def wrap_env(env, test, env_id, monitor, outdir, frame_skip, gray_scale,
frame_stack, disable_action_prior, always_keys, reverse_keys,
exclude_keys, exclude_noop, randomize_action, eval_epsilon):
# wrap env: time limit...
import gym
if isinstance(env, gym.wrappers.TimeLimit):
logger.info(
'Detected `gym.wrappers.TimeLimit`! Unwrap it and re-wrap our own time limit.'
)
env = env.env
max_episode_steps = env.spec.max_episode_steps
env = ContinuingTimeLimit(env, max_episode_steps=max_episode_steps)
# wrap env: observation...
# NOTE: wrapping order matters!
if test and monitor:
env = ContinuingTimeLimitMonitor(
env,
os.path.join(outdir, env.spec.id, 'monitor'),
mode='evaluation' if test else 'training',
video_callable=lambda episode_id: True)
if frame_skip is not None:
env = FrameSkip(env, skip=frame_skip)
if gray_scale:
env = GrayScaleWrapper(env, dict_space_key='pov')
if env_id.startswith('MineRLNavigate'):
env = PoVWithCompassAngleWrapper(env)
else:
env = ObtainPoVWrapper(env)
env = MoveAxisWrapper(
env, source=-1,
destination=0) # convert hwc -> chw as Chainer requires.
env = ScaledFloatFrame(env)
if frame_stack is not None and frame_stack > 0:
env = FrameStack(env, frame_stack, channel_order='chw')
# wrap env: action...
if not disable_action_prior:
env = SerialDiscreteActionWrapper(env,
always_keys=always_keys,
reverse_keys=reverse_keys,
exclude_keys=exclude_keys,
exclude_noop=exclude_noop)
else:
env = CombineActionWrapper(env)
env = SerialDiscreteCombineActionWrapper(env)
if randomize_action:
env = RandomizeAction(env, eval_epsilon)
return env
def wrap_env(env, test):
# wrap env: time limit...
if isinstance(env, gym.wrappers.TimeLimit):
logger.info(
'Detected `gym.wrappers.TimeLimit`! Unwrap it and re-wrap our own time limit.'
)
env = env.env
max_episode_steps = env.spec.max_episode_steps
env = ContinuingTimeLimit(env, max_episode_steps=max_episode_steps)
# wrap env: observation...
# NOTE: wrapping order matters!
if test and args.monitor:
env = ContinuingTimeLimitMonitor(
env,
os.path.join(args.outdir, 'monitor'),
mode='evaluation' if test else 'training',
video_callable=lambda episode_id: True)
if args.frame_skip is not None:
env = FrameSkip(env, skip=args.frame_skip)
if args.gray_scale:
env = GrayScaleWrapper(env, dict_space_key='pov')
if args.env.startswith('MineRLNavigate'):
env = PoVWithCompassAngleWrapper(env)
else:
env = ObtainPoVWrapper(env)
env = MoveAxisWrapper(
env, source=-1,
destination=0) # convert hwc -> chw as Chainer requires.
env = ScaledFloatFrame(env)
if args.frame_stack is not None and args.frame_stack > 0:
env = FrameStack(env, args.frame_stack, channel_order='chw')
# wrap env: action...
if not args.disable_action_prior:
env = SerialDiscreteActionWrapper(env,
always_keys=args.always_keys,
reverse_keys=args.reverse_keys,
exclude_keys=args.exclude_keys,
exclude_noop=args.exclude_noop)
else:
env = CombineActionWrapper(env)
env = SerialDiscreteCombineActionWrapper(env)
env_seed = test_seed if test else train_seed
# env.seed(int(env_seed)) # TODO: not supported yet
return env