def act(self, obs: rlt.FeatureData) -> rlt.ActorOutput:
""" Act randomly regardless of the observation. """
obs: torch.Tensor = obs.float_features
assert obs.dim() >= 2, f"obs has shape {obs.shape} (dim < 2)"
batch_size = obs.size(0)
# pyre-fixme[6]: Expected `Union[torch.Size, torch.Tensor]` for 1st param
# but got `Tuple[int]`.
action = self.dist.sample((batch_size, ))
log_prob = self.dist.log_prob(action)
return rlt.ActorOutput(action=action, log_prob=log_prob)
def act(self, obs: rlt.FeatureData) -> rlt.ActorOutput:
""" Act randomly regardless of the observation. """
obs: torch.Tensor = obs.float_features
assert obs.dim() >= 2, f"obs has shape {obs.shape} (dim < 2)"
batch_size = obs.shape[0]
weights = torch.ones((batch_size, self.num_actions))
# sample a random action
m = torch.distributions.Categorical(weights)
raw_action = m.sample()
action = F.one_hot(raw_action, self.num_actions)
log_prob = m.log_prob(raw_action).float()
return rlt.ActorOutput(action=action, log_prob=log_prob)
def act(
self, obs: rlt.FeatureData, possible_actions_mask: Optional[np.ndarray] = None
) -> rlt.ActorOutput:
""" Act randomly regardless of the observation. """
obs: torch.Tensor = obs.float_features
assert obs.dim() >= 2, f"obs has shape {obs.shape} (dim < 2)"
batch_size = obs.size(0)
# pyre-fixme[6]: Expected `Union[torch.Size, torch.Tensor]` for 1st param
# but got `Tuple[int]`.
action = self.dist.sample((batch_size,))
# sum over action_dim (since assuming i.i.d. per coordinate)
log_prob = self.dist.log_prob(action).sum(1)
return rlt.ActorOutput(action=action, log_prob=log_prob)
def act(
self, obs: rlt.FeatureData, possible_actions_mask: Optional[np.ndarray] = None
) -> rlt.ActorOutput:
""" Act randomly regardless of the observation. """
obs: torch.Tensor = obs.float_features
assert obs.dim() >= 2, f"obs has shape {obs.shape} (dim < 2)"
assert obs.shape[0] == 1, f"obs has shape {obs.shape} (0th dim != 1)"
batch_size = obs.shape[0]
scores = torch.ones((batch_size, self.num_actions))
scores = apply_possible_actions_mask(
scores, possible_actions_mask, invalid_score=0.0
)
# sample a random action
m = torch.distributions.Categorical(scores)
raw_action = m.sample()
action = F.one_hot(raw_action, self.num_actions)
log_prob = m.log_prob(raw_action).float()
return rlt.ActorOutput(action=action, log_prob=log_prob)