Source code for ax.models.torch.botorch_modular.utils

#!/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.

import warnings
from contextlib import contextmanager
from logging import Logger
from typing import Any, Dict, Generator, List, Optional, Tuple, Type

import torch
from ax.core.search_space import SearchSpaceDigest
from ax.exceptions.core import AxWarning, UnsupportedError
from ax.models.types import TConfig
from ax.utils.common.constants import Keys
from ax.utils.common.logger import get_logger
from ax.utils.common.typeutils import checked_cast
from botorch.acquisition.acquisition import AcquisitionFunction
from botorch.acquisition.monte_carlo import qNoisyExpectedImprovement
from botorch.acquisition.multi_objective.monte_carlo import (
    qNoisyExpectedHypervolumeImprovement,
)
from botorch.fit import fit_fully_bayesian_model_nuts, fit_gpytorch_mll
from botorch.models.gp_regression import FixedNoiseGP, SingleTaskGP
from botorch.models.gp_regression_fidelity import (
    FixedNoiseMultiFidelityGP,
    SingleTaskMultiFidelityGP,
)
from botorch.models.gp_regression_mixed import MixedSingleTaskGP
from botorch.models.gpytorch import BatchedMultiOutputGPyTorchModel, GPyTorchModel
from botorch.models.model import Model, ModelList
from botorch.models.multitask import FixedNoiseMultiTaskGP, MultiTaskGP
from botorch.models.pairwise_gp import PairwiseGP
from botorch.models.transforms.input import ChainedInputTransform
from botorch.utils.datasets import FixedNoiseDataset, SupervisedDataset
from botorch.utils.transforms import is_fully_bayesian
from gpytorch.mlls.marginal_log_likelihood import MarginalLogLikelihood
from torch import Tensor

MIN_OBSERVED_NOISE_LEVEL = 1e-7
logger: Logger = get_logger(__name__)


[docs]def use_model_list( datasets: List[SupervisedDataset], botorch_model_class: Type[Model] ) -> bool: if issubclass(botorch_model_class, MultiTaskGP): # We currently always wrap multi-task models into `ModelListGP`. return True if len(datasets) == 1: # Just one outcome, can use single model. return False if issubclass(botorch_model_class, BatchedMultiOutputGPyTorchModel) and all( torch.equal(datasets[0].X(), ds.X()) for ds in datasets[1:] ): # Single model, batched multi-output case. return False # If there are multiple Xs and they are not all equal, we # use `ListSurrogate` and `ModelListGP`. return True
[docs]def choose_model_class( datasets: List[SupervisedDataset], search_space_digest: SearchSpaceDigest, ) -> Type[Model]: """Chooses a BoTorch `Model` using the given data (currently just Yvars) and its properties (information about task and fidelity features). Args: Yvars: List of tensors, each representing observation noise for a given outcome, where outcomes are in the same order as in Xs. task_features: List of columns of X that are tasks. fidelity_features: List of columns of X that are fidelity parameters. Returns: A BoTorch `Model` class. """ if len(search_space_digest.fidelity_features) > 1: raise NotImplementedError( "Only a single fidelity feature supported " f"(got: {search_space_digest.fidelity_features})." ) if len(search_space_digest.task_features) > 1: raise NotImplementedError( f"Only a single task feature supported " f"(got: {search_space_digest.task_features})." ) if search_space_digest.task_features and search_space_digest.fidelity_features: raise NotImplementedError( "Multi-task multi-fidelity optimization not yet supported." ) is_fixed_noise = [isinstance(ds, FixedNoiseDataset) for ds in datasets] all_inferred = not any(is_fixed_noise) if not all_inferred and not all(is_fixed_noise): raise ValueError( "Mix of known and unknown variances indicates valuation function " "errors. Variances should all be specified, or none should be." ) # Multi-task cases (when `task_features` specified). if search_space_digest.task_features and all_inferred: model_class = MultiTaskGP # Unknown observation noise. elif search_space_digest.task_features: model_class = FixedNoiseMultiTaskGP # Known observation noise. # Single-task multi-fidelity cases. elif search_space_digest.fidelity_features and all_inferred: model_class = SingleTaskMultiFidelityGP # Unknown observation noise. elif search_space_digest.fidelity_features: model_class = FixedNoiseMultiFidelityGP # Known observation noise. # Mixed optimization case. Note that presence of categorical # features in search space digest indicates that downstream in the # stack we chose not to perform continuous relaxation on those # features. elif search_space_digest.categorical_features: if not all_inferred: logger.warning( "Using `MixedSingleTaskGP` despire the known `Yvar` values. This " "is a temporary measure while fixed-noise mixed BO is in the works." ) model_class = MixedSingleTaskGP # Single-task single-fidelity cases. elif all_inferred: # Unknown observation noise. model_class = SingleTaskGP else: model_class = FixedNoiseGP # Known observation noise. logger.debug(f"Chose BoTorch model class: {model_class}.") return model_class
[docs]def choose_botorch_acqf_class( 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, objective_thresholds: Optional[Tensor] = None, objective_weights: Optional[Tensor] = None, ) -> Type[AcquisitionFunction]: """Chooses a BoTorch `AcquisitionFunction` class.""" if objective_thresholds is not None or ( # using objective_weights is a less-than-ideal fix given its ambiguity, # the real fix would be to revisit the infomration passed down via # the modelbridge (and be explicit about whether we scalarize or perform MOO) objective_weights is not None and objective_weights.nonzero().numel() > 1 ): acqf_class = qNoisyExpectedHypervolumeImprovement else: acqf_class = qNoisyExpectedImprovement logger.debug(f"Chose BoTorch acquisition function class: {acqf_class}.") return acqf_class
[docs]def construct_acquisition_and_optimizer_options( acqf_options: TConfig, model_gen_options: Optional[TConfig] = None ) -> Tuple[TConfig, TConfig]: """Extract acquisition and optimizer options from `model_gen_options`.""" acq_options = acqf_options.copy() opt_options = {} if model_gen_options: acq_options.update( checked_cast(dict, model_gen_options.get(Keys.ACQF_KWARGS, {})) ) # TODO: Add this if all acq. functions accept the `subset_model` # kwarg or opt for kwarg filtering. # acq_options[SUBSET_MODEL] = model_gen_options.get(SUBSET_MODEL) opt_options = checked_cast( dict, model_gen_options.get(Keys.OPTIMIZER_KWARGS, {}) ).copy() return acq_options, opt_options
[docs]def convert_to_block_design( datasets: List[SupervisedDataset], metric_names: List[str], force: bool = False, ) -> Tuple[List[SupervisedDataset], List[str]]: # Convert data to "block design". TODO: Figure out a better # solution for this using the data containers (pass outcome # names as properties of the data containers) is_fixed = [isinstance(ds, FixedNoiseDataset) for ds in datasets] if any(is_fixed) and not all(is_fixed): raise UnsupportedError( "Cannot convert mixed data with and without variance " "observaitons to `block design`." ) is_fixed = all(is_fixed) Xs = [dataset.X() for dataset in datasets] metric_names = ["_".join(metric_names)] # TODO: Improve this. if len({X.shape for X in Xs}) != 1 or not all( torch.equal(X, Xs[0]) for X in Xs[1:] ): if not force: raise UnsupportedError( "Cannot convert data to non-block design data. " "To force this and drop data not shared between " "outcomes use `force=True`." ) warnings.warn( "Forcing converion of data not complying to a block design " "to block design by dropping observations that are not shared " "between outcomes.", AxWarning, ) X_shared, idcs_shared = _get_shared_rows(Xs=Xs) Y = torch.cat([ds.Y()[i] for ds, i in zip(datasets, idcs_shared)], dim=-1) if is_fixed: Yvar = torch.cat( [ds.Yvar()[i] for ds, i in zip(datasets, idcs_shared)], dim=-1, ) datasets = [FixedNoiseDataset(X=X_shared, Y=Y, Yvar=Yvar)] else: datasets = [SupervisedDataset(X=X_shared, Y=Y)] return datasets, metric_names # data complies to block design, can concat with impunity Y = torch.cat([ds.Y() for ds in datasets], dim=-1) if is_fixed: Yvar = torch.cat([ds.Yvar() for ds in datasets], dim=-1) # pyre-ignore [16] datasets = [FixedNoiseDataset(X=Xs[0], Y=Y, Yvar=Yvar)] else: datasets = [SupervisedDataset(X=Xs[0], Y=Y)] return datasets, metric_names
def _get_shared_rows(Xs: List[Tensor]) -> Tuple[Tensor, List[Tensor]]: """Extract shared rows from a list of tensors Args: Xs: A list of m two-dimensional tensors with shapes `(n_1 x d), ..., (n_m x d)`. It is not required that the `n_i` are the same. Returns: A two-tuple containing (i) a Tensor with the rows that are shared between all the Tensors in `Xs`, and (ii) a list of index tensors that indicate the location of these rows in the respective elements of `Xs`. """ idcs_shared = [] Xs_sorted = sorted(Xs, key=len) X_shared = Xs_sorted[0].clone() for X in Xs_sorted[1:]: X_shared = X_shared[(X_shared == X.unsqueeze(-2)).all(dim=-1).any(dim=-2)] # get indices for X in Xs: same = (X_shared == X.unsqueeze(-2)).all(dim=-1).any(dim=-1) idcs_shared.append(torch.arange(same.shape[-1], device=X_shared.device)[same]) return X_shared, idcs_shared
[docs]def fit_botorch_model( model: Model, mll_class: Type[MarginalLogLikelihood], mll_options: Optional[Dict[str, Any]] = None, ) -> None: """Fit a BoTorch model.""" models = model.models if isinstance(model, ModelList) else [model] for m in models: # TODO: Support deterministic models when we support `ModelList` if is_fully_bayesian(m): fit_fully_bayesian_model_nuts(m, disable_progbar=True) elif isinstance(m, (GPyTorchModel, PairwiseGP)): mll_options = mll_options or {} mll = mll_class(likelihood=m.likelihood, model=m, **mll_options) fit_gpytorch_mll(mll) else: raise NotImplementedError( f"Model of type {m.__class__.__name__} is currently not supported." )
[docs]@contextmanager def disable_one_to_many_transforms(model: Model) -> Generator[None, None, None]: r"""A context manager for temporarily disabling one-to-many transforms. This can be used to avoid perturbing the user supplied inputs when getting the predictions from the model. NOTE: This currently does not support chained input transforms. Args: model: The BoTorch `Model` to disable the transforms for. """ models = model.models if isinstance(model, ModelList) else [model] input_transforms = [getattr(m, "input_transform", None) for m in models] try: for intf in input_transforms: if intf is None: continue if isinstance(intf, ChainedInputTransform): raise UnsupportedError( "ChainedInputTransforms are currently not supported." ) if intf.is_one_to_many: intf.transform_on_eval = False yield finally: for intf in input_transforms: if intf is not None and intf.is_one_to_many: intf.transform_on_eval = True