ray-project--ray/rllib/algorithms/ppo/ppo_learner.py

import abc
from typing import Any, Dict

from ray.rllib.algorithms.ppo.ppo import (
    LEARNER_RESULTS_CURR_ENTROPY_COEFF_KEY,
    LEARNER_RESULTS_KL_KEY,
    PPOConfig,
)
from ray.rllib.connectors.learner import (
    AddOneTsToEpisodesAndTruncate,
    GeneralAdvantageEstimation,
)
from ray.rllib.core.learner.learner import Learner
from ray.rllib.core.rl_module.apis.value_function_api import ValueFunctionAPI
from ray.rllib.utils.annotations import (
    OverrideToImplementCustomLogic_CallToSuperRecommended,
    override,
)
from ray.rllib.utils.lambda_defaultdict import LambdaDefaultDict
from ray.rllib.utils.metrics import (
    NUM_ENV_STEPS_SAMPLED_LIFETIME,
)
from ray.rllib.utils.numpy import convert_to_numpy
from ray.rllib.utils.schedules.scheduler import Scheduler
from ray.rllib.utils.typing import ModuleID, TensorType


class PPOLearner(Learner):
    @override(Learner)
    def build(self) -> None:
        super().build()

        # Dict mapping module IDs to the respective entropy Scheduler instance.
        self.entropy_coeff_schedulers_per_module: Dict[
            ModuleID, Scheduler
        ] = LambdaDefaultDict(
            lambda module_id: Scheduler(
                fixed_value_or_schedule=(
                    self.config.get_config_for_module(module_id).entropy_coeff
                ),
                framework=self.framework,
                device=self._device,
            )
        )

        # Set up KL coefficient variables (per module).
        # Note that the KL coeff is not controlled by a Scheduler, but seeks
        # to stay close to a given kl_target value.
        self.curr_kl_coeffs_per_module: Dict[ModuleID, TensorType] = LambdaDefaultDict(
            lambda module_id: self._get_tensor_variable(
                self.config.get_config_for_module(module_id).kl_coeff
            )
        )

        # Extend all episodes by one artificial timestep to allow the value function net
        # to compute the bootstrap values (and add a mask to the batch to know, which
        # slots to mask out).
        if (
            self._learner_connector is not None
            and self.config.add_default_connectors_to_learner_pipeline
        ):
            # Before anything, add one ts to each episode (and record this in the loss
            # mask, so that the computations at this extra ts are not used to compute
            # the loss).
            self._learner_connector.prepend(AddOneTsToEpisodesAndTruncate())
            # At the end of the pipeline (when the batch is already completed), add the
            # GAE connector, which performs a vf forward pass, then computes the GAE
            # computations, and puts the results of this (advantages, value targets)
            # directly back in the batch. This is then the batch used for
            # `forward_train` and `compute_losses`.
            self._learner_connector.append(
                GeneralAdvantageEstimation(
                    gamma=self.config.gamma, lambda_=self.config.lambda_
                )
            )

    @override(Learner)
    def remove_module(self, module_id: ModuleID, **kwargs):
        marl_spec = super().remove_module(module_id, **kwargs)

        self.entropy_coeff_schedulers_per_module.pop(module_id, None)
        self.curr_kl_coeffs_per_module.pop(module_id, None)

        return marl_spec

    @OverrideToImplementCustomLogic_CallToSuperRecommended
    @override(Learner)
    def after_gradient_based_update(
        self,
        *,
        timesteps: Dict[str, Any],
    ) -> None:
        super().after_gradient_based_update(timesteps=timesteps)

        for module_id, module in self.module._rl_modules.items():
            config = self.config.get_config_for_module(module_id)

            # Update entropy coefficient via our Scheduler.
            new_entropy_coeff = self.entropy_coeff_schedulers_per_module[
                module_id
            ].update(timestep=timesteps.get(NUM_ENV_STEPS_SAMPLED_LIFETIME, 0))
            self.metrics.log_value(
                (module_id, LEARNER_RESULTS_CURR_ENTROPY_COEFF_KEY),
                new_entropy_coeff,
                window=1,
            )
            if (
                config.use_kl_loss
                and (module_id, LEARNER_RESULTS_KL_KEY) in self.metrics
            ):
                kl_loss = convert_to_numpy(
                    self.metrics.peek((module_id, LEARNER_RESULTS_KL_KEY))
                )
                self._update_module_kl_coeff(
                    module_id=module_id,
                    config=config,
                    kl_loss=kl_loss,
                )

    @classmethod
    @override(Learner)
    def rl_module_required_apis(cls) -> list[type]:
        # In order for a PPOLearner to update an RLModule, it must implement the
        # following APIs:
        return [ValueFunctionAPI]

    @abc.abstractmethod
    def _update_module_kl_coeff(
        self,
        *,
        module_id: ModuleID,
        config: PPOConfig,
        kl_loss: float,
    ) -> None:
        """Dynamically update the KL loss coefficients of each module.

        The update is completed using the mean KL divergence between the action
        distributions current policy and old policy of each module. That action
        distribution is computed during the most recent update/call to `compute_loss`.

        Args:
            module_id: The module whose KL loss coefficient to update.
            config: The AlgorithmConfig specific to the given `module_id`.
            kl_loss: The mean KL loss of the module, computed inside
                `compute_loss_for_module()`.
        """