add PRDC (#431)

* add PRDC

* update README.md

* Remove useless imports

* Fix errors from lint

.gitignore

@@ -21,3 +21,4 @@ build
 dist
 /.idea/
 *.egg-info
+/.vscode/

README.md

@@ -101,6 +101,7 @@ $ docker run -it --gpus all --name d3rlpy takuseno/d3rlpy:latest bash
 | [Critic Reguralized Regression (CRR)](https://arxiv.org/abs/2006.15134) | :no_entry: | :white_check_mark: |
 | [Policy in Latent Action Space (PLAS)](https://arxiv.org/abs/2011.07213) | :no_entry: | :white_check_mark: |
 | [TD3+BC](https://arxiv.org/abs/2106.06860) | :no_entry: | :white_check_mark: |
+| [Policy Regularization with Dataset Constraint (PRDC)](https://arxiv.org/abs/2306.06569) | :no_entry: | :white_check_mark: |
 | [Implicit Q-Learning (IQL)](https://arxiv.org/abs/2110.06169) | :no_entry: | :white_check_mark: |
 | [Calibrated Q-Learning (Cal-QL)](https://arxiv.org/abs/2303.05479) | :no_entry: | :white_check_mark: |
 | [ReBRAC](https://arxiv.org/abs/2305.09836) | :no_entry: | :white_check_mark: |

d3rlpy/algos/qlearning/__init__.py

@@ -12,6 +12,7 @@
 from .iql import *
 from .nfq import *
 from .plas import *
+from .prdc import *
 from .random_policy import *
 from .rebrac import *
 from .sac import *

d3rlpy/algos/qlearning/prdc.py

@@ -0,0 +1,259 @@
+import dataclasses
+from typing import Callable, Optional
+
+import numpy as np
+import torch
+from sklearn.neighbors import NearestNeighbors
+from typing_extensions import Self
+
+from ...base import DeviceArg, LearnableConfig, register_learnable
+from ...constants import ActionSpace, LoggingStrategy
+from ...dataset import ReplayBufferBase
+from ...logging import FileAdapterFactory, LoggerAdapterFactory
+from ...metrics import EvaluatorProtocol
+from ...models.builders import (
+    create_continuous_q_function,
+    create_deterministic_policy,
+)
+from ...models.encoders import EncoderFactory, make_encoder_field
+from ...models.q_functions import QFunctionFactory, make_q_func_field
+from ...optimizers.optimizers import OptimizerFactory, make_optimizer_field
+from ...types import Shape
+from ..utility import build_scalers_with_transition_picker
+from .base import QLearningAlgoBase
+from .torch.ddpg_impl import DDPGModules
+from .torch.prdc_impl import PRDCImpl
+
+__all__ = ["PRDCConfig", "PRDC"]
+
+
+@dataclasses.dataclass()
+class PRDCConfig(LearnableConfig):
+    r"""Config of PRDC algorithm.
+
+    PRDC is an simple offline RL algorithm built on top of TD3.
+    PRDC introduces Dataset Constraint (DC)-reguralized policy objective function.
+
+    .. math::
+
+        J(\phi) = \mathbb{E}_{s \sim D}
+            [\lambda Q(s, \pi(s)) - d^\beta_D(s, \pi(s))]
+
+    where
+
+    .. math::
+
+        \lambda = \frac{\alpha}{\frac{1}{N} \sum_(s_i, a_i) |Q(s_i, a_i)|}
+
+    and `d^\beta_\mathcal{D}(s,\pi(s))` is the DC loss, defined as
+
+    .. math::
+
+        d^\beta_\mathcal{D}(s,\pi(s)) = \min_{\hat{s}, \hat{a} \sim D}
+            [\| (\beta s) \oplus \pi(s) - (\beta \hat{s}) \oplus \hat{a} \|]
+
+    References:
+        * `Ran et al., Policy Regularization with Dataset Constraint for Offline Reinforcement Learning
+          Learning. <https://arxiv.org/abs/2306.06569>`_
+
+    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.optimizers.OptimizerFactory):
+            Optimizer factory for the actor.
+        critic_optim_factory (d3rlpy.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.
+        beta (float): :math:`\beta` value.
+        update_actor_interval (int): Interval to update policy function
+            described as `delayed policy update` in the paper.
+        compile_graph (bool): Flag to enable JIT compilation and CUDAGraph.
+    """
+
+    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
+    beta: float = 2.0
+    update_actor_interval: int = 2
+
+    def create(self, device: DeviceArg = False, enable_ddp: bool = False) -> "PRDC":
+        return PRDC(self, device, enable_ddp)
+
+    @staticmethod
+    def get_type() -> str:
+        return "prdc"
+
+
+class PRDC(QLearningAlgoBase[PRDCImpl, PRDCConfig]):
+    _nbsr = NearestNeighbors(n_neighbors=1, algorithm="auto", n_jobs=-1)
+
+    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,
+            enable_ddp=self._enable_ddp,
+        )
+        targ_policy = create_deterministic_policy(
+            observation_shape,
+            action_size,
+            self._config.actor_encoder_factory,
+            device=self._device,
+            enable_ddp=self._enable_ddp,
+        )
+        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,
+            enable_ddp=self._enable_ddp,
+        )
+        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,
+            enable_ddp=self._enable_ddp,
+        )
+
+        actor_optim = self._config.actor_optim_factory.create(
+            policy.named_modules(),
+            lr=self._config.actor_learning_rate,
+            compiled=self.compiled,
+        )
+        critic_optim = self._config.critic_optim_factory.create(
+            q_funcs.named_modules(),
+            lr=self._config.critic_learning_rate,
+            compiled=self.compiled,
+        )
+
+        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 = PRDCImpl(
+            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,
+            beta=self._config.beta,
+            update_actor_interval=self._config.update_actor_interval,
+            compiled=self.compiled,
+            nbsr=self._nbsr,
+            device=self._device,
+        )
+
+    def fit(
+        self,
+        dataset: ReplayBufferBase,
+        n_steps: int,
+        n_steps_per_epoch: int = 10000,
+        experiment_name: Optional[str] = None,
+        with_timestamp: bool = True,
+        logging_steps: int = 500,
+        logging_strategy: LoggingStrategy = LoggingStrategy.EPOCH,
+        logger_adapter: LoggerAdapterFactory = FileAdapterFactory(),
+        show_progress: bool = True,
+        save_interval: int = 1,
+        evaluators: Optional[dict[str, EvaluatorProtocol]] = None,
+        callback: Optional[Callable[[Self, int, int], None]] = None,
+        epoch_callback: Optional[Callable[[Self, int, int], None]] = None,
+    ) -> list[tuple[int, dict[str, float]]]:
+        observations = []
+        actions = []
+        for episode in dataset.buffer.episodes:
+            for i in range(episode.transition_count):
+                transition = dataset.transition_picker(episode, i)
+                observations.append(np.reshape(transition.observation, (1, -1)))
+                actions.append(np.reshape(transition.action, (1, -1)))
+        observations = np.concatenate(observations, axis=0)
+        actions = np.concatenate(actions, axis=0)
+
+        build_scalers_with_transition_picker(self, dataset)
+        if self.observation_scaler and self.observation_scaler.built:
+            observations = (
+                self.observation_scaler.transform(
+                    torch.tensor(observations, device=self._device)
+                )
+                .cpu()
+                .numpy()
+            )
+
+        if self.action_scaler and self.action_scaler.built:
+            actions = (
+                self.action_scaler.transform(torch.tensor(actions, device=self._device))
+                .cpu()
+                .numpy()
+            )
+
+        self._nbsr.fit(
+            np.concatenate(
+                [np.multiply(observations, self._config.beta), actions],
+                axis=1,
+            )
+        )
+
+        return super().fit(
+            dataset=dataset,
+            n_steps=n_steps,
+            n_steps_per_epoch=n_steps_per_epoch,
+            logging_steps=logging_steps,
+            logging_strategy=logging_strategy,
+            experiment_name=experiment_name,
+            with_timestamp=with_timestamp,
+            logger_adapter=logger_adapter,
+            show_progress=show_progress,
+            save_interval=save_interval,
+            evaluators=evaluators,
+            callback=callback,
+            epoch_callback=epoch_callback,
+        )
+
+    def get_action_type(self) -> ActionSpace:
+        return ActionSpace.CONTINUOUS
+
+
+register_learnable(PRDCConfig)

d3rlpy/algos/qlearning/torch/__init__.py

@@ -9,6 +9,7 @@
 from .dqn_impl import *
 from .iql_impl import *
 from .plas_impl import *
+from .prdc_impl import *
 from .rebrac_impl import *
 from .sac_impl import *
 from .td3_impl import *

d3rlpy/algos/qlearning/torch/prdc_impl.py

@@ -0,0 +1,84 @@
+# pylint: disable=too-many-ancestors
+import dataclasses
+
+import torch
+from sklearn.neighbors import NearestNeighbors
+
+from ....models.torch import ActionOutput, ContinuousEnsembleQFunctionForwarder
+from ....torch_utility import TorchMiniBatch
+from ....types import Shape
+from .ddpg_impl import DDPGBaseActorLoss, DDPGModules
+from .td3_impl import TD3Impl
+
+__all__ = ["PRDCImpl"]
+
+
+@dataclasses.dataclass(frozen=True)
+class PRDCActorLoss(DDPGBaseActorLoss):
+    dc_loss: torch.Tensor
+
+
+class PRDCImpl(TD3Impl):
+    _alpha: float
+    _beta: float
+    _nbsr: NearestNeighbors
+
+    def __init__(
+        self,
+        observation_shape: Shape,
+        action_size: int,
+        modules: DDPGModules,
+        q_func_forwarder: ContinuousEnsembleQFunctionForwarder,
+        targ_q_func_forwarder: ContinuousEnsembleQFunctionForwarder,
+        gamma: float,
+        tau: float,
+        target_smoothing_sigma: float,
+        target_smoothing_clip: float,
+        alpha: float,
+        beta: float,
+        update_actor_interval: int,
+        compiled: bool,
+        nbsr: NearestNeighbors,
+        device: str,
+    ):
+        super().__init__(
+            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=gamma,
+            tau=tau,
+            target_smoothing_sigma=target_smoothing_sigma,
+            target_smoothing_clip=target_smoothing_clip,
+            update_actor_interval=update_actor_interval,
+            compiled=compiled,
+            device=device,
+        )
+        self._alpha = alpha
+        self._beta = beta
+        self._nbsr = nbsr
+
+    def compute_actor_loss(
+        self, batch: TorchMiniBatch, action: ActionOutput
+    ) -> PRDCActorLoss:
+        q_t = self._q_func_forwarder.compute_expected_q(
+            batch.observations, action.squashed_mu, "none"
+        )[0]
+        lam = self._alpha / (q_t.abs().mean()).detach()
+        key = (
+            torch.cat(
+                [torch.mul(batch.observations, self._beta), action.squashed_mu], dim=-1
+            )
+            .detach()
+            .cpu()
+            .numpy()
+        )
+        idx = self._nbsr.kneighbors(key, n_neighbors=1, return_distance=False)
+        nearest_neightbour = torch.tensor(
+            self._nbsr._fit_X[idx][:, :, -self.action_size :],
+            device=self.device,
+            dtype=action.squashed_mu.dtype,
+        ).squeeze(dim=1)
+        dc_loss = torch.nn.functional.mse_loss(action.squashed_mu, nearest_neightbour)
+        return PRDCActorLoss(actor_loss=lam * -q_t.mean() + dc_loss, dc_loss=dc_loss)

mypy.ini

@@ -71,3 +71,6 @@ follow_imports_for_stubs = True
 ignore_missing_imports = True
 follow_imports = skip
 follow_imports_for_stubs = True
+
+[mypy-sklearn.*]
+ignore_missing_imports = True