Source code for ax.models.torch.botorch_defaults

#!/usr/bin/env python3
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

from typing import Any, Callable, Dict, List, Optional, Tuple

import torch
from ax.core.types import TConfig
from ax.models.model_utils import best_observed_point, get_observed
from ax.models.torch.utils import (  # noqa F401
    _to_inequality_constraints,
    predict_from_model,
)
from ax.models.torch_base import TorchModel
from botorch.acquisition.acquisition import AcquisitionFunction
from botorch.acquisition.fixed_feature import FixedFeatureAcquisitionFunction
from botorch.acquisition.objective import ConstrainedMCObjective, GenericMCObjective
from botorch.acquisition.utils import get_acquisition_function, get_infeasible_cost
from botorch.exceptions.errors import UnsupportedError
from botorch.fit import fit_gpytorch_model
from botorch.models.gp_regression import FixedNoiseGP, SingleTaskGP
from botorch.models.gp_regression_fidelity import SingleTaskMultiFidelityGP
from botorch.models.gpytorch import GPyTorchModel
from botorch.models.model import Model
from botorch.models.model_list_gp_regression import ModelListGP
from botorch.models.multitask import FixedNoiseMultiTaskGP, MultiTaskGP
from botorch.optim.optimize import optimize_acqf
from botorch.utils import (
    get_objective_weights_transform,
    get_outcome_constraint_transforms,
)
from botorch.utils.multi_objective.scalarization import get_chebyshev_scalarization
from gpytorch.mlls.exact_marginal_log_likelihood import ExactMarginalLogLikelihood
from gpytorch.mlls.sum_marginal_log_likelihood import SumMarginalLogLikelihood
from gpytorch.priors.lkj_prior import LKJCovariancePrior
from gpytorch.priors.torch_priors import GammaPrior
from torch import Tensor


MIN_OBSERVED_NOISE_LEVEL = 1e-7


[docs]def get_and_fit_model( Xs: List[Tensor], Ys: List[Tensor], Yvars: List[Tensor], task_features: List[int], fidelity_features: List[int], metric_names: List[str], state_dict: Optional[Dict[str, Tensor]] = None, refit_model: bool = True, **kwargs: Any, ) -> GPyTorchModel: r"""Instantiates and fits a botorch GPyTorchModel using the given data. N.B. Currently, the logic for choosing ModelListGP vs other models is handled using if-else statements in lines 96-137. In the future, this logic should be taken care of by modular botorch. Args: Xs: List of X data, one tensor per outcome. Ys: List of Y data, one tensor per outcome. Yvars: List of observed variance of Ys. task_features: List of columns of X that are tasks. fidelity_features: List of columns of X that are fidelity parameters. metric_names: Names of each outcome Y in Ys. state_dict: If provided, will set model parameters to this state dictionary. Otherwise, will fit the model. refit_model: Flag for refitting model. Returns: A fitted GPyTorchModel. """ if len(fidelity_features) > 0 and len(task_features) > 0: raise NotImplementedError( "Currently do not support MF-GP models with task_features!" ) if len(fidelity_features) > 1: raise NotImplementedError( "Fidelity MF-GP models currently support only a single fidelity parameter!" ) if len(task_features) > 1: raise NotImplementedError( f"This model only supports 1 task feature (got {task_features})" ) elif len(task_features) == 1: task_feature = task_features[0] else: task_feature = None model = None # TODO: Better logic for deciding when to use a ModelListGP. Currently the # logic is unclear. The two cases in which ModelListGP is used are # (i) the training inputs (Xs) are not the same for the different outcomes, and # (ii) a multi-task model is used if task_feature is None: if len(Xs) == 1: # Use single output, single task GP model = _get_model( X=Xs[0], Y=Ys[0], Yvar=Yvars[0], task_feature=task_feature, fidelity_features=fidelity_features, **kwargs, ) elif all(torch.equal(Xs[0], X) for X in Xs[1:]): # Use batched multioutput, single task GP Y = torch.cat(Ys, dim=-1) Yvar = torch.cat(Yvars, dim=-1) model = _get_model( X=Xs[0], Y=Y, Yvar=Yvar, task_feature=task_feature, fidelity_features=fidelity_features, **kwargs, ) # TODO: Is this equivalent an "else:" here? if model is None: if task_feature is None: models = [ _get_model(X=X, Y=Y, Yvar=Yvar, **kwargs) for X, Y, Yvar in zip(Xs, Ys, Yvars) ] else: # use multi-task GP mtgp_rank_dict = kwargs.pop("multitask_gp_ranks", {}) # assembles list of ranks associated with each metric if len({len(Xs), len(Ys), len(Yvars), len(metric_names)}) > 1: raise ValueError( "Lengths of Xs, Ys, Yvars, and metric_names must match. Your " f"inputs have lengths {len(Xs)}, {len(Ys)}, {len(Yvars)}, and " f"{len(metric_names)}, respectively." ) mtgp_rank_list = [ mtgp_rank_dict.get(metric, None) for metric in metric_names ] models = [ _get_model( X=X, Y=Y, Yvar=Yvar, task_feature=task_feature, rank=mtgp_rank, **kwargs, ) for X, Y, Yvar, mtgp_rank in zip(Xs, Ys, Yvars, mtgp_rank_list) ] model = ModelListGP(*models) model.to(Xs[0]) if state_dict is not None: model.load_state_dict(state_dict) if state_dict is None or refit_model: # TODO: Add bounds for optimization stability - requires revamp upstream bounds = {} if isinstance(model, ModelListGP): mll = SumMarginalLogLikelihood(model.likelihood, model) else: # pyre-ignore: [16] mll = ExactMarginalLogLikelihood(model.likelihood, model) mll = fit_gpytorch_model(mll, bounds=bounds) return model
[docs]def get_NEI( model: Model, objective_weights: Tensor, outcome_constraints: Optional[Tuple[Tensor, Tensor]] = None, X_observed: Optional[Tensor] = None, X_pending: Optional[Tensor] = None, **kwargs: Any, ) -> AcquisitionFunction: r"""Instantiates a qNoisyExpectedImprovement acquisition function. Args: model: The underlying model which the acqusition function uses to estimate acquisition values of candidates. objective_weights: The objective is to maximize a weighted sum of the columns of f(x). These are the weights. outcome_constraints: A tuple of (A, b). For k outcome constraints and m outputs at f(x), A is (k x m) and b is (k x 1) such that A f(x) <= b. (Not used by single task models) X_observed: A tensor containing points observed for all objective outcomes and outcomes that appear in the outcome constraints (if there are any). X_pending: A tensor containing points whose evaluation is pending (i.e. that have been submitted for evaluation) present for all objective outcomes and outcomes that appear in the outcome constraints (if there are any). mc_samples: The number of MC samples to use (default: 512). qmc: If True, use qMC instead of MC (default: True). prune_baseline: If True, prune the baseline points for NEI (default: True). chebyshev_scalarization: Use augmented Chebyshev scalarization. Returns: qNoisyExpectedImprovement: The instantiated acquisition function. """ if X_observed is None: raise ValueError("There are no feasible observed points.") # construct Objective module if kwargs.get("chebyshev_scalarization", False): if "Ys" not in kwargs: raise ValueError("Chebyshev Scalarization requires Ys argument") Y_tensor = torch.cat(kwargs.get("Ys"), dim=-1) obj_tf = get_chebyshev_scalarization(weights=objective_weights, Y=Y_tensor) else: obj_tf = get_objective_weights_transform(objective_weights) if outcome_constraints is None: objective = GenericMCObjective(objective=obj_tf) else: con_tfs = get_outcome_constraint_transforms(outcome_constraints) inf_cost = get_infeasible_cost(X=X_observed, model=model, objective=obj_tf) objective = ConstrainedMCObjective( objective=obj_tf, constraints=con_tfs or [], infeasible_cost=inf_cost ) return get_acquisition_function( acquisition_function_name="qNEI", model=model, objective=objective, X_observed=X_observed, X_pending=X_pending, prune_baseline=kwargs.get("prune_baseline", True), mc_samples=kwargs.get("mc_samples", 512), qmc=kwargs.get("qmc", True), # pyre-fixme[6]: Expected `Optional[int]` for 9th param but got # `Union[float, int]`. seed=torch.randint(1, 10000, (1,)).item(), )
[docs]def scipy_optimizer( acq_function: AcquisitionFunction, bounds: Tensor, n: int, inequality_constraints: Optional[List[Tuple[Tensor, Tensor, float]]] = None, fixed_features: Optional[Dict[int, float]] = None, rounding_func: Optional[Callable[[Tensor], Tensor]] = None, **kwargs: Any, ) -> Tuple[Tensor, Tensor]: r"""Optimizer using scipy's minimize module on a numpy-adpator. Args: acq_function: A botorch AcquisitionFunction. bounds: A `2 x d`-dim tensor, where `bounds[0]` (`bounds[1]`) are the lower (upper) bounds of the feasible hyperrectangle. n: The number of candidates to generate. inequality constraints: A list of tuples (indices, coefficients, rhs), with each tuple encoding an inequality constraint of the form `\sum_i (X[indices[i]] * coefficients[i]) >= rhs` fixed_features: A map {feature_index: value} for features that should be fixed to a particular value during generation. rounding_func: A function that rounds an optimization result appropriately (i.e., according to `round-trip` transformations). Returns: 2-element tuple containing - A `n x d`-dim tensor of generated candidates. - In the case of joint optimization, a scalar tensor containing the joint acquisition value of the `n` points. In the case of sequential optimization, a `n`-dim tensor of conditional acquisition values, where `i`-th element is the expected acquisition value conditional on having observed candidates `0,1,...,i-1`. """ num_restarts: int = kwargs.get("num_restarts", 20) raw_samples: int = kwargs.get("num_raw_samples", 50 * num_restarts) if kwargs.get("joint_optimization", False): sequential = False else: sequential = True # use SLSQP by default for small problems since it yields faster wall times if "method" not in kwargs: kwargs["method"] = "SLSQP" X, expected_acquisition_value = optimize_acqf( acq_function=acq_function, bounds=bounds, q=n, num_restarts=num_restarts, raw_samples=raw_samples, options=kwargs, inequality_constraints=inequality_constraints, fixed_features=fixed_features, sequential=sequential, post_processing_func=rounding_func, ) return X, expected_acquisition_value
[docs]def recommend_best_observed_point( model: TorchModel, bounds: List[Tuple[float, float]], objective_weights: Tensor, outcome_constraints: Optional[Tuple[Tensor, Tensor]] = None, linear_constraints: Optional[Tuple[Tensor, Tensor]] = None, fixed_features: Optional[Dict[int, float]] = None, model_gen_options: Optional[TConfig] = None, target_fidelities: Optional[Dict[int, float]] = None, ) -> Optional[Tensor]: """ A wrapper around `ax.models.model_utils.best_observed_point` for TorchModel that recommends a best point from previously observed points using either a "max_utility" or "feasible_threshold" strategy. Args: model: A TorchModel. bounds: A list of (lower, upper) tuples for each column of X. objective_weights: The objective is to maximize a weighted sum of the columns of f(x). These are the weights. outcome_constraints: A tuple of (A, b). For k outcome constraints and m outputs at f(x), A is (k x m) and b is (k x 1) such that A f(x) <= b. linear_constraints: A tuple of (A, b). For k linear constraints on d-dimensional x, A is (k x d) and b is (k x 1) such that A x <= b. fixed_features: A map {feature_index: value} for features that should be fixed to a particular value in the best point. model_gen_options: A config dictionary that can contain model-specific options. target_fidelities: A map {feature_index: value} of fidelity feature column indices to their respective target fidelities. Used for multi-fidelity optimization. Returns: A d-array of the best point, or None if no feasible point was observed. """ if target_fidelities: raise NotImplementedError( "target_fidelities not implemented for base BotorchModel" ) x_best = best_observed_point( model=model, bounds=bounds, objective_weights=objective_weights, outcome_constraints=outcome_constraints, linear_constraints=linear_constraints, fixed_features=fixed_features, options=model_gen_options, ) if x_best is None: return None return x_best.to(dtype=model.dtype, device=torch.device("cpu"))
[docs]def recommend_best_out_of_sample_point( model: TorchModel, bounds: List[Tuple[float, float]], objective_weights: Tensor, outcome_constraints: Optional[Tuple[Tensor, Tensor]] = None, linear_constraints: Optional[Tuple[Tensor, Tensor]] = None, fixed_features: Optional[Dict[int, float]] = None, model_gen_options: Optional[TConfig] = None, target_fidelities: Optional[Dict[int, float]] = None, ) -> Optional[Tensor]: """ Identify the current best point by optimizing the posterior mean of the model. This is "out-of-sample" because it considers un-observed designs as well. Return None if no such point can be identified. Args: model: A TorchModel. bounds: A list of (lower, upper) tuples for each column of X. objective_weights: The objective is to maximize a weighted sum of the columns of f(x). These are the weights. outcome_constraints: A tuple of (A, b). For k outcome constraints and m outputs at f(x), A is (k x m) and b is (k x 1) such that A f(x) <= b. linear_constraints: A tuple of (A, b). For k linear constraints on d-dimensional x, A is (k x d) and b is (k x 1) such that A x <= b. fixed_features: A map {feature_index: value} for features that should be fixed to a particular value in the best point. model_gen_options: A config dictionary that can contain model-specific options. target_fidelities: A map {feature_index: value} of fidelity feature column indices to their respective target fidelities. Used for multi-fidelity optimization. Returns: A d-array of the best point, or None if no feasible point exists. """ options = model_gen_options or {} fixed_features = fixed_features or {} acf_options = options.get("acquisition_function_kwargs", {}) optimizer_options = options.get("optimizer_kwargs", {}) X_observed = get_observed( Xs=model.Xs, # pyre-ignore: [16] objective_weights=objective_weights, outcome_constraints=outcome_constraints, ) if hasattr(model, "_get_best_point_acqf"): acq_function, non_fixed_idcs = model._get_best_point_acqf( # pyre-ignore: [16] X_observed=X_observed, objective_weights=objective_weights, mc_samples=acf_options.get("mc_samples", 512), fixed_features=fixed_features, target_fidelities=target_fidelities, outcome_constraints=outcome_constraints, seed_inner=acf_options.get("seed_inner", None), qmc=acf_options.get("qmc", True), ) else: raise RuntimeError("The model should implement _get_best_point_acqf.") inequality_constraints = _to_inequality_constraints(linear_constraints) # TODO: update optimizers to handle inequality_constraints # (including transforming constraints b/c of fixed features) if inequality_constraints is not None: raise UnsupportedError("Inequality constraints are not supported!") return_best_only = optimizer_options.get("return_best_only", True) bounds_ = torch.tensor(bounds, dtype=model.dtype, device=model.device) bounds_ = bounds_.transpose(-1, -2) if non_fixed_idcs is not None: bounds_ = bounds_[..., non_fixed_idcs] candidates, _ = optimize_acqf( acq_function=acq_function, bounds=bounds_, q=1, num_restarts=optimizer_options.get("num_restarts", 60), raw_samples=optimizer_options.get("raw_samples", 1024), inequality_constraints=inequality_constraints, fixed_features=None, # handled inside the acquisition function options={ "batch_limit": optimizer_options.get("batch_limit", 8), "maxiter": optimizer_options.get("maxiter", 200), "nonnegative": optimizer_options.get("nonnegative", False), "method": "L-BFGS-B", }, return_best_only=return_best_only, ) rec_point = candidates.detach().cpu() if isinstance(acq_function, FixedFeatureAcquisitionFunction): rec_point = acq_function._construct_X_full(rec_point) if return_best_only: rec_point = rec_point.view(-1) return rec_point
def _get_model( X: Tensor, Y: Tensor, Yvar: Tensor, task_feature: Optional[int] = None, fidelity_features: Optional[List[int]] = None, **kwargs: Any, ) -> GPyTorchModel: """Instantiate a model of type depending on the input data. Args: X: A `n x d` tensor of input features. Y: A `n x m` tensor of input observations. Yvar: A `n x m` tensor of input variances (NaN if unobserved). task_feature: The index of the column pertaining to the task feature (if present). fidelity_features: List of columns of X that are fidelity parameters. Returns: A GPyTorchModel (unfitted). """ Yvar = Yvar.clamp_min_(MIN_OBSERVED_NOISE_LEVEL) is_nan = torch.isnan(Yvar) any_nan_Yvar = torch.any(is_nan) all_nan_Yvar = torch.all(is_nan) if any_nan_Yvar and not all_nan_Yvar: if task_feature: # TODO (jej): Replace with inferred noise before making perf judgements. Yvar[Yvar != Yvar] = MIN_OBSERVED_NOISE_LEVEL else: raise ValueError( "Mix of known and unknown variances indicates valuation function " "errors. Variances should all be specified, or none should be." ) if fidelity_features is None: fidelity_features = [] if len(fidelity_features) == 0: # only pass linear_truncated arg if there are fidelities kwargs = {k: v for k, v in kwargs.items() if k != "linear_truncated"} if len(fidelity_features) > 0: if task_feature: raise NotImplementedError( # pragma: no cover "multi-task multi-fidelity models not yet available" ) # at this point we can assume that there is only a single fidelity parameter gp = SingleTaskMultiFidelityGP( train_X=X, train_Y=Y, data_fidelity=fidelity_features[0], **kwargs ) elif task_feature is None and all_nan_Yvar: gp = SingleTaskGP(train_X=X, train_Y=Y, **kwargs) elif task_feature is None: gp = FixedNoiseGP(train_X=X, train_Y=Y, train_Yvar=Yvar, **kwargs) else: # instantiate multitask GP all_tasks, _, _ = MultiTaskGP.get_all_tasks(X, task_feature) num_tasks = len(all_tasks) prior_dict = kwargs.get("prior") prior = None if prior_dict is not None: prior_type = prior_dict.get("type", None) if issubclass(prior_type, LKJCovariancePrior): sd_prior = prior_dict.get("sd_prior", GammaPrior(1.0, 0.15)) sd_prior._event_shape = torch.Size([num_tasks]) eta = prior_dict.get("eta", 0.5) if not isinstance(eta, float) and not isinstance(eta, int): raise ValueError(f"eta must be a real number, your eta was {eta}") prior = LKJCovariancePrior(num_tasks, eta, sd_prior) else: raise NotImplementedError( "Currently only LKJ prior is supported," f"your prior type was {prior_type}." ) if all_nan_Yvar: gp = MultiTaskGP( train_X=X, train_Y=Y, task_feature=task_feature, rank=kwargs.get("rank"), prior=prior, ) else: gp = FixedNoiseMultiTaskGP( train_X=X, train_Y=Y, train_Yvar=Yvar, task_feature=task_feature, rank=kwargs.get("rank"), prior=prior, ) return gp