actor

Source code: tianshou/highlevel/module/actor.py

---

class ContinuousActorType(value, names=None, *, module=None, qualname=None, type=None, start=1, boundary=None)

Bases: Enum

GAUSSIAN = 'gaussian'

DETERMINISTIC = 'deterministic'

UNSUPPORTED = 'unsupported'

class ActorFuture(actor: Actor | Module | None = None)

Bases: object

Container, which, in the future, will hold an actor instance.

actor: Actor | Module | None = None

class ActorFutureProviderProtocol(*args, **kwargs)

Bases: Protocol

get_actor_future() → ActorFuture

class ActorFactory

Bases: ModuleFactory, ToStringMixin, ABC

abstract create_module(envs: Environments, device: str | device) → Actor | Module

abstract create_dist_fn(envs: Environments) → Callable[[tuple[Tensor, Tensor]], Distribution] | Callable[[Tensor], Distribution] | None

Parameters:

envs – the environments

Returns:

the distribution function, which converts the actor’s output into a distribution, or None if the actor does not output distribution parameters

class ActorFactoryDefault(continuous_actor_type: ~tianshou.highlevel.module.actor.ContinuousActorType, hidden_sizes: ~collections.abc.Sequence[int] = (64, 64), hidden_activation: type[~torch.nn.modules.module.Module] = <class 'torch.nn.modules.activation.ReLU'>, continuous_unbounded: bool = False, continuous_conditioned_sigma: bool = False, discrete_softmax: bool = True)

Bases: ActorFactory

An actor factory which, depending on the type of environment, creates a suitable MLP-based policy.

DEFAULT_HIDDEN_SIZES = (64, 64)

create_module(envs: Environments, device: str | device) → Actor | Module

create_dist_fn(envs: Environments) → Callable[[tuple[Tensor, Tensor]], Distribution] | Callable[[Tensor], Distribution] | None

Parameters:

envs – the environments

Returns:

the distribution function, which converts the actor’s output into a distribution, or None if the actor does not output distribution parameters

class ActorFactoryContinuous

Bases: ActorFactory, ABC

Serves as a type bound for actor factories that are suitable for continuous action spaces.

class ActorFactoryContinuousDeterministicNet(hidden_sizes: ~collections.abc.Sequence[int], activation: type[~torch.nn.modules.module.Module] = <class 'torch.nn.modules.activation.ReLU'>)

Bases: ActorFactoryContinuous

create_module(envs: Environments, device: str | device) → Actor

create_dist_fn(envs: Environments) → Callable[[tuple[Tensor, Tensor]], Distribution] | Callable[[Tensor], Distribution] | None

Parameters:

envs – the environments

Returns:

the distribution function, which converts the actor’s output into a distribution, or None if the actor does not output distribution parameters

class ActorFactoryContinuousGaussianNet(hidden_sizes: ~collections.abc.Sequence[int], unbounded: bool = True, conditioned_sigma: bool = False, activation: type[~torch.nn.modules.module.Module] = <class 'torch.nn.modules.activation.ReLU'>)

Bases: ActorFactoryContinuous

For actors with Gaussian policies.

Parameters:

• hidden_sizes – the sequence of hidden dimensions to use in the network structure

• unbounded – whether to apply tanh activation on final logits

• conditioned_sigma – if True, the standard deviation of continuous actions (sigma) is computed from the input; if False, sigma is an independent parameter

create_module(envs: Environments, device: str | device) → Actor

create_dist_fn(envs: Environments) → Callable[[tuple[Tensor, Tensor]], Distribution] | Callable[[Tensor], Distribution] | None

Parameters:

envs – the environments

Returns:

the distribution function, which converts the actor’s output into a distribution, or None if the actor does not output distribution parameters

class ActorFactoryDiscreteNet(hidden_sizes: ~collections.abc.Sequence[int], softmax_output: bool = True, activation: type[~torch.nn.modules.module.Module] = <class 'torch.nn.modules.activation.ReLU'>)

Bases: ActorFactory

create_module(envs: Environments, device: str | device) → Actor

create_dist_fn(envs: Environments) → Callable[[tuple[Tensor, Tensor]], Distribution] | Callable[[Tensor], Distribution] | None

Parameters:

envs – the environments

Returns:

the distribution function, which converts the actor’s output into a distribution, or None if the actor does not output distribution parameters

class ActorFactoryTransientStorageDecorator(actor_factory: ActorFactory, actor_future: ActorFuture)

Bases: ActorFactory

Wraps an actor factory, storing the most recently created actor instance such that it can be retrieved.

create_module(envs: Environments, device: str | device) → Actor | Module

create_dist_fn(envs: Environments) → Callable[[tuple[Tensor, Tensor]], Distribution] | Callable[[Tensor], Distribution] | None

Parameters:

envs – the environments

Returns:

the distribution function, which converts the actor’s output into a distribution, or None if the actor does not output distribution parameters

class IntermediateModuleFactoryFromActorFactory(actor_factory: ActorFactory)

Bases: IntermediateModuleFactory

create_intermediate_module(envs: Environments, device: str | device) → IntermediateModule