Source code for optuna.samplers._gp.sampler

from __future__ import annotations

import json
from typing import Any
from typing import TYPE_CHECKING

import numpy as np

import optuna
from optuna._experimental import experimental_class
from optuna._experimental import warn_experimental_argument
from optuna.samplers._base import _CONSTRAINTS_KEY
from optuna.samplers._base import _INDEPENDENT_SAMPLING_WARNING_TEMPLATE
from optuna.samplers._base import _process_constraints_after_trial
from optuna.samplers._base import BaseSampler
from optuna.samplers._lazy_random_state import LazyRandomState
from optuna.study import StudyDirection
from optuna.study._multi_objective import _is_pareto_front
from optuna.trial import FrozenTrial
from optuna.trial import TrialState


if TYPE_CHECKING:
    from collections.abc import Callable
    from collections.abc import Sequence

    import torch

    import optuna._gp.acqf as acqf_module
    import optuna._gp.gp as gp
    import optuna._gp.optim_mixed as optim_mixed
    import optuna._gp.prior as prior
    import optuna._gp.search_space as gp_search_space
    from optuna.distributions import BaseDistribution
    from optuna.study import Study
else:
    from optuna._imports import _LazyImport

    torch = _LazyImport("torch")
    gp_search_space = _LazyImport("optuna._gp.search_space")
    gp = _LazyImport("optuna._gp.gp")
    optim_mixed = _LazyImport("optuna._gp.optim_mixed")
    acqf_module = _LazyImport("optuna._gp.acqf")
    prior = _LazyImport("optuna._gp.prior")

import logging


_logger = logging.getLogger(__name__)

EPS = 1e-10

_RELATIVE_PARAMS_KEY = "gp:relative_params"
# The value of system_attrs must be less than 2046 characters on RDBStorage.
_SYSTEM_ATTR_MAX_LENGTH = 2045


def _standardize_values(values: np.ndarray) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
    clipped_values = gp.warn_and_convert_inf(values)
    means = np.mean(clipped_values, axis=0)
    stds = np.std(clipped_values, axis=0)
    standardized_values = (clipped_values - means) / np.maximum(EPS, stds)
    return standardized_values, means, stds


[docs] @experimental_class("3.6.0") class GPSampler(BaseSampler): """Sampler using Gaussian process-based Bayesian optimization. This sampler fits a Gaussian process (GP) to the objective function and optimizes the acquisition function to suggest the next parameters. The current implementation uses Matern kernel with nu=2.5 (twice differentiable) with automatic relevance determination (ARD) for the length scale of each parameter. The hyperparameters of the kernel are obtained by maximizing the marginal log-likelihood of the hyperparameters given the past trials. To prevent overfitting, Gamma prior is introduced for kernel scale and noise variance and a hand-crafted prior is introduced for inverse squared lengthscales. As an acquisition function, we use: - log expected improvement (logEI) for single-objective optimization, - log expected hypervolume improvement (logEHVI) for Multi-objective optimization, and - the summation of logEI and the logarithm of the feasible probability with the independent assumption of each constraint for (black-box inequality) constrained optimization. For further information about these acquisition functions, please refer to the following papers: - `Unexpected Improvements to Expected Improvement for Bayesian Optimization <https://arxiv.org/abs/2310.20708>`__ - `Differentiable Expected Hypervolume Improvement for Parallel Multi-Objective Bayesian Optimization <https://arxiv.org/abs/2006.05078>`__ - `Bayesian Optimization with Inequality Constraints <https://proceedings.mlr.press/v32/gardner14.pdf>`__ Please also check our articles: - `[Optuna v4.5] Gaussian Process-Based Sampler (GPSampler) Can Now Perform Constrained Multi-Objective Optimization <https://medium.com/optuna/optuna-v4-5-81e78d8e077a>`__ - `[Optuna v4.2] Gaussian Process-Based Sampler (GPSampler) Can Now Handle Inequality Constraints <https://medium.com/optuna/optuna-v4-2-gaussian-process-based-sampler-can-now-handle-inequality-constraints-a4f68e8ee810>`__ - `Introducing Optuna's Native GPSampler <https://medium.com/optuna/introducing-optunas-native-gpsampler-0aa9aa3b4840>`__ The optimization of the acquisition function is performed via: 1. Collect the best param from the past trials, 2. Collect ``n_preliminary_samples`` points using Quasi-Monte Carlo (QMC) sampling, 3. Choose the best point from the collected points, 4. Choose ``n_local_search - 2`` points from the collected points using the roulette selection, 5. Perform a local search for each chosen point as an initial point, and 6. Return the point with the best acquisition function value as the next parameter. Decoupled optimizer update with a batched evaluation is employed to perform a batch of local searches simultaneously, specifically speeding up Step 5 above. The following paper details the methodology: - `Batch Acquisition Function Evaluations and Decouple Optimizer Updates for Faster Bayesian Optimization <https://arxiv.org/abs/2511.13625>`__ Note that the procedures for non single-objective optimization setups are slightly different from the single-objective version described above, but we omit the descriptions for the others for brevity. The local search iteratively optimizes the acquisition function by repeating: 1. Gradient ascent using l-BFGS-B for continuous parameters, and 2. Line search or exhaustive search for each discrete parameter independently. The local search is terminated if the routine stops updating the best parameter set or the maximum number of iterations is reached. We use line search instead of rounding the results from the continuous optimization since EI typically yields a high value between one grid and its adjacent grid. .. note:: This sampler requires ``scipy`` and ``torch``. You can install these dependencies with ``pip install scipy torch``. Args: seed: Random seed to initialize internal random number generator. Defaults to :obj:`None` (a seed is picked randomly). independent_sampler: Sampler used for initial sampling (for the first ``n_startup_trials`` trials) and for conditional parameters. Defaults to :obj:`None` (a random sampler with the same ``seed`` is used). n_startup_trials: Number of initial trials. Defaults to 10. deterministic_objective: Whether the objective function is deterministic or not. If :obj:`True`, the sampler will fix the noise variance of the surrogate model to the minimum value (slightly above 0 to ensure numerical stability). Defaults to :obj:`False`. Currently, all the objectives will be assume to be deterministic if :obj:`True`. constraints_func: An optional function that computes the objective constraints. It must take a :class:`~optuna.trial.FrozenTrial` and return the constraints. The return value must be a sequence of :obj:`float` s. A value strictly larger than 0 means that a constraints is violated. A value equal to or smaller than 0 is considered feasible. If ``constraints_func`` returns more than one value for a trial, that trial is considered feasible if and only if all values are equal to 0 or smaller. The ``constraints_func`` will be evaluated after each successful trial. The function won't be called when trials fail or are pruned, but this behavior is subject to change in future releases. warn_independent_sampling: If this is :obj:`True`, a warning message is emitted when the value of a parameter is sampled by using an independent sampler, meaning that no GP model is used in the sampling. Note that the parameters of the first trial in a study are always sampled via an independent sampler, so no warning messages are emitted in this case. """ def __init__( self, *, seed: int | None = None, independent_sampler: BaseSampler | None = None, n_startup_trials: int = 10, deterministic_objective: bool = False, constraints_func: Callable[[FrozenTrial], Sequence[float]] | None = None, warn_independent_sampling: bool = True, ) -> None: self._rng = LazyRandomState(seed) self._independent_sampler = independent_sampler or optuna.samplers.RandomSampler(seed=seed) self._intersection_search_space = optuna.search_space.IntersectionSearchSpace() self._n_startup_trials = n_startup_trials self._log_prior: Callable[[gp.GPRegressor], torch.Tensor] = prior.default_log_prior self._minimum_noise: float = prior.DEFAULT_MINIMUM_NOISE_VAR # We cache the kernel parameters for initial values of fitting the next time. # TODO(nabenabe): Make the cache lists system_attrs to make GPSampler stateless. self._gprs_cache_list: list[gp.GPRegressor] | None = None self._constraints_gprs_cache_list: list[gp.GPRegressor] | None = None self._deterministic = deterministic_objective self._constraints_func = constraints_func self._warn_independent_sampling = warn_independent_sampling if constraints_func is not None: warn_experimental_argument("constraints_func") # Control parameters of the acquisition function optimization. self._n_preliminary_samples: int = 2048 # NOTE(nabenabe): ehvi in BoTorchSampler uses 20. self._n_local_search = 10 self._tol = 1e-4 def _log_independent_sampling(self, trial: FrozenTrial, param_name: str) -> None: msg = _INDEPENDENT_SAMPLING_WARNING_TEMPLATE.format( param_name=param_name, trial_number=trial.number, independent_sampler_name=self._independent_sampler.__class__.__name__, sampler_name=self.__class__.__name__, fallback_reason="dynamic search space is not supported by GPSampler", ) _logger.warning(msg)
[docs] def reseed_rng(self) -> None: self._rng.rng.seed() self._independent_sampler.reseed_rng()
[docs] def infer_relative_search_space( self, study: Study, trial: FrozenTrial ) -> dict[str, BaseDistribution]: search_space = {} for name, distribution in self._intersection_search_space.calculate(study).items(): if distribution.single(): continue search_space[name] = distribution return search_space
def _optimize_acqf( self, acqf: acqf_module.BaseAcquisitionFunc, best_params: np.ndarray | None ) -> np.ndarray: # Advanced users can override this method to change the optimization algorithm. # However, we do not make any effort to keep backward compatibility between versions. # Particularly, we may remove this function in future refactoring. assert best_params is None or len(best_params.shape) == 2 normalized_params, _acqf_val = optim_mixed.optimize_acqf_mixed( acqf, warmstart_normalized_params_array=best_params, n_preliminary_samples=self._n_preliminary_samples, n_local_search=self._n_local_search, tol=self._tol, rng=self._rng.rng, ) return normalized_params def _get_constraints_acqf_args( self, constraint_vals: np.ndarray, internal_search_space: gp_search_space.SearchSpace, normalized_params: np.ndarray, ) -> tuple[list[gp.GPRegressor], list[float]]: # NOTE(nabenabe): Flip the sign of constraints since they are always to be minimized. standardized_constraint_vals, means, stds = _standardize_values(-constraint_vals) if ( self._gprs_cache_list is not None and len(self._gprs_cache_list[0].inverse_squared_lengthscales) != internal_search_space.dim ): # Clear cache if the search space changes. self._constraints_gprs_cache_list = None is_categorical = internal_search_space.is_categorical constraints_gprs = [] constraints_threshold_list = [] constraints_threshold_list = (-means / np.maximum(EPS, stds)).tolist() for i, vals in enumerate(standardized_constraint_vals.T): cache = ( self._constraints_gprs_cache_list[i] if self._constraints_gprs_cache_list is not None else None ) gpr = gp.fit_kernel_params( X=normalized_params, Y=vals, is_categorical=is_categorical, log_prior=self._log_prior, minimum_noise=self._minimum_noise, gpr_cache=cache, deterministic_objective=self._deterministic, ) constraints_gprs.append(gpr) self._constraints_gprs_cache_list = constraints_gprs return constraints_gprs, constraints_threshold_list def _get_best_params_for_multi_objective( self, normalized_params: np.ndarray, standardized_score_vals: np.ndarray, ) -> np.ndarray: pareto_params = normalized_params[ _is_pareto_front(-standardized_score_vals, assume_unique_lexsorted=False) ] n_pareto_sols = len(pareto_params) # TODO(nabenabe): Verify the validity of this choice. size = min(self._n_local_search // 2, n_pareto_sols) chosen_indices = self._rng.rng.choice(n_pareto_sols, size=size, replace=False) return pareto_params[chosen_indices] def _get_normalized_params_of_running_trials( self, trials: list[FrozenTrial], internal_search_space: gp_search_space.SearchSpace, ) -> np.ndarray | None: running_trials = [ t for t in trials if internal_search_space._optuna_search_space.keys() <= _get_params(t).keys() and t.state == TrialState.RUNNING ] if len(running_trials) == 0: return None return internal_search_space.get_normalized_params( running_trials, [_get_params(t) for t in running_trials] )
[docs] def sample_relative( self, study: Study, trial: FrozenTrial, search_space: dict[str, BaseDistribution] ) -> dict[str, Any]: if search_space == {}: return {} states = (TrialState.COMPLETE, TrialState.RUNNING) # At present, running trials are taken into account only in single-objective # unconstrained optimization. use_cache = len(study.directions) > 1 or self._constraints_func is not None trials = study._get_trials(deepcopy=False, states=states, use_cache=use_cache) completed_trials = [t for t in trials if t.state == TrialState.COMPLETE] if len(completed_trials) < self._n_startup_trials: return {} # Force CPU device for all torch operations to avoid issues when # torch.set_default_device("cuda") is set globally (issue #6113). with torch.device("cpu"): params = self._sample_relative_impl(study, completed_trials, trials, search_space) if params != {}: # Share the params obtained by the relative sampling with the other processes. params_str = json.dumps(params) for i in range(0, len(params_str), _SYSTEM_ATTR_MAX_LENGTH): study._storage.set_trial_system_attr( trial._trial_id, f"{_RELATIVE_PARAMS_KEY}:{i // _SYSTEM_ATTR_MAX_LENGTH}", params_str[i : i + _SYSTEM_ATTR_MAX_LENGTH], ) return params
def _sample_relative_impl( self, study: Study, completed_trials: list[FrozenTrial], trials: list[FrozenTrial], search_space: dict[str, BaseDistribution], ) -> dict[str, Any]: internal_search_space = gp_search_space.SearchSpace(search_space) normalized_params = internal_search_space.get_normalized_params(completed_trials) _sign = np.array([-1.0 if d == StudyDirection.MINIMIZE else 1.0 for d in study.directions]) standardized_score_vals, _, _ = _standardize_values( _sign * np.array([trial.values for trial in completed_trials]) ) if ( self._gprs_cache_list is not None and len(self._gprs_cache_list[0].inverse_squared_lengthscales) != internal_search_space.dim ): # Clear cache if the search space changes. self._gprs_cache_list = None gprs_list = [] n_objectives = standardized_score_vals.shape[-1] is_categorical = internal_search_space.is_categorical for i in range(n_objectives): cache = self._gprs_cache_list[i] if self._gprs_cache_list is not None else None gprs_list.append( gp.fit_kernel_params( X=normalized_params, Y=standardized_score_vals[:, i], is_categorical=is_categorical, log_prior=self._log_prior, minimum_noise=self._minimum_noise, gpr_cache=cache, deterministic_objective=self._deterministic, ) ) self._gprs_cache_list = gprs_list best_params: np.ndarray | None acqf: acqf_module.BaseAcquisitionFunc if self._constraints_func is None: if n_objectives == 1: assert len(gprs_list) == 1 acqf = acqf_module.LogEI( gpr=gprs_list[0], search_space=internal_search_space, threshold=standardized_score_vals[:, 0].max(), normalized_params_of_running_trials=( self._get_normalized_params_of_running_trials( trials, internal_search_space ) ), ) best_params = normalized_params[np.argmax(standardized_score_vals), np.newaxis] else: acqf = acqf_module.LogEHVI( gpr_list=gprs_list, search_space=internal_search_space, Y_train=torch.from_numpy(standardized_score_vals), n_qmc_samples=128, # NOTE(nabenabe): The BoTorch default value. qmc_seed=self._rng.rng.randint(1 << 30), ) best_params = self._get_best_params_for_multi_objective( normalized_params, standardized_score_vals ) else: if n_objectives == 1: assert len(gprs_list) == 1 constraint_vals, is_feasible = _get_constraint_vals_and_feasibility( study, completed_trials ) y_with_neginf = np.where(is_feasible, standardized_score_vals[:, 0], -np.inf) # TODO(kAIto47802): If all trials are infeasible, the acquisition function # for the objective function can be ignored, so skipping the computation # of gpr can speed up. # TODO(kAIto47802): Consider the case where all trials are feasible. # We can ignore constraints in this case. constr_gpr_list, constr_threshold_list = self._get_constraints_acqf_args( constraint_vals, internal_search_space, normalized_params ) i_opt = np.argmax(y_with_neginf) best_feasible_y = y_with_neginf[i_opt] acqf = acqf_module.ConstrainedLogEI( gpr=gprs_list[0], search_space=internal_search_space, threshold=best_feasible_y, constraints_gpr_list=constr_gpr_list, constraints_threshold_list=constr_threshold_list, ) assert normalized_params.shape[:-1] == y_with_neginf.shape best_params = ( None if np.isneginf(best_feasible_y) else normalized_params[i_opt, np.newaxis] ) else: constraint_vals, is_feasible = _get_constraint_vals_and_feasibility( study, completed_trials ) constr_gpr_list, constr_threshold_list = self._get_constraints_acqf_args( constraint_vals, internal_search_space, normalized_params ) is_all_infeasible = not any(is_feasible) acqf = acqf_module.ConstrainedLogEHVI( gpr_list=gprs_list, search_space=internal_search_space, Y_feasible=( torch.from_numpy(standardized_score_vals[is_feasible]) if not is_all_infeasible else None ), n_qmc_samples=128, # NOTE(nabenabe): The BoTorch default value. qmc_seed=self._rng.rng.randint(1 << 30), constraints_gpr_list=constr_gpr_list, constraints_threshold_list=constr_threshold_list, ) best_params = ( self._get_best_params_for_multi_objective( normalized_params[is_feasible], standardized_score_vals[is_feasible], ) if not is_all_infeasible else None ) normalized_param = self._optimize_acqf(acqf, best_params) return internal_search_space.get_unnormalized_param(normalized_param)
[docs] def sample_independent( self, study: Study, trial: FrozenTrial, param_name: str, param_distribution: BaseDistribution, ) -> Any: if self._warn_independent_sampling: states = (TrialState.COMPLETE,) completed_trials = study._get_trials(deepcopy=False, states=states, use_cache=True) if len(completed_trials) >= self._n_startup_trials: self._log_independent_sampling(trial, param_name) return self._independent_sampler.sample_independent( study, trial, param_name, param_distribution )
[docs] def before_trial(self, study: Study, trial: FrozenTrial) -> None: self._independent_sampler.before_trial(study, trial)
[docs] def after_trial( self, study: Study, trial: FrozenTrial, state: TrialState, values: Sequence[float] | None, ) -> None: if self._constraints_func is not None: _process_constraints_after_trial(self._constraints_func, study, trial, state) self._independent_sampler.after_trial(study, trial, state, values)
def _get_constraint_vals_and_feasibility( study: Study, trials: list[FrozenTrial] ) -> tuple[np.ndarray, np.ndarray]: _constraint_vals = [ study._storage.get_trial_system_attrs(trial._trial_id).get(_CONSTRAINTS_KEY, ()) for trial in trials ] if any(len(_constraint_vals[0]) != len(c) for c in _constraint_vals): raise ValueError("The number of constraints must be the same for all trials.") constraint_vals = np.array(_constraint_vals) assert len(constraint_vals.shape) == 2, "constraint_vals must be a 2d array." is_feasible = np.all(constraint_vals <= 0, axis=1) assert not isinstance(is_feasible, np.bool_), "MyPy Redefinition for NumPy v2.2.0." return constraint_vals, is_feasible def _get_params(trial: FrozenTrial) -> dict[str, Any]: if trial.state.is_finished(): return trial.params params_strs = [] i = 0 while params_str_i := trial.system_attrs.get(f"{_RELATIVE_PARAMS_KEY}:{i}"): params_strs.append(params_str_i) i += 1 if len(params_strs) == 0: return trial.params params = json.loads("".join(params_strs)) params.update(trial.params) return params