def forward(self, state, deterministic):
with torch.no_grad():
if isinstance(state, np.ndarray):
state = from_np(state, 'cpu')
action, dist_params = self.policy.select_action(
state, deterministic)
return CentralizedOutput(
action=LiteralActionTransformation(action),
dist_params=dist_params)
def forward(self, state, deterministic):
with torch.no_grad():
if isinstance(state, np.ndarray):
state = from_np(state, 'cpu')
action, dist_params = self.policy.select_action(
state, deterministic)
if self.policy.discrete:
action = self.transformations[action]
else:
# must be a leaf policy
action = LiteralActionTransformation(action)
return CentralizedOutput(action=action, dist_params=dist_params)
def create_transformation_builder(cls, state_dim, action_dim, args):
if args.hrl:
transform_policy = cls.centralized_policy_switch(
state_dim=state_dim, action_dim=action_dim, args=args)
transform_valuefn = cls.value_switch(state_dim=state_dim,
args=args)
if args.shared_vfn:
shared_transform_valuefn = transform_valuefn()
transformation_builder = lambda id_num: SubpolicyTransformation(
id_num=id_num,
networks=dict(
policy=transform_policy(),
valuefn=shared_transform_valuefn,
),
transformations=OrderedDict([(
i, LiteralActionTransformation(id_num=i))
for i in range(action_dim)]),
replay_buffer=PathMemory(max_replay_buffer_size=args.
max_buffer_size),
args=args)
else:
transformation_builder = lambda id_num: SubpolicyTransformation(
id_num=id_num,
networks=dict(
policy=transform_policy(),
valuefn=transform_valuefn(),
),
transformations=OrderedDict([(
i, LiteralActionTransformation(id_num=i))
for i in range(action_dim)]),
replay_buffer=PathMemory(max_replay_buffer_size=args.
max_buffer_size),
args=args)
else:
transformation_builder = lambda id_num: LiteralActionTransformation(
id_num)
return transformation_builder