framestack is false and frames are larger than 1."
if frame_stack:
assert frames > 1, 'Frames should not be 1 if framestack is on'
env = FrameStack(env, frames)
if ATARI_labels:
env = AtariARIWrapper(env)
return env
