takuseno/d3rlpy

import dataclasses

from ...base import DeviceArg, LearnableConfig, register_learnable
from ...constants import ActionSpace
from ...models.builders import (
    create_continuous_q_function,
    create_deterministic_policy,
)
from ...models.encoders import EncoderFactory, make_encoder_field
from ...models.optimizers import OptimizerFactory, make_optimizer_field
from ...models.q_functions import QFunctionFactory, make_q_func_field
from ...types import Shape
from .base import QLearningAlgoBase
from .torch.ddpg_impl import DDPGModules
from .torch.td3_plus_bc_impl import TD3PlusBCImpl

__all__ = ["TD3PlusBCConfig", "TD3PlusBC"]


@dataclasses.dataclass()
class TD3PlusBCConfig(LearnableConfig):
    r"""Config of TD3+BC algorithm.

    TD3+BC is an simple offline RL algorithm built on top of TD3.
    TD3+BC introduces BC-reguralized policy objective function.

    .. math::

        J(\phi) = \mathbb{E}_{s,a \sim D}
            [\lambda Q(s, \pi(s)) - (a - \pi(s))^2]

    where

    .. math::

        \lambda = \frac{\alpha}{\frac{1}{N} \sum_(s_i, a_i) |Q(s_i, a_i)|}

    References:
        * `Fujimoto et al., A Minimalist Approach to Offline Reinforcement
          Learning. <https://arxiv.org/abs/2106.06860>`_

    Args:
        observation_scaler (d3rlpy.preprocessing.ObservationScaler):
            Observation preprocessor.
        action_scaler (d3rlpy.preprocessing.ActionScaler): Action preprocessor.
        reward_scaler (d3rlpy.preprocessing.RewardScaler): Reward preprocessor.
        actor_learning_rate (float): Learning rate for a policy function.
        critic_learning_rate (float): Learning rate for Q functions.
        actor_optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
            Optimizer factory for the actor.
        critic_optim_factory (d3rlpy.models.optimizers.OptimizerFactory):
            Optimizer factory for the critic.
        actor_encoder_factory (d3rlpy.models.encoders.EncoderFactory):
            Encoder factory for the actor.
        critic_encoder_factory (d3rlpy.models.encoders.EncoderFactory):
            Encoder factory for the critic.
        q_func_factory (d3rlpy.models.q_functions.QFunctionFactory):
            Q function factory.
        batch_size (int): Mini-batch size.
        gamma (float): Discount factor.
        tau (float): Target network synchronization coefficiency.
        n_critics (int): Number of Q functions for ensemble.
        target_smoothing_sigma (float): Standard deviation for target noise.
        target_smoothing_clip (float): Clipping range for target noise.
        alpha (float): :math:`\alpha` value.
        update_actor_interval (int): Interval to update policy function
            described as `delayed policy update` in the paper.
    """

    actor_learning_rate: float = 3e-4
    critic_learning_rate: float = 3e-4
    actor_optim_factory: OptimizerFactory = make_optimizer_field()
    critic_optim_factory: OptimizerFactory = make_optimizer_field()
    actor_encoder_factory: EncoderFactory = make_encoder_field()
    critic_encoder_factory: EncoderFactory = make_encoder_field()
    q_func_factory: QFunctionFactory = make_q_func_field()
    batch_size: int = 256
    gamma: float = 0.99
    tau: float = 0.005
    n_critics: int = 2
    target_smoothing_sigma: float = 0.2
    target_smoothing_clip: float = 0.5
    alpha: float = 2.5
    update_actor_interval: int = 2

    def create(self, device: DeviceArg = False) -> "TD3PlusBC":
        return TD3PlusBC(self, device)

    @staticmethod
    def get_type() -> str:
        return "td3_plus_bc"


class TD3PlusBC(QLearningAlgoBase[TD3PlusBCImpl, TD3PlusBCConfig]):
    def inner_create_impl(
        self, observation_shape: Shape, action_size: int
    ) -> None:
        policy = create_deterministic_policy(
            observation_shape,
            action_size,
            self._config.actor_encoder_factory,
            device=self._device,
        )
        targ_policy = create_deterministic_policy(
            observation_shape,
            action_size,
            self._config.actor_encoder_factory,
            device=self._device,
        )
        q_funcs, q_func_forwarder = create_continuous_q_function(
            observation_shape,
            action_size,
            self._config.critic_encoder_factory,
            self._config.q_func_factory,
            n_ensembles=self._config.n_critics,
            device=self._device,
        )
        targ_q_funcs, targ_q_func_forwarder = create_continuous_q_function(
            observation_shape,
            action_size,
            self._config.critic_encoder_factory,
            self._config.q_func_factory,
            n_ensembles=self._config.n_critics,
            device=self._device,
        )

        actor_optim = self._config.actor_optim_factory.create(
            policy.named_modules(), lr=self._config.actor_learning_rate
        )
        critic_optim = self._config.critic_optim_factory.create(
            q_funcs.named_modules(), lr=self._config.critic_learning_rate
        )

        modules = DDPGModules(
            policy=policy,
            targ_policy=targ_policy,
            q_funcs=q_funcs,
            targ_q_funcs=targ_q_funcs,
            actor_optim=actor_optim,
            critic_optim=critic_optim,
        )

        self._impl = TD3PlusBCImpl(
            observation_shape=observation_shape,
            action_size=action_size,
            modules=modules,
            q_func_forwarder=q_func_forwarder,
            targ_q_func_forwarder=targ_q_func_forwarder,
            gamma=self._config.gamma,
            tau=self._config.tau,
            target_smoothing_sigma=self._config.target_smoothing_sigma,
            target_smoothing_clip=self._config.target_smoothing_clip,
            alpha=self._config.alpha,
            update_actor_interval=self._config.update_actor_interval,
            device=self._device,
        )

    def get_action_type(self) -> ActionSpace:
        return ActionSpace.CONTINUOUS


register_learnable(TD3PlusBCConfig)