def heuristic_target_entropy(action_space):
if is_continuous_space(action_space):
heuristic_target_entropy = -np.prod(action_space.shape)
elif is_discrete_space(action_space):
raise NotImplementedError(
"TODO(hartikainen): implement for discrete spaces.")
else:
raise NotImplementedError((type(action_space), action_space))
return heuristic_target_entropy
def __init__(self,
domain,
task,
*args,
env=None,
rescale_action_range=(-1.0, 1.0),
rescale_observation_range=None,
observation_keys=(),
goal_keys=(),
unwrap_time_limit=True,
pixel_wrapper_kwargs=None,
**kwargs):
assert not args, (
"Gym environments don't support args. Use kwargs instead.")
self.rescale_action_range = rescale_action_range
self.rescale_observation_range = rescale_observation_range
self.unwrap_time_limit = unwrap_time_limit
super(GymAdapter, self).__init__(domain,
task,
*args,
goal_keys=goal_keys,
**kwargs)
if env is None:
assert (domain is not None and task is not None), (domain, task)
try:
env_id = f"{domain}-{task}"
env = gym.envs.make(env_id, **kwargs)
except gym.error.UnregisteredEnv:
env_id = f"{domain}{task}"
env = gym.envs.make(env_id, **kwargs)
self._env_kwargs = kwargs
else:
assert not kwargs
assert domain is None and task is None, (domain, task)
if isinstance(env, wrappers.TimeLimit) and unwrap_time_limit:
# Remove the TimeLimit wrapper that sets 'done = True' when
# the time limit specified for each environment has been passed and
# therefore the environment is not Markovian (terminal condition
# depends on time rather than state).
env = env.env
if rescale_observation_range:
env = RescaleObservation(env, *rescale_observation_range)
if rescale_action_range and is_continuous_space(env.action_space):
env = wrappers.RescaleAction(env, *rescale_action_range)
# TODO(hartikainen): We need the clip action wrapper because sometimes
# the tfp.bijectors.Tanh() produces values strictly greater than 1 or
# strictly less than -1, which causes the env fail without clipping.
# The error is in the order of 1e-7, which should not cause issues.
# See https://github.com/tensorflow/probability/issues/664.
env = wrappers.ClipAction(env)
if pixel_wrapper_kwargs is not None:
env = wrappers.PixelObservationWrapper(env, **pixel_wrapper_kwargs)
self._env = env
if isinstance(self._env.observation_space, spaces.Dict):
dict_observation_space = self._env.observation_space
self.observation_keys = (observation_keys or
(*self.observation_space.spaces.keys(), ))
elif isinstance(self._env.observation_space, spaces.Box):
dict_observation_space = spaces.Dict(
OrderedDict(((DEFAULT_OBSERVATION_KEY,
self._env.observation_space), )))
self.observation_keys = (DEFAULT_OBSERVATION_KEY, )
self._observation_space = type(dict_observation_space)([
(name, copy.deepcopy(space))
for name, space in dict_observation_space.spaces.items()
if name in self.observation_keys + self.goal_keys
])
if len(self._env.action_space.shape) > 1:
raise NotImplementedError(
"Shape of the action space ({}) is not flat, make sure to"
" check the implemenation.".format(self._env.action_space))
self._action_space = self._env.action_space
