#!/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"""Hypervolume Utilities.
References
.. [Fonseca2006]
C. M. Fonseca, L. Paquete, and M. Lopez-Ibanez. An improved dimension-sweep
algorithm for the hypervolume indicator. In IEEE Congress on Evolutionary
Computation, pages 1157-1163, Vancouver, Canada, July 2006.
.. [Ishibuchi2011]
H. Ishibuchi, N. Akedo, and Y. Nojima. A many-objective test problem
for visually examining diversity maintenance behavior in a decision
space. Proc. 13th Annual Conf. Genetic Evol. Comput., 2011.
"""
from __future__ import annotations
import warnings
from copy import deepcopy
from itertools import combinations
from typing import Callable, List, Optional, Union
import torch
from botorch.acquisition.cached_cholesky import CachedCholeskyMCSamplerMixin
from botorch.acquisition.multi_objective.objective import MCMultiOutputObjective
from botorch.acquisition.multi_objective.utils import (
prune_inferior_points_multi_objective,
)
from botorch.exceptions.errors import (
BotorchError,
BotorchTensorDimensionError,
UnsupportedError,
)
from botorch.exceptions.warnings import BotorchWarning
from botorch.models.model import Model
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_feasibility_indicator
from botorch.utils.torch import BufferDict
from torch import Tensor
MIN_Y_RANGE = 1e-7
[docs]
def infer_reference_point(
pareto_Y: Tensor,
max_ref_point: Optional[Tensor] = None,
scale: float = 0.1,
scale_max_ref_point: bool = False,
) -> Tensor:
r"""Get reference point for hypervolume computations.
This sets the reference point to be `ref_point = nadir - scale * range`
when there is no `pareto_Y` that is better than `max_ref_point`.
If there's `pareto_Y` better than `max_ref_point`, the reference point
will be set to `max_ref_point - scale * range` if `scale_max_ref_point`
is true and to `max_ref_point` otherwise.
[Ishibuchi2011]_ find 0.1 to be a robust multiplier for scaling the
nadir point.
Note: this assumes maximization of all objectives.
Args:
pareto_Y: A `n x m`-dim tensor of Pareto-optimal points.
max_ref_point: A `m` dim tensor indicating the maximum reference point.
Some elements can be NaN, except when `pareto_Y` is empty,
in which case these dimensions will be treated as if no
`max_ref_point` was provided and set to `nadir - scale * range`.
scale: A multiplier used to scale back the reference point based on the
range of each objective.
scale_max_ref_point: A boolean indicating whether to apply scaling to
the max_ref_point based on the range of each objective.
Returns:
A `m`-dim tensor containing the reference point.
"""
if pareto_Y.shape[0] == 0:
if max_ref_point is None:
raise BotorchError("Empty pareto set and no max ref point provided")
if max_ref_point.isnan().any():
raise BotorchError("Empty pareto set and max ref point includes NaN.")
if scale_max_ref_point:
return max_ref_point - scale * max_ref_point.abs()
return max_ref_point
if max_ref_point is not None:
non_nan_idx = ~max_ref_point.isnan()
# Count all points exceeding non-NaN reference point as being better.
better_than_ref = (pareto_Y[:, non_nan_idx] > max_ref_point[non_nan_idx]).all(
dim=-1
)
else:
non_nan_idx = torch.ones(
pareto_Y.shape[-1], dtype=torch.bool, device=pareto_Y.device
)
better_than_ref = torch.ones(
pareto_Y.shape[:1], dtype=torch.bool, device=pareto_Y.device
)
if max_ref_point is not None and better_than_ref.any() and non_nan_idx.all():
Y_range = pareto_Y[better_than_ref].max(dim=0).values - max_ref_point
if scale_max_ref_point:
return max_ref_point - scale * Y_range
return max_ref_point
elif pareto_Y.shape[0] == 1:
# no points better than max_ref_point and only a single observation
# subtract MIN_Y_RANGE to handle the case that pareto_Y is a singleton
# with objective value of 0.
Y_range = pareto_Y.abs().clamp_min(MIN_Y_RANGE).view(-1)
ref_point = pareto_Y.view(-1) - scale * Y_range
else:
# no points better than max_ref_point and multiple observations
# make sure that each dimension of the nadir point is no greater than
# the max_ref_point
nadir = pareto_Y.min(dim=0).values
if max_ref_point is not None:
nadir[non_nan_idx] = torch.min(
nadir[non_nan_idx], max_ref_point[non_nan_idx]
)
ideal = pareto_Y.max(dim=0).values
# handle case where all values for one objective are the same
Y_range = (ideal - nadir).clamp_min(MIN_Y_RANGE)
ref_point = nadir - scale * Y_range
# Set not-nan indices - if any - to max_ref_point.
if non_nan_idx.any() and not non_nan_idx.all() and better_than_ref.any():
if scale_max_ref_point:
ref_point[non_nan_idx] = (max_ref_point - scale * Y_range)[non_nan_idx]
else:
ref_point[non_nan_idx] = max_ref_point[non_nan_idx]
return ref_point
[docs]
class Hypervolume:
r"""Hypervolume computation dimension sweep algorithm from [Fonseca2006]_.
Adapted from Simon Wessing's implementation of the algorithm
(Variant 3, Version 1.2) in [Fonseca2006]_ in PyMOO:
https://github.com/msu-coinlab/pymoo/blob/master/pymoo/vendor/hv.py
Maximization is assumed.
TODO: write this in C++ for faster looping.
"""
def __init__(self, ref_point: Tensor) -> None:
r"""Initialize hypervolume object.
Args:
ref_point: `m`-dim Tensor containing the reference point.
"""
self.ref_point = ref_point
@property
def ref_point(self) -> Tensor:
r"""Get reference point (for maximization).
Returns:
A `m`-dim tensor containing the reference point.
"""
return -self._ref_point
@ref_point.setter
def ref_point(self, ref_point: Tensor) -> None:
r"""Set the reference point for maximization
Args:
ref_point: A `m`-dim tensor containing the reference point.
"""
self._ref_point = -ref_point
[docs]
def compute(self, pareto_Y: Tensor) -> float:
r"""Compute hypervolume.
Args:
pareto_Y: A `n x m`-dim tensor of pareto optimal outcomes
Returns:
The hypervolume.
"""
if pareto_Y.shape[-1] != self._ref_point.shape[0]:
raise BotorchTensorDimensionError(
"pareto_Y must have the same number of objectives as ref_point. "
f"Got {pareto_Y.shape[-1]}, expected {self._ref_point.shape[0]}."
)
elif pareto_Y.ndim != 2:
raise BotorchTensorDimensionError(
f"pareto_Y must have exactly two dimensions, got {pareto_Y.ndim}."
)
# This assumes maximization, but internally flips the sign of the pareto front
# and the reference point and computes hypervolume for the minimization problem.
pareto_Y = -pareto_Y
better_than_ref = (pareto_Y <= self._ref_point).all(dim=-1)
pareto_Y = pareto_Y[better_than_ref]
# shift the pareto front so that reference point is all zeros
pareto_Y = pareto_Y - self._ref_point
self._initialize_multilist(pareto_Y)
bounds = torch.full_like(self._ref_point, float("-inf"))
return self._hv_recursive(
i=self._ref_point.shape[0] - 1, n_pareto=pareto_Y.shape[0], bounds=bounds
)
def _hv_recursive(self, i: int, n_pareto: int, bounds: Tensor) -> float:
r"""Recursive method for hypervolume calculation.
This assumes minimization (internally).
In contrast to the paper, this code assumes that the reference point
is the origin. This enables pruning a few operations.
Args:
i: objective index
n_pareto: number of pareto points
bounds: objective bounds
Returns:
The hypervolume.
"""
hvol = torch.tensor(0.0, dtype=bounds.dtype, device=bounds.device)
sentinel = self.list.sentinel
if n_pareto == 0:
# base case: one dimension
return hvol.item()
elif i == 0:
# base case: one dimension
return -sentinel.next[0].data[0].item()
elif i == 1:
# two dimensions, end recursion
q = sentinel.next[1]
h = q.data[0]
p = q.next[1]
while p is not sentinel:
hvol += h * (q.data[1] - p.data[1])
if p.data[0] < h:
h = p.data[0]
q = p
p = q.next[1]
hvol += h * q.data[1]
return hvol.item()
else:
p = sentinel
q = p.prev[i]
while q.data is not None:
if q.ignore < i:
q.ignore = 0
q = q.prev[i]
q = p.prev[i]
while n_pareto > 1 and (
q.data[i] > bounds[i] or q.prev[i].data[i] >= bounds[i]
):
p = q
self.list.remove(p, i, bounds)
q = p.prev[i]
n_pareto -= 1
q_prev = q.prev[i]
if n_pareto > 1:
hvol = q_prev.volume[i] + q_prev.area[i] * (q.data[i] - q_prev.data[i])
else:
q.area[0] = 1
q.area[1 : i + 1] = q.area[:i] * -(q.data[:i])
q.volume[i] = hvol
if q.ignore >= i:
q.area[i] = q_prev.area[i]
else:
q.area[i] = self._hv_recursive(i - 1, n_pareto, bounds)
if q.area[i] <= q_prev.area[i]:
q.ignore = i
while p is not sentinel:
p_data = p.data[i]
hvol += q.area[i] * (p_data - q.data[i])
bounds[i] = p_data
self.list.reinsert(p, i, bounds)
n_pareto += 1
q = p
p = p.next[i]
q.volume[i] = hvol
if q.ignore >= i:
q.area[i] = q.prev[i].area[i]
else:
q.area[i] = self._hv_recursive(i - 1, n_pareto, bounds)
if q.area[i] <= q.prev[i].area[i]:
q.ignore = i
hvol -= q.area[i] * q.data[i]
return hvol.item()
def _initialize_multilist(self, pareto_Y: Tensor) -> None:
r"""Sets up the multilist data structure needed for calculation.
Note: this assumes minimization.
Args:
pareto_Y: A `n x m`-dim tensor of pareto optimal objectives.
"""
m = pareto_Y.shape[-1]
nodes = [
Node(m=m, dtype=pareto_Y.dtype, device=pareto_Y.device, data=point)
for point in pareto_Y
]
self.list = MultiList(m=m, dtype=pareto_Y.dtype, device=pareto_Y.device)
for i in range(m):
sort_by_dimension(nodes, i)
self.list.extend(nodes, i)
[docs]
def sort_by_dimension(nodes: List[Node], i: int) -> None:
r"""Sorts the list of nodes in-place by the specified objective.
Args:
nodes: A list of Nodes
i: The index of the objective to sort by
"""
# build a list of tuples of (point[i], node)
decorated = [(node.data[i], index, node) for index, node in enumerate(nodes)]
# sort by this value
decorated.sort()
# write back to original list
nodes[:] = [node for (_, _, node) in decorated]
[docs]
class Node:
r"""Node in the MultiList data structure."""
def __init__(
self,
m: int,
dtype: torch.dtype,
device: torch.device,
data: Optional[Tensor] = None,
) -> None:
r"""Initialize MultiList.
Args:
m: The number of objectives
dtype: The dtype
device: The device
data: The tensor data to be stored in this Node.
"""
self.data = data
self.next = [None] * m
self.prev = [None] * m
self.ignore = 0
self.area = torch.zeros(m, dtype=dtype, device=device)
self.volume = torch.zeros_like(self.area)
[docs]
class MultiList:
r"""A special data structure used in hypervolume computation.
It consists of several doubly linked lists that share common nodes.
Every node has multiple predecessors and successors, one in every list.
"""
def __init__(self, m: int, dtype: torch.dtype, device: torch.device) -> None:
r"""Initialize `m` doubly linked lists.
Args:
m: number of doubly linked lists
dtype: the dtype
device: the device
"""
self.m = m
self.sentinel = Node(m=m, dtype=dtype, device=device)
self.sentinel.next = [self.sentinel] * m
self.sentinel.prev = [self.sentinel] * m
[docs]
def append(self, node: Node, index: int) -> None:
r"""Appends a node to the end of the list at the given index.
Args:
node: the new node
index: the index where the node should be appended.
"""
last = self.sentinel.prev[index]
node.next[index] = self.sentinel
node.prev[index] = last
# set the last element as the new one
self.sentinel.prev[index] = node
last.next[index] = node
[docs]
def extend(self, nodes: List[Node], index: int) -> None:
r"""Extends the list at the given index with the nodes.
Args:
nodes: list of nodes to append at the given index.
index: the index where the nodes should be appended.
"""
for node in nodes:
self.append(node=node, index=index)
[docs]
def remove(self, node: Node, index: int, bounds: Tensor) -> Node:
r"""Removes and returns 'node' from all lists in [0, 'index'].
Args:
node: The node to remove
index: The upper bound on the range of indices
bounds: A `2 x m`-dim tensor bounds on the objectives
"""
for i in range(index):
predecessor = node.prev[i]
successor = node.next[i]
predecessor.next[i] = successor
successor.prev[i] = predecessor
bounds.data = torch.min(bounds, node.data)
return node
[docs]
def reinsert(self, node: Node, index: int, bounds: Tensor) -> None:
r"""Re-inserts the node at its original position.
Re-inserts the node at its original position in all lists in [0, 'index']
before it was removed. This method assumes that the next and previous
nodes of the node that is reinserted are in the list.
Args:
node: The node
index: The upper bound on the range of indices
bounds: A `2 x m`-dim tensor bounds on the objectives
"""
for i in range(index):
node.prev[i].next[i] = node
node.next[i].prev[i] = node
bounds.data = torch.min(bounds, node.data)
[docs]
class SubsetIndexCachingMixin:
"""A Mixin class that adds q-subset index computations and caching."""
def __init__(self):
"""Initializes the class with q_out = -1 and an empty q_subset_indices dict."""
self.q_out: int = -1
self.q_subset_indices: BufferDict[str, Tensor] = BufferDict()
[docs]
def compute_q_subset_indices(
self, q_out: int, device: torch.device
) -> BufferDict[str, Tensor]:
r"""Returns and caches a dict of indices equal to subsets of `{1, ..., q_out}`.
This means that consecutive calls to `self.compute_q_subset_indices` with
the same `q_out` do 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`.
Returns:
A dict that maps "q choose i" to all size-i subsets of `{1, ..., q_out}`.
"""
if q_out != self.q_out:
self.q_subset_indices = compute_subset_indices(q_out, device=device)
self.q_out = q_out
return self.q_subset_indices
[docs]
def compute_subset_indices(
q: int, device: Optional[torch.device] = None
) -> BufferDict[str, Tensor]:
r"""Compute all (2^q - 1) distinct subsets of {1, ..., `q`}.
Args:
q: An integer defininig the set {1, ..., `q`} whose subsets to compute.
Returns:
A dict that maps "q choose i" to all size-i subsets of {1, ..., `q_out`}.
"""
indices = torch.arange(q, dtype=torch.long, device=device)
return BufferDict(
{
f"q_choose_{i}": torch.tensor(
list(combinations(indices, i)), dtype=torch.long, device=device
)
for i in range(1, q + 1)
}
)
class NoisyExpectedHypervolumeMixin(CachedCholeskyMCSamplerMixin):
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,
prune_baseline: bool = False,
alpha: float = 0.0,
cache_pending: bool = True,
max_iep: int = 0,
incremental_nehvi: bool = True,
cache_root: bool = True,
marginalize_dim: Optional[int] = None,
):
"""Initialize a mixin that contains functions for the batched Pareto-frontier
partitioning used by the noisy hypervolume-improvement-based acquisition
functions, i.e. qNEHVI and qLogNEHVI.
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 box decomposition is
of the Pareto front is created for each MC sample, an operation that
scales as `O(n^m)` and thus becomes particularly costly for `m` > 2.
objective: The MCMultiOutputObjective under which the samples are
evaluated. Defaults to `IdentityMCMultiOutputObjective()`.
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.
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.
marginalize_dim: A batch dimension that should be marginalized. For example,
this is useful when using a batched fully Bayesian model.
"""
super().__init__(model=model, cache_root=cache_root, sampler=sampler)
if len(ref_point) < 2:
raise ValueError(
"NoisyExpectedHypervolumeMixin supports m>=2 outcomes "
f"but ref_point has length {len(ref_point)}, which is smaller than 2."
)
tkwargs = {"dtype": X_baseline.dtype, "device": X_baseline.device}
ref_point = torch.as_tensor(ref_point, **tkwargs)
self.register_buffer("ref_point", ref_point)
if X_baseline.ndim > 2:
raise UnsupportedError(
f"NoisyExpectedHypervolumeMixin 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=marginalize_dim,
)
self.alpha = alpha
self.q_in = -1
self.q_out = -1
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, **tkwargs),
)
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
# is this called twice, once here, once in MultiObjectiveMCAcquisitionFunction?
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.
"""
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)
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,
)
# compute feasibility indicator if there are constraints
if self.constraints is None or self.X_baseline.shape[0] == 0:
feas = None
else:
feas = compute_feasibility_indicator(
constraints=self.constraints, samples=samples
)
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 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 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:]
) # 2 x batch_shape x sample_shape x num_cells x m
self.register_buffer("cell_lower_bounds", cell_bounds[0])
self.register_buffer("cell_upper_bounds", cell_bounds[1])
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 `sample_shape`-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)
)
