Source code for ax.models.torch.fully_bayesian

#!/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


"""
Models and utilities for fully bayesian inference.

TODO: move some of this into botorch.

References

.. [Eriksson2021saasbo]
    D. Eriksson, M. Jankowiak. High-Dimensional Bayesian Optimization
    with Sparse Axis-Aligned Subspaces. Proceedings of the Thirty-
    Seventh Conference on Uncertainty in Artificial Intelligence, 2021.

.. [Eriksson2021nas]
    D. Eriksson, P. Chuang, S. Daulton, et al. Latency-Aware Neural
    Architecture Search with Multi-Objective Bayesian Optimization.
    ICML AutoML Workshop, 2021.

"""

import math
import sys
import time
import types
import warnings

from logging import Logger
from typing import Any, Callable, Dict, List, Optional, Tuple

import numpy as np

import pyro
import torch
from ax.exceptions.core import AxError
from ax.models.torch.botorch import (
    BotorchModel,
    TAcqfConstructor,
    TBestPointRecommender,
    TModelConstructor,
    TModelPredictor,
    TOptimizer,
)
from ax.models.torch.botorch_defaults import (
    get_qLogNEI,
    MIN_OBSERVED_NOISE_LEVEL,
    recommend_best_observed_point,
    scipy_optimizer,
)
from ax.models.torch.botorch_moo import MultiObjectiveBotorchModel
from ax.models.torch.botorch_moo_defaults import (
    get_qLogNEHVI,
    pareto_frontier_evaluator,
)
from ax.models.torch.frontier_utils import TFrontierEvaluator
from ax.models.torch.fully_bayesian_model_utils import (
    _get_single_task_gpytorch_model,
    load_mcmc_samples_to_model,
    pyro_sample_input_warping,
    pyro_sample_mean,
    pyro_sample_noise,
    pyro_sample_outputscale,
    pyro_sample_saas_lengthscales,
)
from ax.utils.common.docutils import copy_doc
from ax.utils.common.logger import get_logger
from ax.utils.common.typeutils import checked_cast
from botorch.acquisition import AcquisitionFunction
from botorch.models.gpytorch import GPyTorchModel
from botorch.models.model import Model
from botorch.models.model_list_gp_regression import ModelListGP
from botorch.posteriors.gpytorch import GPyTorchPosterior
from botorch.utils.safe_math import logmeanexp
from gpytorch.kernels.kernel import dist
from torch import Tensor

logger: Logger = get_logger(__name__)


SAAS_DEPRECATION_MSG = (
    "Passing `use_saas` is no longer supported and has no effect. "
    "SAAS priors are used by default. "
    "This will become an error in the future."
)


[docs]def predict_from_model_mcmc(model: Model, X: Tensor) -> Tuple[Tensor, Tensor]: r"""Predicts outcomes given a model and input tensor. This method integrates over the hyperparameter posterior. Args: model: A batched botorch Model where the batch dimension corresponds to sampled hyperparameters. X: A `n x d` tensor of input parameters. 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(): # compute the batch (independent posterior over the inputs) posterior = checked_cast(GPyTorchPosterior, model.posterior(X.unsqueeze(-3))) # the mean and variance both have shape: n x num_samples x m (after squeezing) mean = posterior.mean.cpu().detach() # TODO: Allow Posterior to (optionally) return the full covariance matrix variance = posterior.variance.cpu().detach().clamp_min(0) # marginalize over samples t1 = variance.sum(dim=0) / variance.shape[0] t2 = mean.pow(2).sum(dim=0) / variance.shape[0] t3 = -(mean.sum(dim=0) / variance.shape[0]).pow(2) variance = t1 + t2 + t3 mean = mean.mean(dim=0) cov = torch.diag_embed(variance) return mean, cov
[docs]def compute_dists(X: Tensor, Z: Tensor, lengthscale: Tensor) -> Tensor: """Compute kernel distances.""" mean = X.mean(dim=0) x1 = (X - mean).div(lengthscale) x2 = (Z - mean).div(lengthscale) return dist(x1=x1, x2=x2, x1_eq_x2=torch.equal(x1, x2))
[docs]def matern_kernel(X: Tensor, Z: Tensor, lengthscale: Tensor, nu: float = 2.5) -> Tensor: """Scaled Matern kernel.""" dist = compute_dists(X=X, Z=Z, lengthscale=lengthscale) exp_component = torch.exp(-math.sqrt(nu * 2) * dist) if nu == 0.5: constant_component = 1 elif nu == 1.5: constant_component = (math.sqrt(3) * dist).add(1) elif nu == 2.5: # pyre-fixme[58]: `**` is not supported for operand types `Tensor` and `int`. constant_component = (math.sqrt(5) * dist).add(1).add(5.0 / 3.0 * (dist**2)) else: raise AxError(f"Unsupported value of nu: {nu}") return constant_component * exp_component
[docs]def rbf_kernel(X: Tensor, Z: Tensor, lengthscale: Tensor) -> Tensor: """Scaled RBF kernel.""" dist = compute_dists(X=X, Z=Z, lengthscale=lengthscale) # pyre-fixme[6]: For 1st param expected `Tensor` but got `float`. # pyre-fixme[58]: `**` is not supported for operand types `Tensor` and `int`. return torch.exp(-0.5 * (dist**2))
[docs]def single_task_pyro_model( X: Tensor, Y: Tensor, Yvar: Tensor, use_input_warping: bool = False, eps: float = 1e-7, gp_kernel: str = "matern", task_feature: Optional[int] = None, rank: Optional[int] = None, ) -> None: r"""Instantiates a single task pyro model for running fully bayesian inference. Args: X: A `n x d` tensor of input parameters. Y: A `n x 1` tensor of output. Yvar: A `n x 1` tensor of observed noise. use_input_warping: A boolean indicating whether to use input warping task_feature: Column index of task feature in X. gp_kernel: kernel name. Currently only two kernels are supported: "matern" for Matern Kernel and "rbf" for RBFKernel. rank: num of latent task features to learn for task covariance. """ Y = Y.view(-1) Yvar = Yvar.view(-1) tkwargs = {"dtype": X.dtype, "device": X.device} dim = X.shape[-1] # TODO: test alternative outputscale priors outputscale = pyro_sample_outputscale(concentration=2.0, rate=0.15, **tkwargs) mean = pyro_sample_mean(**tkwargs) if torch.isnan(Yvar).all(): # infer noise level noise = MIN_OBSERVED_NOISE_LEVEL + pyro_sample_noise(**tkwargs) else: noise = Yvar.clamp_min(MIN_OBSERVED_NOISE_LEVEL) # pyre-fixme[6]: For 2nd param expected `float` but got `Union[device, dtype]`. lengthscale = pyro_sample_saas_lengthscales(dim=dim, **tkwargs) # transform inputs through kumaraswamy cdf if use_input_warping: c0, c1 = pyro_sample_input_warping(dim=dim, **tkwargs) # unnormalize X from [0, 1] to [eps, 1-eps] X = (X * (1 - 2 * eps) + eps).clamp(eps, 1 - eps) X_tf = 1 - torch.pow((1 - torch.pow(X, c1)), c0) else: X_tf = X # compute kernel if gp_kernel == "matern": K = matern_kernel(X=X_tf, Z=X_tf, lengthscale=lengthscale) elif gp_kernel == "rbf": K = rbf_kernel(X=X_tf, Z=X_tf, lengthscale=lengthscale) else: raise ValueError(f"Expected kernel to be 'rbf' or 'matern', got {gp_kernel}") # add noise K = outputscale * K + noise * torch.eye(X.shape[0], dtype=X.dtype, device=X.device) pyro.sample( "Y", # pyre-fixme[16]: Module `distributions` has no attribute `MultivariateNormal`. pyro.distributions.MultivariateNormal( loc=mean.view(-1).expand(X.shape[0]), covariance_matrix=K, ), obs=Y, )
def _get_model_mcmc_samples( Xs: List[Tensor], Ys: List[Tensor], Yvars: List[Tensor], task_features: List[int], fidelity_features: List[int], metric_names: List[str], state_dict: Optional[Dict[str, Tensor]] = None, refit_model: bool = True, use_input_warping: bool = False, use_loocv_pseudo_likelihood: bool = False, num_samples: int = 256, warmup_steps: int = 512, thinning: int = 16, max_tree_depth: int = 6, disable_progbar: bool = False, gp_kernel: str = "matern", verbose: bool = False, jit_compile: bool = False, # pyre-fixme[24]: Generic type `Callable` expects 2 type parameters. pyro_model: Callable = single_task_pyro_model, # pyre-fixme[24]: Generic type `Callable` expects 2 type parameters. get_gpytorch_model: Callable = _get_single_task_gpytorch_model, rank: Optional[int] = 1, **kwargs: Any, # pyre-fixme[24]: Generic type `dict` expects 2 type parameters, use `typing.Dict` # to avoid runtime subscripting errors. ) -> Tuple[ModelListGP, List[Dict]]: r"""Instantiates a batched GPyTorchModel(ModelListGP) based on the given data and fit the model based on MCMC in pyro. Args: pyro_model: callable to instantiate a pyro model for running MCMC get_gpytorch_model: callable to instantiate a coupled GPyTorchModel to load the returned MCMC samples. """ model = get_gpytorch_model( Xs=Xs, Ys=Ys, Yvars=Yvars, task_features=task_features, fidelity_features=fidelity_features, state_dict=state_dict, num_samples=num_samples, thinning=thinning, use_input_warping=use_input_warping, gp_kernel=gp_kernel, **kwargs, ) if state_dict is not None: # Expected `OrderedDict[typing.Any, typing.Any]` for 1st # param but got `Dict[str, Tensor]`. model.load_state_dict(state_dict) mcmc_samples_list = [] if len(task_features) > 0: task_feature = task_features[0] else: task_feature = None if state_dict is None or refit_model: for X, Y, Yvar in zip(Xs, Ys, Yvars): mcmc_samples = run_inference( pyro_model=pyro_model, X=X, Y=Y, Yvar=Yvar, num_samples=num_samples, warmup_steps=warmup_steps, thinning=thinning, use_input_warping=use_input_warping, max_tree_depth=max_tree_depth, disable_progbar=disable_progbar, gp_kernel=gp_kernel, verbose=verbose, task_feature=task_feature, rank=rank, jit_compile=jit_compile, ) mcmc_samples_list.append(mcmc_samples) return model, mcmc_samples_list
[docs]def get_and_fit_model_mcmc( Xs: List[Tensor], Ys: List[Tensor], Yvars: List[Tensor], task_features: List[int], fidelity_features: List[int], metric_names: List[str], state_dict: Optional[Dict[str, Tensor]] = None, refit_model: bool = True, use_input_warping: bool = False, use_loocv_pseudo_likelihood: bool = False, num_samples: int = 256, warmup_steps: int = 512, thinning: int = 16, max_tree_depth: int = 6, disable_progbar: bool = False, gp_kernel: str = "matern", verbose: bool = False, jit_compile: bool = False, **kwargs: Any, ) -> GPyTorchModel: r"""Instantiates a batched GPyTorchModel(ModelListGP) based on the given data and fit the model based on MCMC in pyro. The batch dimension corresponds to sampled hyperparameters from MCMC. """ model, mcmc_samples_list = _get_model_mcmc_samples( Xs=Xs, Ys=Ys, Yvars=Yvars, task_features=task_features, fidelity_features=fidelity_features, metric_names=metric_names, state_dict=state_dict, refit_model=refit_model, use_input_warping=use_input_warping, use_loocv_pseudo_likelihood=use_loocv_pseudo_likelihood, num_samples=num_samples, warmup_steps=warmup_steps, thinning=thinning, max_tree_depth=max_tree_depth, disable_progbar=disable_progbar, gp_kernel=gp_kernel, verbose=verbose, jit_compile=jit_compile, pyro_model=single_task_pyro_model, get_gpytorch_model=_get_single_task_gpytorch_model, ) for i, mcmc_samples in enumerate(mcmc_samples_list): load_mcmc_samples_to_model(model=model.models[i], mcmc_samples=mcmc_samples) return model
[docs]def run_inference( # pyre-fixme[24]: Generic type `Callable` expects 2 type parameters. pyro_model: Callable, X: Tensor, Y: Tensor, Yvar: Tensor, num_samples: int = 256, warmup_steps: int = 512, thinning: int = 16, use_input_warping: bool = False, max_tree_depth: int = 6, disable_progbar: bool = False, gp_kernel: str = "matern", verbose: bool = False, task_feature: Optional[int] = None, rank: Optional[int] = None, jit_compile: bool = False, ) -> Dict[str, Tensor]: start = time.time() try: from pyro.infer.mcmc import MCMC, NUTS from pyro.infer.mcmc.util import print_summary except ImportError: raise RuntimeError("Cannot call run_inference without pyro installed!") kernel = NUTS( pyro_model, jit_compile=jit_compile, full_mass=True, ignore_jit_warnings=True, max_tree_depth=max_tree_depth, ) mcmc = MCMC( kernel, warmup_steps=warmup_steps, num_samples=num_samples, disable_progbar=disable_progbar, ) mcmc.run( X, Y, Yvar, use_input_warping=use_input_warping, gp_kernel=gp_kernel, task_feature=task_feature, rank=rank, ) # compute the true lengthscales and get rid of the temporary variables samples = mcmc.get_samples() inv_length_sq = ( samples["kernel_tausq"].unsqueeze(-1) * samples["_kernel_inv_length_sq"] ) samples["lengthscale"] = (1.0 / inv_length_sq).sqrt() # pyre-ignore [16] del samples["kernel_tausq"], samples["_kernel_inv_length_sq"] # this prints the summary if verbose: orig_std_out = sys.stdout.write sys.stdout.write = logger.info # pyre-fixme[8] print_summary(samples, prob=0.9, group_by_chain=False) sys.stdout.write = orig_std_out logger.info(f"MCMC elapsed time: {time.time() - start}") # thin for k, v in samples.items(): samples[k] = v[::thinning] # apply thinning return samples
[docs]def get_fully_bayesian_acqf( model: Model, objective_weights: Tensor, outcome_constraints: Optional[Tuple[Tensor, Tensor]] = None, X_observed: Optional[Tensor] = None, X_pending: Optional[Tensor] = None, **kwargs: Any, ) -> AcquisitionFunction: """NOTE: An `acqf_constructor` with which the underlying acquisition function is constructed is optionally extracted from `kwargs` and defaults to NEI. We did not add `acqf_constructor` directly to the argument list of `get_fully_bayesian_acqf` so that it satisfies the `TAcqfConstructor` Protocol that is shared by all other legacy Ax acquisition function constructors. """ kwargs["marginalize_dim"] = -3 acqf_constructor: TAcqfConstructor = kwargs.pop("acqf_constructor", get_qLogNEI) acqf = acqf_constructor( model=model, objective_weights=objective_weights, outcome_constraints=outcome_constraints, X_observed=X_observed, X_pending=X_pending, **kwargs, ) base_forward = acqf.forward # enabling manual override, default to True for LogEI log = kwargs.pop("log", acqf._log) sample_reduction = torch.mean if not log else logmeanexp # pyre-fixme[53]: Captured variable `base_forward` is not annotated. # pyre-fixme[3]: Return type must be annotated. # pyre-fixme[2]: Parameter must be annotated. def forward(self, X): # unsqueeze dim for GP hyperparameter samples return sample_reduction(base_forward(X.unsqueeze(-3)), dim=-1) acqf.forward = types.MethodType(forward, acqf) # pyre-ignore[8] return acqf
[docs]def get_fully_bayesian_acqf_nehvi( model: Model, objective_weights: Tensor, outcome_constraints: Optional[Tuple[Tensor, Tensor]] = None, X_observed: Optional[Tensor] = None, X_pending: Optional[Tensor] = None, **kwargs: Any, ) -> AcquisitionFunction: return get_fully_bayesian_acqf( model=model, objective_weights=objective_weights, outcome_constraints=outcome_constraints, X_observed=X_observed, X_pending=X_pending, acqf_constructor=get_qLogNEHVI, **kwargs, )
[docs]class FullyBayesianBotorchModelMixin: _model: Optional[Model] = None
[docs] def feature_importances(self) -> np.ndarray: if self._model is None: raise RuntimeError( "Cannot calculate feature_importances without a fitted model" ) elif isinstance(self._model, ModelListGP): models = self._model.models else: models = [self._model] lengthscales = [] for m in models: ls = m.covar_module.base_kernel.lengthscale lengthscales.append(ls) lengthscales = torch.stack(lengthscales, dim=0) # take mean over MCMC samples lengthscales = torch.quantile(lengthscales, 0.5, dim=1) # pyre-ignore [16] # pyre-fixme[58]: `/` is not supported for operand types `int` and `Tensor`. return (1 / lengthscales).detach().cpu().numpy()
[docs]class FullyBayesianBotorchModel(FullyBayesianBotorchModelMixin, BotorchModel): r"""Fully Bayesian Model that uses NUTS to sample from hyperparameter posterior. This includes support for using sparse axis-aligned subspace priors (SAAS). See [Eriksson2021saasbo]_ for details. """ def __init__( self, model_constructor: TModelConstructor = get_and_fit_model_mcmc, model_predictor: TModelPredictor = predict_from_model_mcmc, acqf_constructor: TAcqfConstructor = get_fully_bayesian_acqf, # pyre-fixme[9]: acqf_optimizer declared/used type mismatch acqf_optimizer: TOptimizer = scipy_optimizer, best_point_recommender: TBestPointRecommender = recommend_best_observed_point, refit_on_cv: bool = False, warm_start_refitting: bool = True, use_input_warping: bool = False, # use_saas is deprecated. TODO: remove use_saas: Optional[bool] = None, num_samples: int = 256, warmup_steps: int = 512, thinning: int = 16, max_tree_depth: int = 6, disable_progbar: bool = False, gp_kernel: str = "matern", verbose: bool = False, jit_compile: bool = False, **kwargs: Any, ) -> None: """Initialize Fully Bayesian Botorch Model. Args: model_constructor: A callable that instantiates and fits a model on data, with signature as described below. model_predictor: A callable that predicts using the fitted model, with signature as described below. acqf_constructor: A callable that creates an acquisition function from a fitted model, with signature as described below. acqf_optimizer: A callable that optimizes the acquisition function, with signature as described below. best_point_recommender: A callable that recommends the best point, with signature as described below. refit_on_cv: If True, refit the model for each fold when performing cross-validation. warm_start_refitting: If True, start model refitting from previous model parameters in order to speed up the fitting process. use_input_warping: A boolean indicating whether to use input warping use_saas: [deprecated] A boolean indicating whether to use the SAAS model num_samples: The number of MCMC samples. Note that with thinning, num_samples/thinning samples are retained. warmup_steps: The number of burn-in steps for NUTS. thinning: The amount of thinning. Every nth sample is retained. max_tree_depth: The max_tree_depth for NUTS. disable_progbar: A boolean indicating whether to print the progress bar and diagnostics during MCMC. gp_kernel: The type of ARD base kernel. "matern" corresponds to a Matern-5/2 kernel and "rbf" corresponds to an RBF kernel. verbose: A boolean indicating whether to print summary stats from MCMC. """ # use_saas is deprecated. TODO: remove if use_saas is not None: warnings.warn(SAAS_DEPRECATION_MSG, DeprecationWarning) BotorchModel.__init__( self, model_constructor=model_constructor, model_predictor=model_predictor, acqf_constructor=acqf_constructor, acqf_optimizer=acqf_optimizer, best_point_recommender=best_point_recommender, refit_on_cv=refit_on_cv, warm_start_refitting=warm_start_refitting, use_input_warping=use_input_warping, num_samples=num_samples, warmup_steps=warmup_steps, thinning=thinning, max_tree_depth=max_tree_depth, disable_progbar=disable_progbar, gp_kernel=gp_kernel, verbose=verbose, jit_compile=jit_compile, )
[docs]class FullyBayesianMOOBotorchModel( FullyBayesianBotorchModelMixin, MultiObjectiveBotorchModel ): r"""Fully Bayesian Model that uses qNEHVI. This includes support for using qNEHVI + SAASBO as in [Eriksson2021nas]_. """ @copy_doc(FullyBayesianBotorchModel.__init__) def __init__( self, model_constructor: TModelConstructor = get_and_fit_model_mcmc, model_predictor: TModelPredictor = predict_from_model_mcmc, acqf_constructor: TAcqfConstructor = get_fully_bayesian_acqf_nehvi, # pyre-fixme[9]: acqf_optimizer has type `Callable[[AcquisitionFunction, # Tensor, int, Optional[Dict[int, float]], Optional[Callable[[Tensor], # Tensor]], Any], Tensor]`; used as `Callable[[AcquisitionFunction, Tensor, # int, Optional[Dict[int, float]], Optional[Callable[[Tensor], Tensor]], # **(Any)], Tensor]`. acqf_optimizer: TOptimizer = scipy_optimizer, # TODO: Remove best_point_recommender for botorch_moo. Used in modelbridge._gen. best_point_recommender: TBestPointRecommender = recommend_best_observed_point, frontier_evaluator: TFrontierEvaluator = pareto_frontier_evaluator, refit_on_cv: bool = False, warm_start_refitting: bool = False, use_input_warping: bool = False, num_samples: int = 256, warmup_steps: int = 512, thinning: int = 16, max_tree_depth: int = 6, # use_saas is deprecated. TODO: remove use_saas: Optional[bool] = None, disable_progbar: bool = False, gp_kernel: str = "matern", verbose: bool = False, jit_compile: bool = False, **kwargs: Any, ) -> None: # use_saas is deprecated. TODO: remove if use_saas is not None: warnings.warn(SAAS_DEPRECATION_MSG, DeprecationWarning) MultiObjectiveBotorchModel.__init__( self, model_constructor=model_constructor, model_predictor=model_predictor, acqf_constructor=acqf_constructor, acqf_optimizer=acqf_optimizer, best_point_recommender=best_point_recommender, frontier_evaluator=frontier_evaluator, refit_on_cv=refit_on_cv, warm_start_refitting=warm_start_refitting, use_input_warping=use_input_warping, num_samples=num_samples, warmup_steps=warmup_steps, thinning=thinning, max_tree_depth=max_tree_depth, disable_progbar=disable_progbar, gp_kernel=gp_kernel, verbose=verbose, jit_compile=jit_compile, )