def build_normalizer(env: EnvWrapper) -> Dict[str, NormalizationData]:
try:
return env.normalization_data
except AttributeError:
# TODO: make this a property of EnvWrapper?
# pyre-fixme[16]: Module `envs` has no attribute `RecSim`.
if HAS_RECSIM and isinstance(env, RecSim):
return {
NormalizationKey.STATE: NormalizationData(
dense_normalization_parameters=only_continuous_normalizer(
list(range(env.observation_space["user"].shape[0]))
)
),
NormalizationKey.ITEM: NormalizationData(
dense_normalization_parameters=only_continuous_normalizer(
list(range(env.observation_space["doc"]["0"].shape[0]))
)
),
}
return {
NormalizationKey.STATE: NormalizationData(
dense_normalization_parameters=build_state_normalizer(env)
),
NormalizationKey.ACTION: NormalizationData(
dense_normalization_parameters=build_action_normalizer(env)
),
}
def normalization_action(self):
if self.action_type == EnvType.DISCRETE_ACTION:
return only_continuous_normalizer(
list(range(self.state_dim, self.state_dim + self.action_dim)),
min_value=0,
max_value=1,
)
return only_continuous_normalizer(
list(range(self.state_dim, self.state_dim + self.action_dim)),
min_value=self.env.action_space.low,
max_value=self.env.action_space.high,
)
def normalization_data(self):
return {
NormalizationKey.STATE:
NormalizationData(
dense_normalization_parameters=only_continuous_normalizer(
list(range(self.num_arms)), MU_LOW, MU_HIGH))
}
def normalization(self):
if self.img:
return None
else:
return only_continuous_normalizer(
list(range(self.state_dim)),
self.env.observation_space.low,
self.env.observation_space.high,
)
def build_state_normalizer(env: EnvWrapper):
if isinstance(env.observation_space, spaces.Box):
assert (
len(env.observation_space.shape) == 1
), f"{env.observation_space.shape} has dim > 1, and is not supported."
return only_continuous_normalizer(
list(range(env.observation_space.shape[0])),
env.observation_space.low,
env.observation_space.high,
)
elif isinstance(env.observation_space, spaces.Dict):
# assuming env.observation_space is image
return None
else:
raise NotImplementedError(f"{env.observation_space} not supported")
def build_action_normalizer(env):
action_space = env.action_space
if isinstance(action_space, spaces.Discrete):
return only_continuous_normalizer(list(range(action_space.n)),
min_value=0,
max_value=1)
elif isinstance(action_space, spaces.Box):
assert action_space.shape == (
1, ), f"Box action shape {action_space.shape} not supported."
return only_continuous_action_normalizer(
[0],
min_value=action_space.low.item(),
max_value=action_space.high.item())
else:
raise NotImplementedError(f"{action_space} not supported.")
def build_normalizer(env):
if isinstance(env.observation_space, gym.spaces.Box):
assert (len(env.observation_space.shape) == 1
), f"{env.observation_space} not supported."
return {
"state":
NormalizationData(
dense_normalization_parameters=only_continuous_normalizer(
list(range(env.observation_space.shape[0])),
env.observation_space.low,
env.observation_space.high,
))
}
elif isinstance(env.observation_space, gym.spaces.Dict):
# assuming env.observation_space is image
return None
else:
raise NotImplementedError(f"{env.observation_space} not supported")
def build_action_normalizer(env: Env):
action_space = env.action_space
if isinstance(action_space, spaces.Discrete):
return only_continuous_normalizer(list(range(action_space.n)),
min_value=0,
max_value=1)
elif isinstance(action_space, spaces.Box):
assert (len(action_space.shape) == 1
), f"Box action shape {action_space.shape} not supported."
action_dim = action_space.shape[0]
return only_continuous_action_normalizer(
list(range(action_dim)),
min_value=action_space.low,
max_value=action_space.high,
)
else:
raise NotImplementedError(f"{action_space} not supported.")
def normalization_action(self):
return only_continuous_normalizer(
list(range(self.num_states, self.num_states + self.num_actions)),
min_value=0,
max_value=1,
)
