#!/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"""
Utilities for MC and qMC sampling.
References
.. [Trikalinos2014polytope]
T. A. Trikalinos and G. van Valkenhoef. Efficient sampling from uniform
density n-polytopes. Technical report, Brown University, 2014.
"""
from __future__ import annotations
import warnings
from abc import ABC, abstractmethod
from collections.abc import Callable, Generator, Iterable
from contextlib import contextmanager
from typing import TYPE_CHECKING
import numpy as np
import numpy.typing as npt
import scipy
import torch
from botorch.exceptions.errors import BotorchError
from botorch.exceptions.warnings import UserInputWarning
from botorch.sampling.qmc import NormalQMCEngine
from botorch.utils.transforms import unnormalize
from scipy.spatial import Delaunay, HalfspaceIntersection
from torch import LongTensor, Tensor
from torch.quasirandom import SobolEngine
if TYPE_CHECKING:
from botorch.models.deterministic import ( # pragma: no cover
GenericDeterministicModel,
)
[docs]
@contextmanager
def manual_seed(seed: int | None = None) -> Generator[None, None, None]:
r"""Contextmanager for manual setting the torch.random seed.
Args:
seed: The seed to set the random number generator to.
Returns:
Generator
Example:
>>> with manual_seed(1234):
>>> X = torch.rand(3)
"""
old_state = torch.random.get_rng_state()
try:
if seed is not None:
torch.random.manual_seed(seed)
yield
finally:
if seed is not None:
torch.random.set_rng_state(old_state)
[docs]
def draw_sobol_samples(
bounds: Tensor,
n: int,
q: int,
batch_shape: Iterable[int] | torch.Size | None = None,
seed: int | None = None,
) -> Tensor:
r"""Draw qMC samples from the box defined by bounds.
Args:
bounds: A `2 x d` dimensional tensor specifying box constraints on a
`d`-dimensional space, where bounds[0, :] and bounds[1, :] correspond
to lower and upper bounds, respectively.
n: The number of (q-batch) samples. As a best practice, use powers of 2.
q: The size of each q-batch.
batch_shape: The batch shape of the samples. If given, returns samples
of shape `n x batch_shape x q x d`, where each batch is an
`n x q x d`-dim tensor of qMC samples.
seed: The seed used for initializing Owen scrambling. If None (default),
use a random seed.
Returns:
A `n x batch_shape x q x d`-dim tensor of qMC samples from the box
defined by bounds.
Example:
>>> bounds = torch.stack([torch.zeros(3), torch.ones(3)])
>>> samples = draw_sobol_samples(bounds, 16, 2)
"""
batch_shape = batch_shape or torch.Size()
batch_size = int(torch.prod(torch.tensor(batch_shape)))
d = bounds.shape[-1]
lower = bounds[0]
rng = bounds[1] - bounds[0]
sobol_engine = SobolEngine(q * d, scramble=True, seed=seed)
samples_raw = sobol_engine.draw(batch_size * n, dtype=lower.dtype)
samples_raw = samples_raw.view(*batch_shape, n, q, d).to(device=lower.device)
if batch_shape != torch.Size():
samples_raw = samples_raw.permute(-3, *range(len(batch_shape)), -2, -1)
return lower + rng * samples_raw
[docs]
def draw_sobol_normal_samples(
d: int,
n: int,
device: torch.device | None = None,
dtype: torch.dtype | None = None,
seed: int | None = None,
) -> Tensor:
r"""Draw qMC samples from a multi-variate standard normal N(0, I_d).
A primary use-case for this functionality is to compute an QMC average
of f(X) over X where each element of X is drawn N(0, 1).
Args:
d: The dimension of the normal distribution.
n: The number of samples to return. As a best practice, use powers of 2.
device: The torch device.
dtype: The torch dtype.
seed: The seed used for initializing Owen scrambling. If None (default),
use a random seed.
Returns:
A tensor of qMC standard normal samples with dimension `n x d` with device
and dtype specified by the input.
Example:
>>> samples = draw_sobol_normal_samples(2, 16)
"""
normal_qmc_engine = NormalQMCEngine(d=d, seed=seed, inv_transform=True)
samples = normal_qmc_engine.draw(n, dtype=dtype)
return samples.to(device=device)
[docs]
def sample_hypersphere(
d: int,
n: int = 1,
qmc: bool = False,
seed: int | None = None,
device: torch.device | None = None,
dtype: torch.dtype | None = None,
) -> Tensor:
r"""Sample uniformly from a unit d-sphere.
Args:
d: The dimension of the hypersphere.
n: The number of samples to return.
qmc: If True, use QMC Sobol sampling (instead of i.i.d. uniform).
seed: If provided, use as a seed for the RNG.
device: The torch device.
dtype: The torch dtype.
Returns:
An `n x d` tensor of uniform samples from from the d-hypersphere.
Example:
>>> sample_hypersphere(d=5, n=10)
"""
if d == 1:
rnd = torch.randint(0, 2, (n, 1), device=device, dtype=dtype)
return 2 * rnd - 1
if qmc:
rnd = draw_sobol_normal_samples(d=d, n=n, device=device, dtype=dtype, seed=seed)
else:
with manual_seed(seed=seed):
rnd = torch.randn(n, d, dtype=dtype)
samples = rnd / torch.linalg.norm(rnd, dim=-1, keepdim=True)
if device is not None:
samples = samples.to(device)
return samples
[docs]
def sample_simplex(
d: int,
n: int = 1,
qmc: bool = False,
seed: int | None = None,
device: torch.device | None = None,
dtype: torch.dtype | None = None,
) -> Tensor:
r"""Sample uniformly from a d-simplex.
Args:
d: The dimension of the simplex.
n: The number of samples to return.
qmc: If True, use QMC Sobol sampling (instead of i.i.d. uniform).
seed: If provided, use as a seed for the RNG.
device: The torch device.
dtype: The torch dtype.
Returns:
An `n x d` tensor of uniform samples from from the d-simplex.
Example:
>>> sample_simplex(d=3, n=10)
"""
if d == 1:
return torch.ones(n, 1, device=device, dtype=dtype)
if qmc:
sobol_engine = SobolEngine(d - 1, scramble=True, seed=seed)
rnd = sobol_engine.draw(n, dtype=dtype)
else:
with manual_seed(seed=seed):
rnd = torch.rand(n, d - 1, dtype=dtype)
srnd, _ = torch.sort(rnd, dim=-1)
zeros = torch.zeros(n, 1, dtype=dtype)
ones = torch.ones(n, 1, dtype=dtype)
srnd = torch.cat([zeros, srnd, ones], dim=-1)
if device is not None:
srnd = srnd.to(device)
return srnd[..., 1:] - srnd[..., :-1]
[docs]
def sample_polytope(
A: Tensor,
b: Tensor,
x0: Tensor,
n: int = 10000,
n0: int = 100,
n_thinning: int = 1,
seed: int | None = None,
) -> Tensor:
r"""
Hit and run sampler from uniform sampling points from a polytope,
described via inequality constraints A*x<=b.
Args:
A: A `m x d`-dim Tensor describing inequality constraints
so that all samples satisfy `Ax <= b`.
b: A `m`-dim Tensor describing the inequality constraints
so that all samples satisfy `Ax <= b`.
x0: A `d`-dim Tensor representing a starting point of the chain
satisfying the constraints.
n: The number of resulting samples kept in the output.
n0: The number of burn-in samples. The chain will produce
n+n0 samples but the first n0 samples are not saved.
n_thinning: The amount of thinnning. This function will return every
`n_thinning`-th sample from the chain (after burn-in).
seed: The seed for the sampler. If omitted, use a random seed.
Returns:
(n, d) dim Tensor containing the resulting samples.
"""
# Check that starting point satisfies the constraints.
if not ((slack := A @ x0 - b) <= 0).all():
raise ValueError(
f"Starting point does not satisfy the constraints. Inputs: {A=},"
f"{b=}, {x0=}, A@x0-b={slack}."
)
# Remove rows where all elements of A are 0. This avoids nan and infs later.
# A may have zero rows in it when this is called from PolytopeSampler
# with equality constraints (which are absorbed into A & b).
non_zero_rows = torch.any(A != 0, dim=-1)
A = A[non_zero_rows]
b = b[non_zero_rows]
n_tot = n0 + n * n_thinning
seed = seed if seed is not None else torch.randint(0, 1000000, (1,)).item()
with manual_seed(seed=seed):
rands = torch.rand(n_tot, dtype=A.dtype, device=A.device)
# Sample uniformly from unit hypersphere in d dims.
# Increment seed by +1 to avoid correlation with step size, see #2156 for details.
Rs = sample_hypersphere(
d=x0.shape[0], n=n_tot, dtype=A.dtype, device=A.device, seed=seed + 1
).unsqueeze(-1)
# Use batch operations for matrix multiplication.
ARs = (A @ Rs).squeeze(-1)
out = torch.empty(n, A.size(-1), dtype=A.dtype, device=A.device)
x = x0.clone()
large_constant = torch.finfo().max
for i, (ar, r, rnd) in enumerate(zip(ARs, Rs, rands)):
# Given x, the next point in the chain is x+alpha*r.
# It must satisfy A(x+alpha*r)<=b, which implies A*alpha*r<=b-Ax,
# so alpha<=(b-Ax)/ar for ar>0, and alpha>=(b-Ax)/ar for ar<0.
# If x is at the boundary, b - Ax = 0. If ar > 0, then we must
# have alpha <= 0. If ar < 0, we must have alpha >= 0.
# ar == 0 is an unlikely event that provides no signal.
# b - A @ x is always >= 0, clamping for numerical tolerances.
w = (b - A @ x).squeeze().clamp(min=0.0) / ar
# Find upper bound for alpha. If there are no constraints on
# the upper bound of alpha, set it to a large value.
pos = w > 0
alpha_max = w[pos].min().item() if pos.any() else large_constant
# Find lower bound for alpha.
neg = w < 0
alpha_min = w[neg].max().item() if neg.any() else -large_constant
# Handle the boundary case.
if (w_eq_0 := (w == 0)).any():
# If ar > 0 at the boundary, alpha <= 0.
if w_eq_0.logical_and(ar > 0).any():
alpha_max = min(alpha_max, 0.0)
# If ar < 0 at the boundary, alpha >= 0.
if w_eq_0.logical_and(ar < 0).any():
alpha_min = max(alpha_min, 0.0)
# alpha ~ Uniform[alpha_min, alpha_max]
alpha = alpha_min + rnd * (alpha_max - alpha_min)
x = x + alpha * r
if (k := i - n0) >= 0: # save samples after burn-in period
idx, rem = divmod(k, n_thinning)
if rem == 0:
out[idx] = x.squeeze()
return out
[docs]
def batched_multinomial(
weights: Tensor,
num_samples: int,
replacement: bool = False,
generator: torch.Generator | None = None,
out: Tensor | None = None,
) -> LongTensor:
r"""Sample from multinomial with an arbitrary number of batch dimensions.
Args:
weights: A `batch_shape x num_categories` tensor of weights. For each batch
index `i, j, ...`, this functions samples from a multinomial with `input`
`weights[i, j, ..., :]`. Note that the weights need not sum to one, but must
be non-negative, finite and have a non-zero sum.
num_samples: The number of samples to draw for each batch index. Must be smaller
than `num_categories` if `replacement=False`.
replacement: If True, samples are drawn with replacement.
generator: A a pseudorandom number generator for sampling.
out: The output tensor (optional). If provided, must be of size
`batch_shape x num_samples`.
Returns:
A `batch_shape x num_samples` tensor of samples.
This is a thin wrapper around `torch.multinomial` that allows weight (`input`)
tensors with an arbitrary number of batch dimensions (`torch.multinomial` only
allows a single batch dimension). The calling signature is the same as for
`torch.multinomial`.
Example:
>>> weights = torch.rand(2, 3, 10)
>>> samples = batched_multinomial(weights, 4) # shape is 2 x 3 x 4
"""
batch_shape, n_categories = weights.shape[:-1], weights.size(-1)
flat_samples = torch.multinomial(
input=weights.view(-1, n_categories),
num_samples=num_samples,
replacement=replacement,
generator=generator,
out=None if out is None else out.view(-1, num_samples),
)
return flat_samples.view(*batch_shape, num_samples)
def _convert_bounds_to_inequality_constraints(bounds: Tensor) -> tuple[Tensor, Tensor]:
r"""Convert bounds into inequality constraints of the form Ax <= b.
Args:
bounds: A `2 x d`-dim tensor of bounds
Returns:
A two-element tuple containing
- A: A `2d x d`-dim tensor of coefficients
- b: A `2d x 1`-dim tensor containing the right hand side
"""
d = bounds.shape[-1]
eye = torch.eye(d, dtype=bounds.dtype, device=bounds.device)
lower, upper = bounds
lower_finite, upper_finite = bounds.isfinite()
A = torch.cat([-eye[lower_finite], eye[upper_finite]], dim=0)
b = torch.cat([-lower[lower_finite], upper[upper_finite]], dim=0).unsqueeze(-1)
return A, b
[docs]
def find_interior_point(
A: npt.NDArray,
b: npt.NDArray,
A_eq: npt.NDArray | None = None,
b_eq: npt.NDArray | None = None,
) -> npt.NDArray:
r"""Find an interior point of a polytope via linear programming.
Args:
A: A `n_ineq x d`-dim numpy array containing the coefficients of the
constraint inequalities.
b: A `n_ineq x 1`-dim numpy array containing the right hand sides of
the constraint inequalities.
A_eq: A `n_eq x d`-dim numpy array containing the coefficients of the
constraint equalities.
b_eq: A `n_eq x 1`-dim numpy array containing the right hand sides of
the constraint equalities.
Returns:
A `d`-dim numpy array containing an interior point of the polytope.
This function will raise a ValueError if there is no such point.
This method solves the following Linear Program:
min -s subject to A @ x <= b - 2 * s, s >= 0, A_eq @ x = b_eq
In case the polytope is unbounded, then it will also constrain the slack
variable `s` to `s<=1`.
"""
# augment inequality constraints: A @ (x, s) <= b
d = A.shape[-1]
ncon = A.shape[-2] + 1
c = np.zeros(d + 1)
c[-1] = -1
b_ub = np.zeros(ncon)
b_ub[:-1] = b.reshape(-1)
A_ub = np.zeros((ncon, d + 1))
A_ub[:-1, :-1] = A
A_ub[:-1, -1] = 2.0
A_ub[-1, -1] = -1.0
result = scipy.optimize.linprog(
c=c,
A_ub=A_ub,
b_ub=b_ub,
A_eq=A_eq,
b_eq=b_eq,
bounds=(None, None),
method="highs",
)
if result.status == 3:
# problem is unbounded - to find a bounded solution we constrain the
# slack variable `s` to `s <= 1.0`.
A_s = np.concatenate([np.zeros((1, d)), np.ones((1, 1))], axis=-1)
A_ub = np.concatenate([A_ub, A_s], axis=0)
b_ub = np.concatenate([b_ub, np.ones(1)], axis=-1)
result = scipy.optimize.linprog(
c=c,
A_ub=A_ub,
b_ub=b_ub,
A_eq=A_eq,
b_eq=b_eq,
bounds=(None, None),
method="highs",
)
if result.status == 2:
raise ValueError(
"No feasible point found. Constraint polytope appears empty. "
+ "Check your constraints."
)
elif result.status > 0:
raise ValueError(
"Problem checking constraint specification. "
+ f"linprog status: {result.message}"
)
# the x in the result is really (x, s)
return result.x[:-1]
[docs]
class PolytopeSampler(ABC):
"""Base class for samplers that sample points from a polytope."""
def __init__(
self,
inequality_constraints: tuple[Tensor, Tensor] | None = None,
equality_constraints: tuple[Tensor, Tensor] | None = None,
bounds: Tensor | None = None,
interior_point: Tensor | None = None,
) -> None:
r"""
Args:
inequality_constraints: Tensors `(A, b)` describing inequality
constraints `A @ x <= b`, where `A` is a `n_ineq_con x d`-dim
Tensor and `b` is a `n_ineq_con x 1`-dim Tensor, with `n_ineq_con`
the number of inequalities and `d` the dimension of the sample space.
equality_constraints: Tensors `(C, d)` describing the equality constraints
`C @ x = d`, where `C` is a `n_eq_con x d`-dim Tensor and `d` is a
`n_eq_con x 1`-dim Tensor with `n_eq_con` the number of equalities.
bounds: A `2 x d`-dim tensor of box bounds, where `inf` (`-inf`) means
that the respective dimension is unbounded above (below).
interior_point: A `d x 1`-dim Tensor presenting a point in the
(relative) interior of the polytope. If omitted, determined
automatically by solving a Linear Program.
"""
if inequality_constraints is None:
if bounds is None:
raise BotorchError(
"PolytopeSampler requires either inequality constraints or bounds."
)
A = torch.empty(
0, bounds.shape[-1], dtype=bounds.dtype, device=bounds.device
)
b = torch.empty(0, 1, dtype=bounds.dtype, device=bounds.device)
else:
A, b = inequality_constraints
if bounds is not None:
# add inequality constraints for bounds
# TODO: make sure there are not deduplicate constraints
A2, b2 = _convert_bounds_to_inequality_constraints(bounds=bounds)
A = torch.cat([A, A2], dim=0)
b = torch.cat([b, b2], dim=0)
self.A = A
self.b = b
self.equality_constraints = equality_constraints
if equality_constraints is not None:
self.C, self.d = equality_constraints
U, S, Vh = torch.linalg.svd(self.C)
r = torch.nonzero(S).size(0) # rank of matrix C
self.nullC = Vh[r:, :].transpose(-1, -2) # orthonormal null space of C,
# satisfying # C @ nullC = 0 and nullC.T @ nullC = I
# using the change of variables x=x0+nullC*y,
# sample y satisfies A*nullC*y<=b-A*x0.
# the linear constraint is automatically satisfied as x0 satisfies it.
else:
self.C = None
self.d = None
self.nullC = torch.eye(
self.A.size(-1), dtype=self.A.dtype, device=self.A.device
)
self.new_A = self.A @ self.nullC # doesn't depend on the initial point
# initial point for the original, not transformed, problem
if interior_point is not None:
if self.feasible(interior_point):
self.x0 = interior_point
else:
raise ValueError("The given input point is not feasible.")
else:
self.x0 = self.find_interior_point()
[docs]
def feasible(self, x: Tensor) -> bool:
r"""Check whether a point is contained in the polytope.
Args:
x: A `d x 1`-dim Tensor.
Returns:
True if `x` is contained inside the polytope (incl. its boundary),
False otherwise.
"""
ineq = (self.A @ x - self.b <= 0).all()
if self.equality_constraints is not None:
eq = (self.C @ x - self.d == 0).all()
return ineq & eq
return ineq
[docs]
def find_interior_point(self) -> Tensor:
r"""Find an interior point of the polytope.
Returns:
A `d x 1`-dim Tensor representing a point contained in the polytope.
This function will raise a ValueError if there is no such point.
"""
if self.equality_constraints:
# equality constraints: A_eq * (x, s) = b_eq
A_eq = np.zeros((self.C.size(0), self.C.size(-1) + 1))
A_eq[:, :-1] = self.C.cpu().numpy()
b_eq = self.d.cpu().numpy()
else:
A_eq = None
b_eq = None
x0 = find_interior_point(
A=self.A.cpu().numpy(), b=self.b.cpu().numpy(), A_eq=A_eq, b_eq=b_eq
)
return torch.from_numpy(x0).to(self.A).unsqueeze(-1)
# -------- Abstract methods to be implemented by subclasses -------- #
[docs]
@abstractmethod
def draw(self, n: int = 1) -> Tensor:
r"""Draw samples from the polytope.
Args:
n: The number of samples.
Returns:
A `n x d` Tensor of samples from the polytope.
"""
pass # pragma: no cover
[docs]
class HitAndRunPolytopeSampler(PolytopeSampler):
r"""A sampler for sampling from a polyope using a hit-and-run algorithm."""
def __init__(
self,
inequality_constraints: tuple[Tensor, Tensor] | None = None,
equality_constraints: tuple[Tensor, Tensor] | None = None,
bounds: Tensor | None = None,
interior_point: Tensor | None = None,
n_burnin: int = 200,
n_thinning: int = 20,
seed: int | None = None,
) -> None:
r"""A sampler for sampling from a polyope using a hit-and-run algorithm.
Args:
inequality_constraints: Tensors `(A, b)` describing inequality
constraints `A @ x <= b`, where `A` is a `n_ineq_con x d`-dim
Tensor and `b` is a `n_ineq_con x 1`-dim Tensor, with `n_ineq_con`
the number of inequalities and `d` the dimension of the sample space.
equality_constraints: Tensors `(C, d)` describing the equality constraints
`C @ x = d`, where `C` is a `n_eq_con x d`-dim Tensor and `d` is a
`n_eq_con x 1`-dim Tensor with `n_eq_con` the number of equalities.
bounds: A `2 x d`-dim tensor of box bounds, where `inf` (`-inf`) means
that the respective dimension is unbounded from above (below). If
omitted, no bounds (in addition to the above constraints) are applied.
interior_point: A `d x 1`-dim Tensor representing a point in the
(relative) interior of the polytope. If omitted, determined
automatically by solving a Linear Program.
n_burnin: The number of burn in samples. The sampler will discard
`n_burnin` samples before returning the first sample.
n_thinning: The amount of thinning. The sampler will return every
`n_thinning` sample (after burn-in). This may need to be increased
for sets of constraints that are difficult to satisfy (i.e. in which
case the volume of the constraint polytope is small relative to that
of its bounding box).
seed: The random seed.
"""
if inequality_constraints is None and bounds is None:
raise BotorchError(
"HitAndRunPolytopeSampler requires either inequality constraints "
"or bounds."
)
# Normalize constraints to avoid the following issue:
# https://github.com/pytorch/botorch/issues/1225
offset, scale = None, None
if inequality_constraints or equality_constraints:
if bounds is None:
warnings.warn(
"HitAndRunPolytopeSampler did not receive `bounds`, which can "
"lead to non-uniform sampling if the parameter ranges are very "
"different (see https://github.com/pytorch/botorch/issues/1225).",
UserInputWarning,
stacklevel=3,
)
else:
if inequality_constraints:
inequality_constraints = normalize_dense_linear_constraints(
bounds=bounds, constraints=inequality_constraints
)
if equality_constraints:
equality_constraints = normalize_dense_linear_constraints(
bounds=bounds, constraints=equality_constraints
)
lower, upper = bounds
offset = lower
scale = upper - lower
if interior_point is not None:
# If provided, we also need to normalize the interior point
interior_point = (interior_point - offset[:, None]) / scale[:, None]
bounds = torch.zeros_like(bounds)
bounds[1, :] = 1.0
super().__init__(
inequality_constraints=inequality_constraints,
equality_constraints=equality_constraints,
bounds=bounds,
interior_point=interior_point,
)
self.n_burnin: int = n_burnin
self.n_thinning: int = n_thinning
self.num_samples_generated: int = 0
self._seed: int | None = seed
self._offset: Tensor | None = offset
self._scale: Tensor | None = scale
[docs]
def draw(self, n: int = 1) -> Tensor:
r"""Draw samples from the polytope.
Args:
n: The number of samples.
Returns:
A `n x d` Tensor of samples from the polytope.
"""
# There are two layers of normalization. In the outer layer, the space
# has been normalized to the unit cube. In the inner layer, we remove
# any equality constraints and sample on the subspace defined by those
# equality constraints, with an additional shift to normalize the interior
# point to the origin. Below, after sampling in that inner layer, we have
# to reverse both layers of normalization.
transformed_samples = sample_polytope(
# Run this on the cpu since there is a lot of looping going on
A=self.new_A.cpu(),
b=(self.b - self.A @ self.x0).cpu(),
x0=torch.zeros(
(self.nullC.size(1), 1), dtype=self.A.dtype, device=torch.device("cpu")
),
n=n,
n0=self.n_burnin if self.num_samples_generated == 0 else 0,
n_thinning=self.n_thinning,
seed=self._seed,
).to(self.b)
# Update the seed for the next call in a deterministic fashion
if self._seed is not None:
self._seed += n
# Unnormalize the inner layer
init_shift = self.x0.transpose(-1, -2)
samples = init_shift + transformed_samples @ self.nullC.transpose(-1, -2)
# Keep the last element as the beginning of the next chain
self.x0 = samples[-1].reshape(-1, 1)
# Unnormalize the outer layer
if self._scale is not None:
samples = self._offset + self._scale * samples
self.num_samples_generated += n
return samples
[docs]
class DelaunayPolytopeSampler(PolytopeSampler):
r"""A polytope sampler using Delaunay triangulation.
This sampler first enumerates the vertices of the constraint polytope and
then uses a Delaunay triangulation to tesselate its convex hull.
The sampling happens in two stages:
1. First, we sample from the set of hypertriangles generated by the
Delaunay triangulation (i.e. which hyper-triangle to draw the sample
from) with probabilities proportional to the triangle volumes.
2. Then, we sample uniformly from the chosen hypertriangle by sampling
uniformly from the unit simplex of the appropriate dimension, and
then computing the convex combination of the vertices of the
hypertriangle according to that draw from the simplex.
The best reference (not exactly the same, but functionally equivalent) is
[Trikalinos2014polytope]_. A simple R implementation is available at
https://github.com/gertvv/tesselample.
"""
def __init__(
self,
inequality_constraints: tuple[Tensor, Tensor] | None = None,
equality_constraints: tuple[Tensor, Tensor] | None = None,
bounds: Tensor | None = None,
interior_point: Tensor | None = None,
) -> None:
r"""Initialize DelaunayPolytopeSampler.
Args:
inequality_constraints: Tensors `(A, b)` describing inequality
constraints `A @ x <= b`, where `A` is a `n_ineq_con x d`-dim
Tensor and `b` is a `n_ineq_con x 1`-dim Tensor, with `n_ineq_con`
the number of inequalities and `d` the dimension of the sample space.
equality_constraints: Tensors `(C, d)` describing the equality constraints
`C @ x = d`, where `C` is a `n_eq_con x d`-dim Tensor and `d` is a
`n_eq_con x 1`-dim Tensor with `n_eq_con` the number of equalities.
bounds: A `2 x d`-dim tensor of box bounds, where `inf` (`-inf`) means
that the respective dimension is unbounded from above (below).
interior_point: A `d x 1`-dim Tensor representing a point in the
(relative) interior of the polytope. If omitted, determined
automatically by solving a Linear Program.
Warning: The vertex enumeration performed in this algorithm can become
extremely costly if there are a large number of inequalities. Similarly,
the triangulation can get very expensive in high dimensions. Only use
this algorithm for moderate dimensions / moderately complex constraint sets.
An alternative is the `HitAndRunPolytopeSampler`.
"""
super().__init__(
inequality_constraints=inequality_constraints,
equality_constraints=equality_constraints,
bounds=bounds,
interior_point=interior_point,
)
# shift coordinate system to be anchored at x0
new_b = self.b - self.A @ self.x0
if self.new_A.shape[-1] < 2:
# if the polytope is in dim 1 (i.e. a line segment) Qhull won't work
tshlds = new_b / self.new_A
neg = self.new_A < 0
self.y_min = tshlds[neg].max()
self.y_max = tshlds[~neg].min()
self.dim = 1
else:
# Qhull expects inputs of the form A @ x + b <= 0, so we need to negate here
halfspaces = torch.cat([self.new_A, -new_b], dim=-1).cpu().numpy()
vertices = HalfspaceIntersection(
halfspaces=halfspaces, interior_point=np.zeros(self.new_A.shape[-1])
).intersections
self.dim = vertices.shape[-1]
try:
delaunay = Delaunay(vertices)
except ValueError as e:
if "Points cannot contain NaN" in str(e):
raise ValueError("Polytope is unbounded.")
raise e # pragma: no cover
polytopes = torch.from_numpy(
np.array([delaunay.points[s] for s in delaunay.simplices]),
).to(self.A)
volumes = torch.stack([torch.det(p[1:] - p[0]).abs() for p in polytopes])
self._polytopes = polytopes
self._p = volumes / volumes.sum()
[docs]
def draw(self, n: int = 1, seed: int | None = None) -> Tensor:
r"""Draw samples from the polytope.
Args:
n: The number of samples.
seed: The random seed.
Returns:
A `n x d` Tensor of samples from the polytope.
"""
if self.dim == 1:
with manual_seed(seed):
e = torch.rand(n, 1, device=self.new_A.device, dtype=self.new_A.dtype)
transformed_samples = self.y_min + (self.y_max - self.y_min) * e
else:
if seed is None:
generator = None
else:
generator = torch.Generator(device=self.A.device)
generator.manual_seed(seed)
index_rvs = torch.multinomial(
self._p,
num_samples=n,
replacement=True,
generator=generator,
)
simplex_rvs = sample_simplex(
d=self.dim + 1, n=n, seed=seed, device=self.A.device, dtype=self.A.dtype
)
transformed_samples = torch.stack(
[rv @ self._polytopes[idx] for rv, idx in zip(simplex_rvs, index_rvs)]
)
init_shift = self.x0.transpose(-1, -2)
samples = init_shift + transformed_samples @ self.nullC.transpose(-1, -2)
return samples
[docs]
def normalize_sparse_linear_constraints(
bounds: Tensor, constraints: list[tuple[Tensor, Tensor, float]]
) -> list[tuple[Tensor, Tensor, float]]:
r"""Normalize sparse linear constraints to the unit cube.
Args:
bounds: A `2 x d`-dim tensor containing the box bounds.
constraints: A list of tuples (`indices`, `coefficients`, `rhs`), with
`indices` and `coefficients` one-dimensional tensors and `rhs` a
scalar, where each tuple encodes an inequality constraint of the form
`\sum_i (X[indices[i]] * coefficients[i]) >= rhs` or
`\sum_i (X[indices[i]] * coefficients[i]) = rhs`.
"""
new_constraints = []
for index, coefficient, rhs in constraints:
if index.ndim != 1:
raise ValueError(
"`indices` must be a one-dimensional tensor. This method does not "
"support the kind of 'inter-point constraints' that are supported by "
"`optimize_acqf()`. To achieve this behavior, you need define the "
"problem on the joint space over `q` points and impose use constraints,"
"see https://github.com/pytorch/botorch/issues/2468#issuecomment-2287706461" # noqa: E501
)
lower, upper = bounds[:, index]
s = upper - lower
new_constraints.append(
(index, s * coefficient, (rhs - torch.dot(coefficient, lower)).item())
)
return new_constraints
[docs]
def normalize_dense_linear_constraints(
bounds: Tensor,
constraints: tuple[Tensor, Tensor],
) -> tuple[Tensor, Tensor]:
r"""Normalize dense linear constraints to the unit cube.
Args:
bounds: A `2 x d`-dim tensor containing the box bounds.
constraints: A tensor tuple `(A, b)` describing constraints
`A @ x (<)= b`, where `A` is a `n_con x d`-dim Tensor and
`b` is a `n_con x 1`-dim Tensor, with `n_con` the number of
constraints and `d` the dimension of the sample space.
Returns:
A tensor tuple `(A_nlz, b_nlz)` of normalized constraints.
"""
lower, upper = bounds
A, b = constraints
A_nlz = (upper - lower) * A
b_nlz = b - (A @ lower).unsqueeze(-1)
return A_nlz, b_nlz
[docs]
def get_polytope_samples(
n: int,
bounds: Tensor,
inequality_constraints: list[tuple[Tensor, Tensor, float]] | None = None,
equality_constraints: list[tuple[Tensor, Tensor, float]] | None = None,
seed: int | None = None,
n_burnin: int = 10_000,
n_thinning: int = 32,
) -> Tensor:
r"""Sample from polytope defined by box bounds and (in)equality constraints.
This uses a hit-and-run Markov chain sampler.
NOTE: Much of the functionality of this method has been moved into
`HitAndRunPolytopeSampler`. If you want to repeatedly draw samples, you should
use `HitAndRunPolytopeSampler` directly in order to avoid repeatedly running
a burn-in of the chain. To do so, you need to convert the sparse constraint
format that `get_polytope_samples` expects to the dense constraint format that
`HitAndRunPolytopeSampler` expects. This can be done via the
`sparse_to_dense_constraints` method (but remember to adjust the constraint
from the `Ax >= b` format expecxted here to the `Ax <= b` format expected by
`PolytopeSampler` by multiplying both `A` and `b` by -1.)
NOTE: This method does not support the kind of "inter-point constraints" that
are supported by `optimize_acqf()`. To achieve this behavior, you need define the
problem on the joint space over `q` points and impose use constraints, see:
https://github.com/pytorch/botorch/issues/2468#issuecomment-2287706461
Args:
n: The number of samples.
bounds: A `2 x d`-dim tensor containing the box bounds.
inequality_constraints: A list of tuples (`indices`, `coefficients`, `rhs`),
with `indices` and `coefficients` one-dimensional tensors and `rhs` a
scalar, where each tuple encodes an inequality constraint of the form
`\sum_i (X[indices[i]] * coefficients[i]) >= rhs`.
equality_constraints: A list of tuples (`indices`, `coefficients`, `rhs`),
with `indices` and `coefficients` one-dimensional tensors and `rhs` a
scalar, where each tuple encodes an equality constraint of the form
`\sum_i (X[indices[i]] * coefficients[i]) = rhs`.
seed: The random seed.
n_burnin: The number of burn-in samples for the Markov chain sampler.
n_thinning: The amount of thinnning. This function will return every
`n_thinning`-th sample from the chain (after burn-in).
Returns:
A `n x d`-dim tensor of samples.
"""
if inequality_constraints:
A, b = sparse_to_dense_constraints(
d=bounds.shape[-1],
constraints=inequality_constraints,
)
# Note that the inequality constraints are of the form Ax >= b,
# but PolytopeSampler expects inequality constraints of the
# form Ax <= b, so we multiply by -1 below.
dense_inequality_constraints = (-A, -b)
else:
dense_inequality_constraints = None
if equality_constraints:
dense_equality_constraints = sparse_to_dense_constraints(
d=bounds.shape[-1], constraints=equality_constraints
)
else:
dense_equality_constraints = None
polytope_sampler = HitAndRunPolytopeSampler(
bounds=bounds,
inequality_constraints=dense_inequality_constraints,
equality_constraints=dense_equality_constraints,
n_burnin=n_burnin,
n_thinning=n_thinning,
seed=seed,
)
return polytope_sampler.draw(n=n)
[docs]
def sparse_to_dense_constraints(
d: int,
constraints: list[tuple[Tensor, Tensor, float]],
) -> tuple[Tensor, Tensor]:
r"""Convert parameter constraints from a sparse format into a dense format.
This method converts sparse triples of the form (indices, coefficients, rhs)
to constraints of the form Ax >= b or Ax = b.
Args:
d: The input dimension.
constraints: A list of tuples (`indices`, `coefficients`, `rhs`),
with `indices` and `coefficients` one-dimensional tensors and `rhs` a
scalar, where each tuple encodes an (in)equality constraint of the form
`\sum_i (X[indices[i]] * coefficients[i]) >= rhs` or
`\sum_i (X[indices[i]] * coefficients[i]) = rhs`.
Returns:
A two-element tuple containing:
- A: A `n_constraints x d`-dim tensor of coefficients.
- b: A `n_constraints x 1`-dim tensor of right hand sides.
"""
_t = constraints[0][1]
A = torch.zeros(len(constraints), d, dtype=_t.dtype, device=_t.device)
b = torch.zeros(len(constraints), 1, dtype=_t.dtype, device=_t.device)
for i, (indices, coefficients, rhs) in enumerate(constraints):
A[i, indices.long()] = coefficients
b[i] = rhs
return A, b
[docs]
def optimize_posterior_samples(
paths: GenericDeterministicModel,
bounds: Tensor,
raw_samples: int = 1024,
num_restarts: int = 20,
sample_transform: Callable[[Tensor], Tensor] | None = None,
return_transformed: bool = False,
) -> tuple[Tensor, Tensor]:
r"""Cheaply maximizes posterior samples by random querying followed by
gradient-based optimization using SciPy's L-BFGS-B routine.
Args:
paths: Random Fourier Feature-based sample paths from the GP
bounds: The bounds on the search space.
raw_samples: The number of samples with which to query the samples initially.
num_restarts: The number of points selected for gradient-based optimization.
sample_transform: A callable transform of the sample outputs (e.g.
MCAcquisitionObjective or ScalarizedPosteriorTransform.evaluate) used to
negate the objective or otherwise transform the output.
return_transformed: A boolean indicating whether to return the transformed
or non-transformed samples.
Returns:
A two-element tuple containing:
- X_opt: A `num_optima x [batch_size] x d`-dim tensor of optimal inputs x*.
- f_opt: A `num_optima x [batch_size] x m`-dim, optionally
`num_optima x [batch_size] x 1`-dim, tensor of optimal outputs f*.
"""
def path_func(x) -> Tensor:
res = paths(x)
if sample_transform:
res = sample_transform(res)
return res.squeeze(-1)
candidate_set = unnormalize(
SobolEngine(dimension=bounds.shape[1], scramble=True).draw(n=raw_samples),
bounds=bounds,
)
# queries all samples on all candidates - output shape
# raw_samples * num_optima * num_models
candidate_queries = path_func(candidate_set)
argtop_k = torch.topk(candidate_queries, num_restarts, dim=-1).indices
X_top_k = candidate_set[argtop_k, :]
# to avoid circular import, the import occurs here
from botorch.generation.gen import gen_candidates_scipy
X_top_k, f_top_k = gen_candidates_scipy(
X_top_k,
path_func,
lower_bounds=bounds[0],
upper_bounds=bounds[1],
)
f_opt, arg_opt = f_top_k.max(dim=-1, keepdim=True)
# For each sample (and possibly for every model in the batch of models), this
# retrieves the argmax. We flatten, pick out the indices and then reshape to
# the original batch shapes (so instead of pickig out the argmax of a
# (3, 7, num_restarts, D)) along the num_restarts dim, we pick it out of a
# (21, num_restarts, D)
final_shape = candidate_queries.shape[:-1]
X_opt = X_top_k.reshape(final_shape.numel(), num_restarts, -1)[
torch.arange(final_shape.numel()), arg_opt.flatten()
].reshape(*final_shape, -1)
# if we return transformed, we do not need to pass the samples through paths
# paths a second time but rather just return the transformed optimal values
if return_transformed:
return X_opt, f_opt
f_opt = paths(X_opt.unsqueeze(-2)).squeeze(-2)
return X_opt, f_opt
