#!/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.
# pyre-strict
from dataclasses import dataclass
from logging import Logger
from typing import Any, Callable, cast, Dict, List, Optional, Tuple, Type
import numpy as np
import torch
from ax.exceptions.core import UnsupportedError
from ax.exceptions.model import ModelError
from ax.models.model_utils import filter_constraints_and_fixed_features, get_observed
from ax.models.random.sobol import SobolGenerator
from ax.models.types import TConfig
from ax.utils.common.constants import Keys
from ax.utils.common.logger import get_logger
from botorch.acquisition.acquisition import AcquisitionFunction
from botorch.acquisition.analytic import PosteriorMean
from botorch.acquisition.fixed_feature import FixedFeatureAcquisitionFunction
from botorch.acquisition.monte_carlo import (
qSimpleRegret,
SampleReducingMCAcquisitionFunction,
)
from botorch.acquisition.multi_objective.analytic import (
MultiObjectiveAnalyticAcquisitionFunction,
)
from botorch.acquisition.multi_objective.monte_carlo import (
MultiObjectiveMCAcquisitionFunction,
)
from botorch.acquisition.multi_objective.multi_output_risk_measures import (
MARS,
MultiOutputRiskMeasureMCObjective,
)
from botorch.acquisition.multi_objective.objective import (
IdentityMCMultiOutputObjective,
WeightedMCMultiOutputObjective,
)
from botorch.acquisition.objective import (
ConstrainedMCObjective,
GenericMCObjective,
IdentityMCObjective,
LinearMCObjective,
MCAcquisitionObjective,
PosteriorTransform,
ScalarizedPosteriorTransform,
)
from botorch.acquisition.risk_measures import RiskMeasureMCObjective
from botorch.acquisition.utils import get_infeasible_cost
from botorch.models import ModelListGP, SingleTaskGP
from botorch.models.model import Model
from botorch.posteriors.fully_bayesian import GaussianMixturePosterior
from botorch.posteriors.gpytorch import GPyTorchPosterior
from botorch.posteriors.posterior_list import PosteriorList
from botorch.sampling.normal import IIDNormalSampler, SobolQMCNormalSampler
from botorch.utils.constraints import get_outcome_constraint_transforms
from botorch.utils.datasets import SupervisedDataset
from botorch.utils.objective import get_objective_weights_transform
from botorch.utils.sampling import sample_hypersphere, sample_simplex
from torch import Tensor
logger: Logger = get_logger(__name__)
NOISELESS_MODELS = {SingleTaskGP}
# Distributions
SIMPLEX = "simplex"
HYPERSPHERE = "hypersphere"
[docs]@dataclass
class SubsetModelData:
model: Model
objective_weights: Tensor
outcome_constraints: Optional[Tuple[Tensor, Tensor]]
objective_thresholds: Optional[Tensor]
indices: Tensor
[docs]def is_noiseless(model: Model) -> bool:
"""Check if a given (single-task) botorch model is noiseless"""
if isinstance(model, ModelListGP):
raise ModelError(
"Checking for noisless models only applies to sub-models of ModelListGP"
)
return model.__class__ in NOISELESS_MODELS
def _filter_X_observed(
Xs: List[Tensor],
objective_weights: Tensor,
bounds: List[Tuple[float, float]],
outcome_constraints: Optional[Tuple[Tensor, Tensor]] = None,
linear_constraints: Optional[Tuple[Tensor, Tensor]] = None,
fixed_features: Optional[Dict[int, float]] = None,
fit_out_of_design: bool = False,
) -> Optional[Tensor]:
r"""Filter input points to those appearing in objective or constraints.
Args:
Xs: The input tensors of a model.
objective_weights: The objective is to maximize a weighted sum of
the columns of f(x). These are the weights.
bounds: A list of (lower, upper) tuples for each column of X.
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)
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. (Not used by single task models)
fixed_features: A map {feature_index: value} for features that
should be fixed to a particular value during generation.
fit_out_of_design: If specified, all training data is returned.
Otherwise, only in design points are returned.
Returns:
Tensor: All points that are feasible and appear in the objective or
the constraints. None if there are no such points.
"""
# Get points observed for all objective and constraint outcomes
X_obs = get_observed(
Xs=Xs,
objective_weights=objective_weights,
outcome_constraints=outcome_constraints,
)
if not fit_out_of_design:
# Filter to those that satisfy constraints.
X_obs = filter_constraints_and_fixed_features(
X=X_obs,
bounds=bounds,
linear_constraints=linear_constraints,
fixed_features=fixed_features,
)
if len(X_obs) > 0:
return torch.as_tensor(X_obs) # please the linter
def _get_X_pending_and_observed(
Xs: List[Tensor],
objective_weights: Tensor,
bounds: List[Tuple[float, float]],
pending_observations: Optional[List[Tensor]] = None,
outcome_constraints: Optional[Tuple[Tensor, Tensor]] = None,
linear_constraints: Optional[Tuple[Tensor, Tensor]] = None,
fixed_features: Optional[Dict[int, float]] = None,
fit_out_of_design: bool = False,
) -> Tuple[Optional[Tensor], Optional[Tensor]]:
r"""Get pending and observed points.
If all points would otherwise be filtered, remove `linear_constraints`
and `fixed_features` from filter and retry.
Args:
Xs: The input tensors of a model.
objective_weights: The objective is to maximize a weighted sum of
the columns of f(x). These are the weights.
bounds: A list of (lower, upper) tuples for each column of X.
pending_observations: A list of m (k_i x d) feature tensors X
for m outcomes and k_i pending observations for outcome i.
(Only used if n > 1).
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)
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. (Not used by single task models)
fixed_features: A map {feature_index: value} for features that
should be fixed to a particular value during generation.
fit_out_of_design: If specified, all training data is returned.
Otherwise, only in design points are returned.
Returns:
Tensor: Pending points that are feasible and appear in the objective or
the constraints. None if there are no such points.
Tensor: Observed points that are feasible and appear in the objective or
the constraints. None if there are no such points.
"""
if pending_observations is None:
X_pending = None
else:
X_pending = _filter_X_observed(
Xs=pending_observations,
objective_weights=objective_weights,
outcome_constraints=outcome_constraints,
bounds=bounds,
linear_constraints=linear_constraints,
fixed_features=fixed_features,
)
filtered_X_observed = _filter_X_observed(
Xs=Xs,
objective_weights=objective_weights,
outcome_constraints=outcome_constraints,
bounds=bounds,
linear_constraints=linear_constraints,
fixed_features=fixed_features,
fit_out_of_design=fit_out_of_design,
)
if filtered_X_observed is not None and len(filtered_X_observed) > 0:
return X_pending, filtered_X_observed
else:
unfiltered_X_observed = _filter_X_observed(
Xs=Xs,
objective_weights=objective_weights,
bounds=bounds,
outcome_constraints=outcome_constraints,
fit_out_of_design=fit_out_of_design,
)
return X_pending, unfiltered_X_observed
def _generate_sobol_points(
n_sobol: int,
bounds: List[Tuple[float, float]],
device: torch.device,
linear_constraints: Optional[Tuple[Tensor, Tensor]] = None,
fixed_features: Optional[Dict[int, float]] = None,
rounding_func: Optional[Callable[[Tensor], Tensor]] = None,
model_gen_options: Optional[TConfig] = None,
) -> Tensor:
linear_constraints_array = None
if linear_constraints is not None:
linear_constraints_array = (
linear_constraints[0].detach().cpu().numpy(),
linear_constraints[1].detach().cpu().numpy(),
)
array_rounding_func = None
if rounding_func is not None:
array_rounding_func = tensor_callable_to_array_callable(
tensor_func=rounding_func, device=device
)
sobol = SobolGenerator(deduplicate=False, seed=np.random.randint(10000))
array_X, _ = sobol.gen(
n=n_sobol,
bounds=bounds,
linear_constraints=linear_constraints_array,
fixed_features=fixed_features,
rounding_func=array_rounding_func,
model_gen_options=model_gen_options,
)
return torch.from_numpy(array_X).to(device)
[docs]def normalize_indices(indices: List[int], d: int) -> List[int]:
r"""Normalize a list of indices to ensure that they are positive.
Args:
indices: A list of indices (may contain negative indices for indexing
"from the back").
d: The dimension of the tensor to index.
Returns:
A normalized list of indices such that each index is between `0` and `d-1`.
"""
normalized_indices = []
for i in indices:
if i < 0:
i = i + d
if i < 0 or i > d - 1:
raise ValueError(f"Index {i} out of bounds for tensor or length {d}.")
normalized_indices.append(i)
return normalized_indices
[docs]def subset_model(
model: Model,
objective_weights: Tensor,
outcome_constraints: Optional[Tuple[Tensor, Tensor]] = None,
objective_thresholds: Optional[Tensor] = None,
) -> SubsetModelData:
"""Subset a botorch model to the outputs used in the optimization.
Args:
model: A BoTorch Model. If the model does not implement the
`subset_outputs` method, this function is a null-op and returns the
input arguments.
objective_weights: The objective is to maximize a weighted sum of
the columns of f(x). These are the weights.
objective_thresholds: The `m`-dim tensor of objective thresholds. There
is one for each modeled metric.
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)
Returns:
A SubsetModelData dataclass containing the model, objective_weights,
outcome_constraints, objective thresholds, all subset to only those
outputs that appear in either the objective weights or the outcome
constraints, along with the indices of the outputs.
"""
nonzero = objective_weights != 0
if outcome_constraints is not None:
A, _ = outcome_constraints
nonzero = nonzero | torch.any(A != 0, dim=0)
idcs_t = torch.arange(nonzero.size(0), device=objective_weights.device)[nonzero]
idcs = idcs_t.tolist()
# note that the number of metrics can be different than
# model.num_outputs which counts multiple tasks per
# outcome as separate outputs
num_outcomes = objective_weights.shape[0]
if len(idcs) == num_outcomes:
# if we use all model outputs, just return the inputs
return SubsetModelData(
model=model,
objective_weights=objective_weights,
outcome_constraints=outcome_constraints,
objective_thresholds=objective_thresholds,
indices=torch.arange(
num_outcomes,
device=objective_weights.device,
),
)
elif len(idcs) > model.num_outputs:
raise RuntimeError(
"Model size inconsistency. Trying to subset a model with "
f"{model.num_outputs} outputs to {len(idcs)} outputs"
)
try:
model = model.subset_output(idcs=idcs)
objective_weights = objective_weights[nonzero]
if outcome_constraints is not None:
A, b = outcome_constraints
outcome_constraints = A[:, nonzero], b
if objective_thresholds is not None:
objective_thresholds = objective_thresholds[nonzero]
except NotImplementedError:
idcs_t = torch.arange(
model.num_outputs,
device=objective_weights.device,
)
return SubsetModelData(
model=model,
objective_weights=objective_weights,
outcome_constraints=outcome_constraints,
objective_thresholds=objective_thresholds,
indices=idcs_t,
)
def _to_inequality_constraints(
linear_constraints: Optional[Tuple[Tensor, Tensor]] = None
) -> Optional[List[Tuple[Tensor, Tensor, float]]]:
if linear_constraints is not None:
A, b = linear_constraints
inequality_constraints = []
k, d = A.shape
for i in range(k):
indices = torch.atleast_1d(A[i, :].nonzero(as_tuple=False).squeeze())
coefficients = torch.atleast_1d(-A[i, indices])
rhs = -b[i, 0].item()
inequality_constraints.append((indices, coefficients, rhs))
else:
inequality_constraints = None
return inequality_constraints
[docs]def tensor_callable_to_array_callable(
tensor_func: Callable[[Tensor], Tensor], device: torch.device
) -> Callable[[np.ndarray], np.ndarray]:
"""transfer a tensor callable to an array callable"""
def array_func(x: np.ndarray) -> np.ndarray:
return tensor_func(torch.from_numpy(x).to(device)).detach().cpu().numpy()
return array_func
def _get_weighted_mo_objective(
objective_weights: Tensor,
) -> WeightedMCMultiOutputObjective:
"""Constructs the `WeightedMCMultiOutputObjective` for the given
objective weights.
"""
nonzero_idcs = torch.nonzero(objective_weights).view(-1)
objective_weights = objective_weights[nonzero_idcs]
return WeightedMCMultiOutputObjective(
weights=objective_weights, outcomes=nonzero_idcs.tolist()
)
def _get_risk_measure(
model: Model,
objective_weights: Tensor,
risk_measure: RiskMeasureMCObjective,
outcome_constraints: Optional[Tuple[Tensor, Tensor]] = None,
X_observed: Optional[Tensor] = None,
) -> RiskMeasureMCObjective:
r"""Processes the risk measure for `get_botorch_objective_and_transform`.
See the docstring of `get_botorch_objective_and_transform` for the arguments.
"""
if outcome_constraints is not None:
# TODO[T131759270]: Handle the constraints via feasibility weighting.
# See `FeasibilityWeightedMCMultiOutputObjective`.
raise NotImplementedError(
"Outcome constraints are not supported with risk measures."
)
# Isinstance doesn't work since it covers subclasses as well.
if risk_measure.preprocessing_function.__class__ not in (
IdentityMCObjective,
IdentityMCMultiOutputObjective,
) or hasattr(risk_measure.preprocessing_function, "outcomes"):
raise UnsupportedError(
"User supplied preprocessing functions for the risk measures are not "
"supported. We construct a new one based on `objective_weights` instead."
)
if isinstance(risk_measure, MultiOutputRiskMeasureMCObjective):
risk_measure.preprocessing_function = _get_weighted_mo_objective(
objective_weights=objective_weights
)
if isinstance(risk_measure, MARS):
risk_measure.chebyshev_weights = sample_simplex(
len(objective_weights.nonzero())
).squeeze()
if X_observed is None:
raise UnsupportedError("X_observed is required when using MARS.")
risk_measure.set_baseline_Y(model=model, X_baseline=X_observed)
else:
risk_measure.preprocessing_function = LinearMCObjective(
weights=objective_weights
)
return risk_measure
[docs]def get_out_of_sample_best_point_acqf(
model: Model,
Xs: List[Tensor],
X_observed: Tensor,
objective_weights: Tensor,
mc_samples: int = 512,
fixed_features: Optional[Dict[int, float]] = None,
fidelity_features: Optional[List[int]] = None,
target_fidelities: Optional[Dict[int, float]] = None,
outcome_constraints: Optional[Tuple[Tensor, Tensor]] = None,
seed_inner: Optional[int] = None,
qmc: bool = True,
risk_measure: Optional[RiskMeasureMCObjective] = None,
**kwargs: Any,
) -> Tuple[AcquisitionFunction, Optional[List[int]]]:
"""Picks an appropriate acquisition function to find the best
out-of-sample (predicted by the given surrogate model) point
and instantiates it.
NOTE: Typically the appropriate function is the posterior mean,
but can differ to account for fidelities etc.
"""
model = model
# subset model only to the outcomes we need for the optimization
if kwargs.get(Keys.SUBSET_MODEL, True):
subset_model_results = subset_model(
model=model,
objective_weights=objective_weights,
outcome_constraints=outcome_constraints,
)
model = subset_model_results.model
objective_weights = subset_model_results.objective_weights
outcome_constraints = subset_model_results.outcome_constraints
fixed_features = fixed_features or {}
target_fidelities = target_fidelities or {}
if fidelity_features:
# we need to optimize at the target fidelities
if any(f in fidelity_features for f in fixed_features):
raise RuntimeError("Fixed features cannot also be fidelity features.")
elif set(fidelity_features) != set(target_fidelities):
raise RuntimeError(
"Must provide a target fidelity for every fidelity feature."
)
# make sure to not modify fixed_features in-place
fixed_features = {**fixed_features, **target_fidelities}
elif target_fidelities:
raise RuntimeError(
"Must specify fidelity_features in fit() when using target fidelities."
)
acqf_class, acqf_options = pick_best_out_of_sample_point_acqf_class(
outcome_constraints=outcome_constraints,
mc_samples=mc_samples,
qmc=qmc,
seed_inner=seed_inner,
risk_measure=risk_measure,
)
objective, posterior_transform = get_botorch_objective_and_transform(
botorch_acqf_class=acqf_class,
model=model,
objective_weights=objective_weights,
outcome_constraints=outcome_constraints,
X_observed=X_observed,
risk_measure=risk_measure,
)
if objective is not None:
if not isinstance(objective, MCAcquisitionObjective):
raise UnsupportedError(
f"Unknown objective type: {objective.__class__}" # pragma: nocover
)
acqf_options = {"objective": objective, **acqf_options}
if posterior_transform is not None:
acqf_options = {"posterior_transform": posterior_transform, **acqf_options}
acqf = acqf_class(model=model, **acqf_options) # pyre-ignore [45]
if fixed_features:
acqf = FixedFeatureAcquisitionFunction(
acq_function=acqf,
d=X_observed.size(-1),
columns=list(fixed_features.keys()),
values=list(fixed_features.values()),
)
non_fixed_idcs = [i for i in range(Xs[0].size(-1)) if i not in fixed_features]
else:
non_fixed_idcs = None
return acqf, non_fixed_idcs
[docs]def pick_best_out_of_sample_point_acqf_class(
outcome_constraints: Optional[Tuple[Tensor, Tensor]] = None,
mc_samples: int = 512,
qmc: bool = True,
seed_inner: Optional[int] = None,
risk_measure: Optional[RiskMeasureMCObjective] = None,
) -> Tuple[Type[AcquisitionFunction], Dict[str, Any]]:
if outcome_constraints is None and risk_measure is None:
acqf_class = PosteriorMean
acqf_options = {}
else:
acqf_class = qSimpleRegret
sampler_class = SobolQMCNormalSampler if qmc else IIDNormalSampler
acqf_options = {
Keys.SAMPLER.value: sampler_class(
sample_shape=torch.Size([mc_samples]), seed=seed_inner
)
}
return cast(Type[AcquisitionFunction], acqf_class), acqf_options
[docs]def predict_from_model(
model: Model, X: Tensor, use_posterior_predictive: bool = False
) -> Tuple[Tensor, Tensor]:
r"""Predicts outcomes given a model and input tensor.
For a `GaussianMixturePosterior` we currently use a Gaussian approximation where we
compute the mean and variance of the Gaussian mixture. This should ideally be
changed to compute quantiles instead when Ax supports non-Gaussian distributions.
Args:
model: A botorch Model.
X: A `n x d` tensor of input parameters.
use_posterior_predictive: A boolean indicating if the predictions
should be from the posterior predictive (i.e. including
observation noise).
Returns:
Tensor: The predicted posterior mean as an `n x o`-dim tensor.
Tensor: The predicted posterior covariance as a `n x o x o`-dim tensor.
"""
with torch.no_grad():
# TODO: Allow Posterior to (optionally) return the full covariance matrix
posterior = model.posterior(X, observation_noise=use_posterior_predictive)
if isinstance(posterior, GaussianMixturePosterior):
mean = posterior.mixture_mean.cpu().detach()
var = posterior.mixture_variance.cpu().detach().clamp_min(0)
elif isinstance(posterior, (GPyTorchPosterior, PosteriorList)):
mean = posterior.mean.cpu().detach()
var = posterior.variance.cpu().detach().clamp_min(0)
else:
raise UnsupportedError(
"Non-Gaussian posteriors are currently not supported."
)
cov = torch.diag_embed(var)
return mean, cov
# TODO(jej): Possibly refactor to use "objective_directions".
[docs]def randomize_objective_weights(
objective_weights: Tensor,
random_scalarization_distribution: str = SIMPLEX,
) -> Tensor:
"""Generate a random weighting based on acquisition function settings.
Args:
objective_weights: Base weights to multiply by random values.
random_scalarization_distribution: "simplex" or "hypersphere".
Returns:
A normalized list of indices such that each index is between `0` and `d-1`.
"""
# Set distribution and sample weights.
distribution = random_scalarization_distribution
dtype = objective_weights.dtype
device = objective_weights.device
if distribution == SIMPLEX:
random_weights = sample_simplex(
len(objective_weights), dtype=dtype, device=device
).squeeze()
elif distribution == HYPERSPHERE:
random_weights = torch.abs(
sample_hypersphere(
len(objective_weights), dtype=dtype, device=device
).squeeze()
)
# pyre-fixme[61]: `random_weights` may not be initialized here.
objective_weights = torch.mul(objective_weights, random_weights)
return objective_weights
def _datasets_to_legacy_inputs(
datasets: List[SupervisedDataset],
) -> Tuple[List[Tensor], List[Tensor], List[Tensor]]:
"""Convert a dictionary of dataset containers to legacy X, Y, Yvar inputs"""
Xs, Ys, Yvars = [], [], []
for dataset in datasets:
if not isinstance(dataset, SupervisedDataset):
raise UnsupportedError("Legacy setup only supports `SupervisedDataset`s")
for i, _ in enumerate(dataset.outcome_names):
Xs.append(dataset.X)
Ys.append(dataset.Y[:, i].unsqueeze(-1))
if dataset.Yvar is not None:
Yvars.append(dataset.Yvar[:, i].unsqueeze(-1))
else:
Yvars.append(torch.full_like(Ys[-1], float("nan")))
return Xs, Ys, Yvars
