OffPolicyStochActorDoubleCritic

source

OffPolicyStochActorDoubleCritic(
   observation_space: gym.Space, action_space: gym.Space, feature_dim: int = 64,
   hidden_dim: int = 1024, opt_class: Type[th.optim.Optimizer] = th.optim.Adam,
   opt_kwargs: Optional[Dict[str, Any]] = None, log_std_range: Tuple = (-5, 2),
   init_fn: str = 'orthogonal'
)

---

Stochastic actor network and double critic network for off-policy algortithms like SAC. Here the 'self.dist' refers to an sampling distribution instance.

Args

• observation_space (gym.Space) : Observation space.
• action_space (gym.Space) : Action space.
• feature_dim (int) : Number of features accepted.
• hidden_dim (int) : Number of units per hidden layer.
• opt_class (Type[th.optim.Optimizer]) : Optimizer class.
• opt_kwargs (Dict[str, Any]) : Optimizer keyword arguments.
• log_std_range (Tuple) : Range of log standard deviation.
• init_fn (str) : Parameters initialization method.

Returns

Actor-Critic network.

Methods:

.describe

source

.describe()

---

Describe the policy.

.freeze

source

.freeze(
   encoder: nn.Module, dist: Distribution
)

---

Freeze all the elements like encoder and dist.

Args

• encoder (nn.Module) : Encoder network.
• dist (Distribution) : Distribution class.

Returns

None.

.forward

source

.forward(
   obs: th.Tensor, training: bool = True
)

---

Sample actions based on observations.

Args

• obs (th.Tensor) : Observations.
• training (bool) : Training mode, True or False.

Returns

Sampled actions.

.get_dist

source

.get_dist(
   obs: th.Tensor
)

---

Get sample distribution.

Args

• obs (th.Tensor) : Observations.
• step (int) : Global training step.

Returns

Action distribution.

.save

source

.save(
   path: Path, pretraining: bool, global_step: int
)

---

Save models.

Args

• path (Path) : Save path.
• pretraining (bool) : Pre-training mode.
• global_step (int) : Global training step.

Returns

None.