examples/atari/tianshou/highlevel/module/actor.py

from abc import ABC, abstractmethod
from collections.abc import Sequence
from dataclasses import dataclass
from enum import Enum
from typing import Protocol

import torch
from torch import nn

from tianshou.highlevel.env import Environments, EnvType
from tianshou.highlevel.module.core import (
    ModuleFactory,
    TDevice,
    init_linear_orthogonal,
)
from tianshou.highlevel.module.intermediate import (
    IntermediateModule,
    IntermediateModuleFactory,
)
from tianshou.highlevel.module.module_opt import ModuleOpt
from tianshou.highlevel.optim import OptimizerFactory
from tianshou.utils.net import continuous, discrete
from tianshou.utils.net.common import BaseActor, ModuleType, Net
from tianshou.utils.string import ToStringMixin


class ContinuousActorType(Enum):
    GAUSSIAN = "gaussian"
    DETERMINISTIC = "deterministic"
    UNSUPPORTED = "unsupported"


@dataclass
class ActorFuture:
    """Container, which, in the future, will hold an actor instance."""

    actor: BaseActor | nn.Module | None = None


class ActorFutureProviderProtocol(Protocol):
    def get_actor_future(self) -> ActorFuture:
        pass


class ActorFactory(ModuleFactory, ToStringMixin, ABC):
    @abstractmethod
    def create_module(self, envs: Environments, device: TDevice) -> BaseActor | nn.Module:
        pass

    def create_module_opt(
        self,
        envs: Environments,
        device: TDevice,
        optim_factory: OptimizerFactory,
        lr: float,
    ) -> ModuleOpt:
        """Creates the actor module along with its optimizer for the given learning rate.

        :param envs: the environments
        :param device: the torch device
        :param optim_factory: the optimizer factory
        :param lr: the learning rate
        :return: a container with the actor module and its optimizer
        """
        module = self.create_module(envs, device)
        optim = optim_factory.create_optimizer(module, lr)
        return ModuleOpt(module, optim)

    @staticmethod
    def _init_linear(actor: torch.nn.Module) -> None:
        """Initializes linear layers of an actor module using default mechanisms.

        :param module: the actor module.
        """
        init_linear_orthogonal(actor)
        if hasattr(actor, "mu"):
            # For continuous action spaces with Gaussian policies
            # do last policy layer scaling, this will make initial actions have (close to)
            # 0 mean and std, and will help boost performances,
            # see https://arxiv.org/abs/2006.05990, Fig.24 for details
            for m in actor.mu.modules():
                if isinstance(m, torch.nn.Linear):
                    m.weight.data.copy_(0.01 * m.weight.data)


class ActorFactoryDefault(ActorFactory):
    """An actor factory which, depending on the type of environment, creates a suitable MLP-based policy."""

    DEFAULT_HIDDEN_SIZES = (64, 64)

    def __init__(
        self,
        continuous_actor_type: ContinuousActorType,
        hidden_sizes: Sequence[int] = DEFAULT_HIDDEN_SIZES,
        hidden_activation: ModuleType = nn.ReLU,
        continuous_unbounded: bool = False,
        continuous_conditioned_sigma: bool = False,
        discrete_softmax: bool = True,
    ):
        self.continuous_actor_type = continuous_actor_type
        self.continuous_unbounded = continuous_unbounded
        self.continuous_conditioned_sigma = continuous_conditioned_sigma
        self.hidden_sizes = hidden_sizes
        self.hidden_activation = hidden_activation
        self.discrete_softmax = discrete_softmax

    def create_module(self, envs: Environments, device: TDevice) -> BaseActor:
        env_type = envs.get_type()
        factory: ActorFactoryContinuousDeterministicNet | ActorFactoryContinuousGaussianNet | ActorFactoryDiscreteNet
        if env_type == EnvType.CONTINUOUS:
            match self.continuous_actor_type:
                case ContinuousActorType.GAUSSIAN:
                    factory = ActorFactoryContinuousGaussianNet(
                        self.hidden_sizes,
                        activation=self.hidden_activation,
                        unbounded=self.continuous_unbounded,
                        conditioned_sigma=self.continuous_conditioned_sigma,
                    )
                case ContinuousActorType.DETERMINISTIC:
                    factory = ActorFactoryContinuousDeterministicNet(
                        self.hidden_sizes,
                        activation=self.hidden_activation,
                    )
                case ContinuousActorType.UNSUPPORTED:
                    raise ValueError("Continuous action spaces are not supported by the algorithm")
                case _:
                    raise ValueError(self.continuous_actor_type)
            return factory.create_module(envs, device)
        elif env_type == EnvType.DISCRETE:
            factory = ActorFactoryDiscreteNet(
                self.DEFAULT_HIDDEN_SIZES,
                softmax_output=self.discrete_softmax,
            )
            return factory.create_module(envs, device)
        else:
            raise ValueError(f"{env_type} not supported")


class ActorFactoryContinuous(ActorFactory, ABC):
    """Serves as a type bound for actor factories that are suitable for continuous action spaces."""


class ActorFactoryContinuousDeterministicNet(ActorFactoryContinuous):
    def __init__(self, hidden_sizes: Sequence[int], activation: ModuleType = nn.ReLU):
        self.hidden_sizes = hidden_sizes
        self.activation = activation

    def create_module(self, envs: Environments, device: TDevice) -> BaseActor:
        net_a = Net(
            state_shape=envs.get_observation_shape(),
            hidden_sizes=self.hidden_sizes,
            activation=self.activation,
            device=device,
        )
        return continuous.Actor(
            preprocess_net=net_a,
            action_shape=envs.get_action_shape(),
            hidden_sizes=(),
            device=device,
        ).to(device)


class ActorFactoryContinuousGaussianNet(ActorFactoryContinuous):
    def __init__(
        self,
        hidden_sizes: Sequence[int],
        unbounded: bool = True,
        conditioned_sigma: bool = False,
        activation: ModuleType = nn.ReLU,
    ):
        """For actors with Gaussian policies.

        :param hidden_sizes: the sequence of hidden dimensions to use in the network structure
        :param unbounded: whether to apply tanh activation on final logits
        :param conditioned_sigma: if True, the standard deviation of continuous actions (sigma) is computed from the
            input; if False, sigma is an independent parameter
        """
        self.hidden_sizes = hidden_sizes
        self.unbounded = unbounded
        self.conditioned_sigma = conditioned_sigma
        self.activation = activation

    def create_module(self, envs: Environments, device: TDevice) -> BaseActor:
        net_a = Net(
            state_shape=envs.get_observation_shape(),
            hidden_sizes=self.hidden_sizes,
            activation=self.activation,
            device=device,
        )
        actor = continuous.ActorProb(
            preprocess_net=net_a,
            action_shape=envs.get_action_shape(),
            unbounded=self.unbounded,
            device=device,
            conditioned_sigma=self.conditioned_sigma,
        ).to(device)

        # init params
        if not self.conditioned_sigma:
            torch.nn.init.constant_(actor.sigma_param, -0.5)
        self._init_linear(actor)

        return actor


class ActorFactoryDiscreteNet(ActorFactory):
    def __init__(
        self,
        hidden_sizes: Sequence[int],
        softmax_output: bool = True,
        activation: ModuleType = nn.ReLU,
    ):
        self.hidden_sizes = hidden_sizes
        self.softmax_output = softmax_output
        self.activation = activation

    def create_module(self, envs: Environments, device: TDevice) -> BaseActor:
        net_a = Net(
            state_shape=envs.get_observation_shape(),
            hidden_sizes=self.hidden_sizes,
            activation=self.activation,
            device=device,
        )
        return discrete.Actor(
            net_a,
            envs.get_action_shape(),
            hidden_sizes=(),
            device=device,
            softmax_output=self.softmax_output,
        ).to(device)


class ActorFactoryTransientStorageDecorator(ActorFactory):
    """Wraps an actor factory, storing the most recently created actor instance such that it can be retrieved."""

    def __init__(self, actor_factory: ActorFactory, actor_future: ActorFuture):
        self.actor_factory = actor_factory
        self._actor_future = actor_future

    def __getstate__(self) -> dict:
        d = dict(self.__dict__)
        del d["_actor_future"]
        return d

    def __setstate__(self, state: dict) -> None:
        self.__dict__ = state
        self._actor_future = ActorFuture()

    def _tostring_excludes(self) -> list[str]:
        return [*super()._tostring_excludes(), "_actor_future"]

    def create_module(self, envs: Environments, device: TDevice) -> BaseActor | nn.Module:
        module = self.actor_factory.create_module(envs, device)
        self._actor_future.actor = module
        return module


class IntermediateModuleFactoryFromActorFactory(IntermediateModuleFactory):
    def __init__(self, actor_factory: ActorFactory):
        self.actor_factory = actor_factory

    def create_intermediate_module(self, envs: Environments, device: TDevice) -> IntermediateModule:
        actor = self.actor_factory.create_module(envs, device)
        assert isinstance(actor, BaseActor)
        return IntermediateModule(actor, actor.get_output_dim())