#!/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. from typing import Any, Callable, Dict, List, Optional, Tuple import numpy as np import torch from ax.exceptions.core import SearchSpaceExhausted from ax.models.base import Model from ax.models.model_utils import ( add_fixed_features, rejection_sample, tunable_feature_indices, validate_bounds, ) from ax.models.types import TConfig from ax.utils.common.docutils import copy_doc from ax.utils.common.logger import get_logger from botorch.utils.sampling import HitAndRunPolytopeSampler from torch import Tensor logger = get_logger(__name__) class RandomModel(Model): """This class specifies the basic skeleton for a random model. As random generators do not make use of models, they do not implement the fit or predict methods. These models do not need data, or optimization configs. To satisfy search space parameter constraints, these models can use rejection sampling. To enable rejection sampling for a subclass, only only `_gen_samples` needs to be implemented, or alternatively, `_gen_unconstrained`/`gen` can be directly implemented. Attributes: deduplicate: If True (defaults to True), a single instantiation of the model will not return the same point twice. This flag is used in rejection sampling. scramble: If True, permutes the parameter values among the elements of the Sobol sequence. Default is True. seed: An optional seed value for scrambling. """ def __init__( self, deduplicate: bool = True, seed: Optional[int] = None, generated_points: Optional[np.ndarray] = None, fallback_to_sample_polytope: bool = False, ) -> None: super().__init__() self.deduplicate = deduplicate self.seed = seed # Used for deduplication. self.generated_points = generated_points self.fallback_to_sample_polytope = fallback_to_sample_polytope def gen( self, n: int, bounds: List[Tuple[float, float]], linear_constraints: Optional[Tuple[np.ndarray, np.ndarray]] = None, fixed_features: Optional[Dict[int, float]] = None, model_gen_options: Optional[TConfig] = None, rounding_func: Optional[Callable[[np.ndarray], np.ndarray]] = None, ) -> Tuple[np.ndarray, np.ndarray]: """Generate new candidates. Args: n: Number of candidates to generate. bounds: A list of (lower, upper) tuples for each column of X. Defined on [0, 1]^d. 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. fixed_features: A map {feature_index: value} for features that should be fixed to a particular value during generation. model_gen_options: A config dictionary that is passed along to the model. rounding_func: A function that rounds an optimization result appropriately (e.g., according to `round-trip` transformations). Returns: 2-element tuple containing - (n x d) array of generated points. - Uniform weights, an n-array of ones for each point. """ tf_indices = tunable_feature_indices( bounds=bounds, fixed_features=fixed_features ) if fixed_features: fixed_feature_indices = np.array(list(fixed_features.keys())) else: fixed_feature_indices = np.array([]) validate_bounds(bounds=bounds, fixed_feature_indices=fixed_feature_indices) attempted_draws = 0 max_draws = None if model_gen_options: max_draws = model_gen_options.get("max_rs_draws") if max_draws is not None: # pyre-fixme[6]: Expected `Union[bytes, str, typing.SupportsInt]` # for 1st param but got # `Union[botorch.acquisition.acquisition.AcquisitionFunction, float, # int, str]`. # pyre-fixme[35]: Target cannot be annotated. max_draws: int = int(max_draws) try: # Always rejection sample, but this only rejects if there are # constraints or actual duplicates and deduplicate is specified. # If rejection sampling fails, fall back to polytope sampling points, attempted_draws = rejection_sample( gen_unconstrained=self._gen_unconstrained, n=n, d=len(bounds), tunable_feature_indices=tf_indices, linear_constraints=linear_constraints, deduplicate=self.deduplicate, max_draws=max_draws, fixed_features=fixed_features, rounding_func=rounding_func, existing_points=self.generated_points, ) except SearchSpaceExhausted as e: if self.fallback_to_sample_polytope: logger.info( "Rejection sampling exceeded specified maximum draws." "Falling back on polytope sampler" ) # If rejection sampling fails, try polytope sampler. polytope_sampler = HitAndRunPolytopeSampler( inequality_constraints=self._convert_inequality_constraints( linear_constraints ), equality_constraints=self._convert_equality_constraints( len(bounds), fixed_features ), interior_point=self._get_last_point(), bounds=self._convert_bounds(bounds), ) points = polytope_sampler.draw(n).numpy() else: raise e # pyre-fixme[16]: `RandomModel` has no attribute `attempted_draws`. self.attempted_draws = attempted_draws if self.deduplicate: if self.generated_points is None: self.generated_points = points else: self.generated_points = np.vstack([self.generated_points, points]) return points, np.ones(len(points)) @copy_doc(Model._get_state) def _get_state(self) -> Dict[str, Any]: state = super()._get_state() if not self.deduplicate: return state state.update({"generated_points": self.generated_points}) return state def _gen_unconstrained( self, n: int, d: int, tunable_feature_indices: np.ndarray, fixed_features: Optional[Dict[int, float]] = None, ) -> np.ndarray: """Generate n points, from an unconstrained parameter space, using _gen_samples. Args: n: Number of points to generate. d: Dimension of parameter space. fixed_features: A map {feature_index: value} for features that should be fixed to a particular value during generation. tunable_feature_indices: Parameter indices (in d) which are tunable. Returns: An (n x d) array of generated points. """ tunable_points = self._gen_samples(n=n, tunable_d=len(tunable_feature_indices)) points = add_fixed_features( tunable_points=tunable_points, d=d, tunable_feature_indices=tunable_feature_indices, fixed_features=fixed_features, ) return points def _gen_samples(self, n: int, tunable_d: int) -> np.ndarray: """Generate n samples on [0, 1]^d. Args: n: Number of points to generate. Returns: (n x d) array of generated points. """ raise NotImplementedError("Base RandomModel can't generate samples.") def _convert_inequality_constraints( self, linear_constraints: Optional[Tuple[np.ndarray, np.ndarray]] ) -> Optional[Tuple[Tensor, Tensor]]: """Helper method to convert inequality constraints used by the rejection sampler to the format required for the polytope sampler. Args: 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. Returns: Optional 2-element tuple containing A and b as tensors """ if linear_constraints is None: return None else: A = torch.tensor(linear_constraints[0], dtype=torch.double) b = torch.tensor(linear_constraints[1], dtype=torch.double) return A, b def _convert_equality_constraints( self, d: int, fixed_features: Optional[Dict[int, float]] ) -> Optional[Tuple[Tensor, Tensor]]: """Helper method to convert the fixed feature dictionary used by the rejection sampler to the corresponding matrix representation required for the polytope sampler. Args: d: dimension of samples fixed_features: A map {feature_index: value} for features that should be fixed to a particular value during generation. Returns: Optional 2-element tuple containing C and c such that the equality constraints are defined by Cx = c """ if fixed_features is None: return None else: n = len(fixed_features) fixed_indices = sorted(fixed_features.keys()) fixed_vals = torch.tensor( [fixed_features[i] for i in fixed_indices], dtype=torch.double ) constraint_matrix = torch.zeros((n, d), dtype=torch.double) for index in range(0, len(fixed_vals)): constraint_matrix[index, fixed_indices[index]] = 1.0 return constraint_matrix, fixed_vals def _convert_bounds(self, bounds: List[Tuple[float, float]]) -> Optional[Tensor]: """Helper method to convert bounds list used by the rejectionsampler to the tensor format required for the polytope sampler. Args: bounds: A list of (lower, upper) tuples for each column of X. Defined on [0, 1]^d. Returns: Optional 2 x d tensor representing the bounds """ if bounds is None: return None else: return torch.tensor(bounds, dtype=torch.double).transpose(-1, -2) def _get_last_point(self) -> Optional[Tensor]: # Return the last sampled point when points have been sampled if self.generated_points is None: return None else: last_point = self.generated_points[-1, :].reshape((-1, 1)) return torch.from_numpy(last_point).double()