Source code for botorch.acquisition.multi_objective.monte_carlo

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

r"""
Monte-Carlo Acquisition Functions for Multi-objective Bayesian optimization.

References

.. [Daulton2020qehvi]
    S. Daulton, M. Balandat, and E. Bakshy. Differentiable Expected Hypervolume
    Improvement for Parallel Multi-Objective Bayesian Optimization. Advances in Neural
    Information Processing Systems 33, 2020.

.. [Daulton2021nehvi]
    S. Daulton, M. Balandat, and E. Bakshy. Parallel Bayesian Optimization of
    Multiple Noisy Objectives with Expected Hypervolume Improvement. Advances
    in Neural Information Processing Systems 34, 2021.

"""

from __future__ import annotations

import warnings
from abc import ABC, abstractmethod
from copy import deepcopy
from itertools import combinations
from typing import Any, Callable, List, Optional, Union

import torch
from botorch.acquisition.acquisition import AcquisitionFunction, MCSamplerMixin
from botorch.acquisition.cached_cholesky import CachedCholeskyMCAcquisitionFunction
from botorch.acquisition.multi_objective.objective import (
    IdentityMCMultiOutputObjective,
    MCMultiOutputObjective,
)
from botorch.acquisition.multi_objective.utils import (
    prune_inferior_points_multi_objective,
)
from botorch.exceptions.errors import UnsupportedError
from botorch.exceptions.warnings import BotorchWarning
from botorch.models.model import Model
from botorch.models.transforms.input import InputPerturbation
from botorch.sampling.base import MCSampler
from botorch.utils.multi_objective.box_decompositions.box_decomposition_list import (
    BoxDecompositionList,
)
from botorch.utils.multi_objective.box_decompositions.dominated import (
    DominatedPartitioning,
)
from botorch.utils.multi_objective.box_decompositions.non_dominated import (
    FastNondominatedPartitioning,
    NondominatedPartitioning,
)
from botorch.utils.multi_objective.box_decompositions.utils import (
    _pad_batch_pareto_frontier,
)
from botorch.utils.objective import compute_smoothed_feasibility_indicator
from botorch.utils.torch import BufferDict
from botorch.utils.transforms import (
    concatenate_pending_points,
    is_fully_bayesian,
    match_batch_shape,
    t_batch_mode_transform,
)
from torch import Tensor


[docs]class MultiObjectiveMCAcquisitionFunction(AcquisitionFunction, MCSamplerMixin, ABC): r"""Abstract base class for Multi-Objective batch acquisition functions. NOTE: This does not inherit from `MCAcquisitionFunction` to avoid circular imports. Args: _default_sample_shape: The `sample_shape` for the default sampler. """ _default_sample_shape = torch.Size([128]) def __init__( self, model: Model, sampler: Optional[MCSampler] = None, objective: Optional[MCMultiOutputObjective] = None, constraints: Optional[List[Callable[[Tensor], Tensor]]] = None, eta: Optional[Union[Tensor, float]] = 1e-3, X_pending: Optional[Tensor] = None, ) -> None: r"""Constructor for the MCAcquisitionFunction base class. Args: model: A fitted model. sampler: The sampler used to draw base samples. If not given, a sampler is generated using `get_sampler`. NOTE: For posteriors that do not support base samples, a sampler compatible with intended use case must be provided. See `ForkedRNGSampler` and `StochasticSampler` as examples. objective: The MCMultiOutputObjective under which the samples are evaluated. Defaults to `IdentityMultiOutputObjective()`. constraints: A list of callables, each mapping a Tensor of dimension `sample_shape x batch-shape x q x m` to a Tensor of dimension `sample_shape x batch-shape x q`, where negative values imply feasibility. eta: The temperature parameter for the sigmoid function used for the differentiable approximation of the constraints. In case of a float the same eta is used for every constraint in constraints. In case of a tensor the length of the tensor must match the number of provided constraints. The i-th constraint is then estimated with the i-th eta value. X_pending: A `m x d`-dim Tensor of `m` design points that have points that have been submitted for function evaluation but have not yet been evaluated. """ super().__init__(model=model) MCSamplerMixin.__init__(self, sampler=sampler) if objective is None: objective = IdentityMCMultiOutputObjective() elif not isinstance(objective, MCMultiOutputObjective): raise UnsupportedError( "Only objectives of type MCMultiOutputObjective are supported for " "Multi-Objective MC acquisition functions." ) if ( hasattr(model, "input_transform") and isinstance(model.input_transform, InputPerturbation) and constraints is not None ): raise UnsupportedError( "Constraints are not supported with input perturbations, due to" "sample q-batch shape being different than that of the inputs." "Use a composite objective that applies feasibility weighting to" "samples before calculating the risk measure." ) self.add_module("objective", objective) self.constraints = constraints if constraints: if type(eta) is not Tensor: eta = torch.full((len(constraints),), eta) self.register_buffer("eta", eta) self.X_pending = None if X_pending is not None: self.set_X_pending(X_pending)
[docs] @abstractmethod def forward(self, X: Tensor) -> Tensor: r"""Takes in a `batch_shape x q x d` X Tensor of t-batches with `q` `d`-dim design points each, and returns a Tensor with shape `batch_shape'`, where `batch_shape'` is the broadcasted batch shape of model and input `X`. Should utilize the result of `set_X_pending` as needed to account for pending function evaluations. """ pass # pragma: no cover
[docs]class qExpectedHypervolumeImprovement(MultiObjectiveMCAcquisitionFunction): def __init__( self, model: Model, ref_point: Union[List[float], Tensor], partitioning: NondominatedPartitioning, sampler: Optional[MCSampler] = None, objective: Optional[MCMultiOutputObjective] = None, constraints: Optional[List[Callable[[Tensor], Tensor]]] = None, X_pending: Optional[Tensor] = None, eta: Optional[Union[Tensor, float]] = 1e-3, ) -> None: r"""q-Expected Hypervolume Improvement supporting m>=2 outcomes. See [Daulton2020qehvi]_ for details. Example: >>> model = SingleTaskGP(train_X, train_Y) >>> ref_point = [0.0, 0.0] >>> qEHVI = qExpectedHypervolumeImprovement(model, ref_point, partitioning) >>> qehvi = qEHVI(test_X) Args: model: A fitted model. ref_point: A list or tensor with `m` elements representing the reference point (in the outcome space) w.r.t. to which compute the hypervolume. This is a reference point for the objective values (i.e. after applying`objective` to the samples). partitioning: A `NondominatedPartitioning` module that provides the non- dominated front and a partitioning of the non-dominated space in hyper- rectangles. If constraints are present, this partitioning must only include feasible points. sampler: The sampler used to draw base samples. If not given, a sampler is generated using `get_sampler`. objective: The MCMultiOutputObjective under which the samples are evaluated. Defaults to `IdentityMultiOutputObjective()`. constraints: A list of callables, each mapping a Tensor of dimension `sample_shape x batch-shape x q x m` to a Tensor of dimension `sample_shape x batch-shape x q`, where negative values imply feasibility. The acqusition function will compute expected feasible hypervolume. X_pending: A `batch_shape x m x d`-dim Tensor of `m` design points that have points that have been submitted for function evaluation but have not yet been evaluated. Concatenated into `X` upon forward call. Copied and set to have no gradient. eta: The temperature parameter for the sigmoid function used for the differentiable approximation of the constraints. In case of a float the same eta is used for every constraint in constraints. In case of a tensor the length of the tensor must match the number of provided constraints. The i-th constraint is then estimated with the i-th eta value. """ if len(ref_point) != partitioning.num_outcomes: raise ValueError( "The length of the reference point must match the number of outcomes. " f"Got ref_point with {len(ref_point)} elements, but expected " f"{partitioning.num_outcomes}." ) ref_point = torch.as_tensor( ref_point, dtype=partitioning.pareto_Y.dtype, device=partitioning.pareto_Y.device, ) super().__init__( model=model, sampler=sampler, objective=objective, constraints=constraints, eta=eta, X_pending=X_pending, ) self.register_buffer("ref_point", ref_point) cell_bounds = partitioning.get_hypercell_bounds() self.register_buffer("cell_lower_bounds", cell_bounds[0]) self.register_buffer("cell_upper_bounds", cell_bounds[1]) self.q_out = -1 self.q_subset_indices = BufferDict() def _cache_q_subset_indices(self, q_out: int) -> None: r"""Cache indices corresponding to all subsets of `q_out`. This means that consecutive calls to `forward` with the same `q_out` will not recompute the indices for all (2^q_out - 1) subsets. Note: this will use more memory than regenerating the indices for each i and then deleting them, but it will be faster for repeated evaluations (e.g. during optimization). Args: q_out: The batch size of the objectives. This is typically equal to the q-batch size of `X`. However, if using a set valued objective (e.g., MVaR) that produces `s` objective values for each point on the q-batch of `X`, we need to properly account for each objective while calculating the hypervolume contributions by using `q_out = q * s`. """ if q_out != self.q_out: indices = list(range(q_out)) tkwargs = {"dtype": torch.long, "device": self.ref_point.device} self.q_subset_indices = BufferDict( { f"q_choose_{i}": torch.tensor( list(combinations(indices, i)), **tkwargs ) for i in range(1, q_out + 1) } ) self.q_out = q_out def _compute_qehvi(self, samples: Tensor, X: Optional[Tensor] = None) -> Tensor: r"""Compute the expected (feasible) hypervolume improvement given MC samples. Args: samples: A `n_samples x batch_shape x q' x m`-dim tensor of samples. X: A `batch_shape x q x d`-dim tensor of inputs. Returns: A `batch_shape x (model_batch_shape)`-dim tensor of expected hypervolume improvement for each batch. """ # Note that the objective may subset the outcomes (e.g. this will usually happen # if there are constraints present). obj = self.objective(samples, X=X) q = obj.shape[-2] if self.constraints is not None: feas_weights = compute_smoothed_feasibility_indicator( constraints=self.constraints, samples=samples, eta=self.eta ) # `sample_shape x batch-shape x q` self._cache_q_subset_indices(q_out=q) batch_shape = obj.shape[:-2] # this is n_samples x input_batch_shape x areas_per_segment = torch.zeros( *batch_shape, self.cell_lower_bounds.shape[-2], dtype=obj.dtype, device=obj.device, ) cell_batch_ndim = self.cell_lower_bounds.ndim - 2 sample_batch_view_shape = torch.Size( [ batch_shape[0] if cell_batch_ndim > 0 else 1, *[1 for _ in range(len(batch_shape) - max(cell_batch_ndim, 1))], *self.cell_lower_bounds.shape[1:-2], ] ) view_shape = ( *sample_batch_view_shape, self.cell_upper_bounds.shape[-2], 1, self.cell_upper_bounds.shape[-1], ) for i in range(1, self.q_out + 1): # TODO: we could use batches to compute (q choose i) and (q choose q-i) # simultaneously since subsets of size i and q-i have the same number of # elements. This would decrease the number of iterations, but increase # memory usage. q_choose_i = self.q_subset_indices[f"q_choose_{i}"] # this tensor is mc_samples x batch_shape x i x q_choose_i x m obj_subsets = obj.index_select(dim=-2, index=q_choose_i.view(-1)) obj_subsets = obj_subsets.view( obj.shape[:-2] + q_choose_i.shape + obj.shape[-1:] ) # since all hyperrectangles share one vertex, the opposite vertex of the # overlap is given by the component-wise minimum. # take the minimum in each subset overlap_vertices = obj_subsets.min(dim=-2).values # add batch-dim to compute area for each segment (pseudo-pareto-vertex) # this tensor is mc_samples x batch_shape x num_cells x q_choose_i x m overlap_vertices = torch.min( overlap_vertices.unsqueeze(-3), self.cell_upper_bounds.view(view_shape) ) # substract cell lower bounds, clamp min at zero lengths_i = ( overlap_vertices - self.cell_lower_bounds.view(view_shape) ).clamp_min(0.0) # take product over hyperrectangle side lengths to compute area # sum over all subsets of size i areas_i = lengths_i.prod(dim=-1) # if constraints are present, apply a differentiable approximation of # the indicator function if self.constraints is not None: feas_subsets = feas_weights.index_select( dim=-1, index=q_choose_i.view(-1) ).view(feas_weights.shape[:-1] + q_choose_i.shape) areas_i = areas_i * feas_subsets.unsqueeze(-3).prod(dim=-1) areas_i = areas_i.sum(dim=-1) # Using the inclusion-exclusion principle, set the sign to be positive # for subsets of odd sizes and negative for subsets of even size areas_per_segment += (-1) ** (i + 1) * areas_i # sum over segments and average over MC samples return areas_per_segment.sum(dim=-1).mean(dim=0)
[docs] @concatenate_pending_points @t_batch_mode_transform() def forward(self, X: Tensor) -> Tensor: posterior = self.model.posterior(X) samples = self.get_posterior_samples(posterior) return self._compute_qehvi(samples=samples, X=X)
[docs]class qNoisyExpectedHypervolumeImprovement( qExpectedHypervolumeImprovement, CachedCholeskyMCAcquisitionFunction ): def __init__( self, model: Model, ref_point: Union[List[float], Tensor], X_baseline: Tensor, sampler: Optional[MCSampler] = None, objective: Optional[MCMultiOutputObjective] = None, constraints: Optional[List[Callable[[Tensor], Tensor]]] = None, X_pending: Optional[Tensor] = None, eta: Optional[Union[Tensor, float]] = 1e-3, prune_baseline: bool = False, alpha: float = 0.0, cache_pending: bool = True, max_iep: int = 0, incremental_nehvi: bool = True, cache_root: bool = True, **kwargs: Any, ) -> None: r"""q-Noisy Expected Hypervolume Improvement supporting m>=2 outcomes. See [Daulton2021nehvi]_ for details. Example: >>> model = SingleTaskGP(train_X, train_Y) >>> ref_point = [0.0, 0.0] >>> qNEHVI = qNoisyExpectedHypervolumeImprovement(model, ref_point, train_X) >>> qnehvi = qNEHVI(test_X) Args: model: A fitted model. ref_point: A list or tensor with `m` elements representing the reference point (in the outcome space) w.r.t. to which compute the hypervolume. This is a reference point for the objective values (i.e. after applying `objective` to the samples). X_baseline: A `r x d`-dim Tensor of `r` design points that have already been observed. These points are considered as potential approximate pareto-optimal design points. sampler: The sampler used to draw base samples. If not given, a sampler is generated using `get_sampler`. Note: a pareto front is created for each mc sample, which can be computationally intensive for `m` > 2. objective: The MCMultiOutputObjective under which the samples are evaluated. Defaults to `IdentityMultiOutputObjective()`. constraints: A list of callables, each mapping a Tensor of dimension `sample_shape x batch-shape x q x m` to a Tensor of dimension `sample_shape x batch-shape x q`, where negative values imply feasibility. The acqusition function will compute expected feasible hypervolume. X_pending: A `batch_shape x m x d`-dim Tensor of `m` design points that have points that have been submitted for function evaluation, but have not yet been evaluated. eta: The temperature parameter for the sigmoid function used for the differentiable approximation of the constraints. In case of a float the same eta is used for every constraint in constraints. In case of a tensor the length of the tensor must match the number of provided constraints. The i-th constraint is then estimated with the i-th eta value. For more details, on this parameter, see the docs of `compute_smoothed_feasibility_indicator`. prune_baseline: If True, remove points in `X_baseline` that are highly unlikely to be the pareto optimal and better than the reference point. This can significantly improve computation time and is generally recommended. In order to customize pruning parameters, instead manually call `prune_inferior_points_multi_objective` on `X_baseline` before instantiating the acquisition function. alpha: The hyperparameter controlling the approximate non-dominated partitioning. The default value of 0.0 means an exact partitioning is used. As the number of objectives `m` increases, consider increasing this parameter in order to limit computational complexity. cache_pending: A boolean indicating whether to use cached box decompositions (CBD) for handling pending points. This is generally recommended. max_iep: The maximum number of pending points before the box decompositions will be recomputed. incremental_nehvi: A boolean indicating whether to compute the incremental NEHVI from the `i`th point where `i=1, ..., q` under sequential greedy optimization, or the full qNEHVI over `q` points. cache_root: A boolean indicating whether to cache the root decomposition over `X_baseline` and use low-rank updates. """ if len(ref_point) < 2: raise ValueError( "qNoisyExpectedHypervolumeImprovement supports m>=2 outcomes " f"but ref_point has length {len(ref_point)}, which is smaller than 2." ) ref_point = torch.as_tensor( ref_point, dtype=X_baseline.dtype, device=X_baseline.device ) super(qExpectedHypervolumeImprovement, self).__init__( model=model, sampler=sampler, objective=objective, constraints=constraints, eta=eta, ) self._setup(model=model, cache_root=cache_root) if X_baseline.ndim > 2: raise UnsupportedError( "qNoisyExpectedHypervolumeImprovement does not support batched " f"X_baseline. Expected 2 dims, got {X_baseline.ndim}." ) if prune_baseline: X_baseline = prune_inferior_points_multi_objective( model=model, X=X_baseline, objective=objective, constraints=constraints, ref_point=ref_point, marginalize_dim=kwargs.get("marginalize_dim"), ) self.register_buffer("ref_point", ref_point) self.alpha = alpha self.q_in = -1 self.q_out = -1 self.q_subset_indices = BufferDict() self.partitioning = None # set partitioning class and args self.p_kwargs = {} if self.alpha > 0: self.p_kwargs["alpha"] = self.alpha self.p_class = NondominatedPartitioning else: self.p_class = FastNondominatedPartitioning self.register_buffer("_X_baseline", X_baseline) self.register_buffer("_X_baseline_and_pending", X_baseline) self.register_buffer( "cache_pending", torch.tensor(cache_pending, dtype=bool), ) self.register_buffer( "_prev_nehvi", torch.tensor(0.0, dtype=ref_point.dtype, device=ref_point.device), ) self.register_buffer( "_max_iep", torch.tensor(max_iep, dtype=torch.long), ) self.register_buffer( "incremental_nehvi", torch.tensor(incremental_nehvi, dtype=torch.bool), ) # Base sampler is initialized in _set_cell_bounds. self.base_sampler = None if X_pending is not None: # This will call self._set_cell_bounds if the number of pending # points is greater than self._max_iep. self.set_X_pending(X_pending) # In the case that X_pending is not None, but there are fewer than # max_iep pending points, the box decompositions are not performed in # set_X_pending. Therefore, we need to perform a box decomposition over # f(X_baseline) here. if X_pending is None or X_pending.shape[-2] <= self._max_iep: self._set_cell_bounds(num_new_points=X_baseline.shape[0]) # Set q_in=-1 to so that self.sampler is updated at the next forward call. self.q_in = -1 @property def X_baseline(self) -> Tensor: r"""Return X_baseline augmented with pending points cached using CBD.""" return self._X_baseline_and_pending def _compute_initial_hvs(self, obj: Tensor, feas: Optional[Tensor] = None) -> None: r"""Compute hypervolume dominated by f(X_baseline) under each sample. Args: obj: A `sample_shape x batch_shape x n x m`-dim tensor of samples of objectives. feas: `sample_shape x batch_shape x n`-dim tensor of samples of feasibility indicators. """ initial_hvs = [] for i, sample in enumerate(obj): if self.constraints is not None: sample = sample[feas[i]] dominated_partitioning = DominatedPartitioning( ref_point=self.ref_point, Y=sample, ) hv = dominated_partitioning.compute_hypervolume() initial_hvs.append(hv) self.register_buffer( "_initial_hvs", torch.tensor(initial_hvs, dtype=obj.dtype, device=obj.device).view( self._batch_sample_shape, *obj.shape[-2:] ), ) def _set_cell_bounds(self, num_new_points: int) -> None: r"""Compute the box decomposition under each posterior sample. Args: num_new_points: The number of new points (beyond the points in X_baseline) that were used in the previous box decomposition. In the first box decomposition, this should be the number of points in X_baseline. """ feas = None if self.X_baseline.shape[0] > 0: with torch.no_grad(): posterior = self.model.posterior(self.X_baseline) # Reset sampler, accounting for possible one-to-many transform. self.q_in = -1 if self.base_sampler is None: # Initialize the base sampler if needed. samples = self.get_posterior_samples(posterior) self.base_sampler = deepcopy(self.sampler) else: samples = self.base_sampler(posterior) n_w = posterior._extended_shape()[-2] // self.X_baseline.shape[-2] self._set_sampler(q_in=num_new_points * n_w, posterior=posterior) # cache posterior if self._cache_root: # Note that this implicitly uses LinearOperator's caching to check if # the proper root decomposition has already been cached to # `posterior.mvn.lazy_covariance_matrix`, which it may have been in # the call to `self.base_sampler`, and computes it if not found self._baseline_L = self._compute_root_decomposition(posterior=posterior) obj = self.objective(samples, X=self.X_baseline) if self.constraints is not None: feas = torch.stack( [c(samples) <= 0 for c in self.constraints], dim=0 ).all(dim=0) else: sample_shape = ( self.sampler.sample_shape if self.sampler is not None else self._default_sample_shape ) obj = torch.empty( *sample_shape, 0, self.ref_point.shape[-1], dtype=self.ref_point.dtype, device=self.ref_point.device, ) self._batch_sample_shape = obj.shape[:-2] # collapse batch dimensions # use numel() rather than view(-1) to handle case of no baseline points new_batch_shape = self._batch_sample_shape.numel() obj = obj.view(new_batch_shape, *obj.shape[-2:]) if self.constraints is not None and feas is not None: feas = feas.view(new_batch_shape, *feas.shape[-1:]) if self.partitioning is None and not self.incremental_nehvi: self._compute_initial_hvs(obj=obj, feas=feas) if self.ref_point.shape[-1] > 2: # the partitioning algorithms run faster on the CPU # due to advanced indexing ref_point_cpu = self.ref_point.cpu() obj_cpu = obj.cpu() if self.constraints is not None and feas is not None: feas_cpu = feas.cpu() obj_cpu = [obj_cpu[i][feas_cpu[i]] for i in range(obj.shape[0])] partitionings = [] for sample in obj_cpu: partitioning = self.p_class( ref_point=ref_point_cpu, Y=sample, **self.p_kwargs ) partitionings.append(partitioning) self.partitioning = BoxDecompositionList(*partitionings) else: # use batched partitioning obj = _pad_batch_pareto_frontier( Y=obj, ref_point=self.ref_point.unsqueeze(0).expand( obj.shape[0], self.ref_point.shape[-1] ), feasibility_mask=feas, ) self.partitioning = self.p_class( ref_point=self.ref_point, Y=obj, **self.p_kwargs ) cell_bounds = self.partitioning.get_hypercell_bounds().to(self.ref_point) cell_bounds = cell_bounds.view( 2, *self._batch_sample_shape, *cell_bounds.shape[-2:] ) self.register_buffer("cell_lower_bounds", cell_bounds[0]) self.register_buffer("cell_upper_bounds", cell_bounds[1])
[docs] def set_X_pending(self, X_pending: Optional[Tensor] = None) -> None: r"""Informs the acquisition function about pending design points. Args: X_pending: `n x d` Tensor with `n` `d`-dim design points that have been submitted for evaluation but have not yet been evaluated. """ if X_pending is None: self.X_pending = None else: if X_pending.requires_grad: warnings.warn( "Pending points require a gradient but the acquisition function" " will not provide a gradient to these points.", BotorchWarning, ) X_pending = X_pending.detach().clone() if self.cache_pending: X_baseline = torch.cat([self._X_baseline, X_pending], dim=-2) # Number of new points is the total number of points minus # (the number of previously cached pending points plus the # of number of baseline points). num_new_points = X_baseline.shape[0] - self.X_baseline.shape[0] if num_new_points > 0: if num_new_points > self._max_iep: # Set the new baseline points to include pending points. self.register_buffer("_X_baseline_and_pending", X_baseline) # Recompute box decompositions. self._set_cell_bounds(num_new_points=num_new_points) if not self.incremental_nehvi: self._prev_nehvi = ( (self._hypervolumes - self._initial_hvs) .clamp_min(0.0) .mean() ) # Set to None so that pending points are not concatenated in # forward. self.X_pending = None # Set q_in=-1 to so that self.sampler is updated at the next # forward call. self.q_in = -1 else: self.X_pending = X_pending[-num_new_points:] else: self.X_pending = X_pending
@property def _hypervolumes(self) -> Tensor: r"""Compute hypervolume over X_baseline under each posterior sample. Returns: A `n_samples`-dim tensor of hypervolumes. """ return ( self.partitioning.compute_hypervolume() .to(self.ref_point) # for m > 2, the partitioning is on the CPU .view(self._batch_sample_shape) )
[docs] @concatenate_pending_points @t_batch_mode_transform() def forward(self, X: Tensor) -> Tensor: X_full = torch.cat([match_batch_shape(self.X_baseline, X), X], dim=-2) # Note: it is important to compute the full posterior over `(X_baseline, X)` # to ensure that we properly sample `f(X)` from the joint distribution ` # `f(X_baseline, X) ~ P(f | D)` given that we can already fixed the sampled # function values for `f(X_baseline)`. # TODO: improve efficiency by not recomputing baseline-baseline # covariance matrix. posterior = self.model.posterior(X_full) # Account for possible one-to-many transform and the MCMC batch dimension in # `SaasFullyBayesianSingleTaskGP` event_shape_lag = 1 if is_fully_bayesian(self.model) else 2 n_w = ( posterior._extended_shape()[X_full.dim() - event_shape_lag] // X_full.shape[-2] ) q_in = X.shape[-2] * n_w self._set_sampler(q_in=q_in, posterior=posterior) samples = self._get_f_X_samples(posterior=posterior, q_in=q_in) # Add previous nehvi from pending points. return self._compute_qehvi(samples=samples, X=X) + self._prev_nehvi