Source code for ax.plot.pareto_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 copy
from copy import deepcopy
from itertools import combinations
from logging import Logger
from typing import cast, Dict, List, NamedTuple, Optional, Tuple, Union

import numpy as np
import torch
from ax.core.batch_trial import BatchTrial
from ax.core.data import Data
from ax.core.experiment import Experiment
from ax.core.metric import Metric
from ax.core.objective import ScalarizedObjective
from ax.core.observation import ObservationFeatures
from ax.core.optimization_config import (
    MultiObjectiveOptimizationConfig,
    OptimizationConfig,
)
from ax.core.outcome_constraint import (
    ComparisonOp,
    ObjectiveThreshold,
    OutcomeConstraint,
)
from ax.core.search_space import RobustSearchSpace, SearchSpace
from ax.core.types import TParameterization
from ax.exceptions.core import AxError, UnsupportedError, UserInputError
from ax.modelbridge.modelbridge_utils import (
    _get_modelbridge_training_data,
    get_pareto_frontier_and_configs,
    observed_pareto_frontier,
)
from ax.modelbridge.registry import Models
from ax.modelbridge.torch import TorchModelBridge
from ax.modelbridge.transforms.search_space_to_float import SearchSpaceToFloat
from ax.models.torch.posterior_mean import get_PosteriorMean
from ax.models.torch_base import TorchModel
from ax.utils.common.logger import get_logger
from ax.utils.stats.statstools import relativize
from botorch.utils.multi_objective import is_non_dominated
from botorch.utils.multi_objective.hypervolume import infer_reference_point

# type aliases
Mu = Dict[str, List[float]]
Cov = Dict[str, Dict[str, List[float]]]


logger: Logger = get_logger(__name__)


def _extract_observed_pareto_2d(
    Y: np.ndarray,
    reference_point: Optional[Tuple[float, float]],
    minimize: Union[bool, Tuple[bool, bool]] = True,
) -> np.ndarray:
    if Y.shape[1] != 2:
        raise NotImplementedError("Currently only the 2-dim case is handled.")
    # If `minimize` is a bool, apply to both dimensions
    if isinstance(minimize, bool):
        minimize = (minimize, minimize)
    Y_copy = deepcopy(torch.from_numpy(Y).to())
    if reference_point:
        ref_point = torch.tensor(reference_point, dtype=Y_copy.dtype)
        for i in range(2):
            # Filter based on reference point
            Y_copy = (
                Y_copy[Y_copy[:, i] < ref_point[i]]
                if minimize[i]
                else Y_copy[Y_copy[:, i] > ref_point[i]]
            )
    for i in range(2):
        # Flip sign in each dimension based on minimize
        Y_copy[:, i] *= (-1) ** minimize[i]
    Y_pareto = Y_copy[is_non_dominated(Y_copy)]
    Y_pareto = Y_pareto[torch.argsort(input=Y_pareto[:, 0], descending=True)]
    for i in range(2):
        # Flip sign back
        Y_pareto[:, i] *= (-1) ** minimize[i]

    assert Y_pareto.shape[1] == 2  # Y_pareto should have two outcomes.
    return Y_pareto.detach().cpu().numpy()


[docs]class ParetoFrontierResults(NamedTuple): """Container for results from Pareto frontier computation. Fields are: - param_dicts: The parameter dicts of the points generated on the Pareto Frontier. - means: The posterior mean predictions of the model for each metric (same order as the param dicts). These must be as a percent change relative to status quo for any metric not listed in absolute_metrics. - sems: The posterior sem predictions of the model for each metric (same order as the param dicts). Also must be relativized wrt status quo for any metric not listed in absolute_metrics. - primary_metric: The name of the primary metric. - secondary_metric: The name of the secondary metric. - absolute_metrics: List of outcome metrics that are NOT be relativized w.r.t. the status quo. All other metrics are assumed to be given here as % relative to status_quo. - objective_thresholds: Threshold for each objective. Must be on the same scale as means, so if means is relativized it should be the relative value, otherwise it should be absolute. - arm_names: Optional list of arm names for each parameterization. """ param_dicts: List[TParameterization] means: Dict[str, List[float]] sems: Dict[str, List[float]] primary_metric: str secondary_metric: str absolute_metrics: List[str] objective_thresholds: Optional[Dict[str, float]] arm_names: Optional[List[Optional[str]]]
def _extract_sq_data( experiment: Experiment, data: Data ) -> Tuple[Dict[str, float], Dict[str, float]]: """ Returns sq_means and sq_sems, each a mapping from metric name to, respectively, mean and sem of the status quo arm. Empty dictionaries if no SQ arm. """ sq_means = {} sq_sems = {} if experiment.status_quo is not None: # Extract SQ values sq_df = data.df[ data.df["arm_name"] == experiment.status_quo.name # pyre-ignore ] for metric, metric_df in sq_df.groupby("metric_name"): sq_means[metric] = metric_df["mean"].values[0] sq_sems[metric] = metric_df["sem"].values[0] return sq_means, sq_sems def _relativize_values( means: List[float], sq_mean: float, sems: List[float], sq_sem: float ) -> Tuple[List[float], List[float]]: """ Relativize values, using delta method if SEMs provided, or just by relativizing means if not. Relativization is as percent. """ if np.isnan(sq_sem) or np.isnan(sems).any(): # Just relativize means means = [(mu / sq_mean - 1) * 100 for mu in means] else: # Use delta method means_arr, sems_arr = relativize( means_t=np.array(means), sems_t=np.array(sems), mean_c=sq_mean, sem_c=sq_sem, as_percent=True, ) means, sems = list(means), list(sems) return means, sems
[docs]def get_observed_pareto_frontiers( experiment: Experiment, data: Optional[Data] = None, rel: Optional[bool] = None, arm_names: Optional[List[str]] = None, ) -> List[ParetoFrontierResults]: """ Find all Pareto points from an experiment. Uses only values as observed in the data; no modeling is involved. Makes no assumption about the search space or types of parameters. If "data" is provided will use that, otherwise will use all data already attached to the experiment. Uses all arms present in data; does not filter according to experiment search space. If arm_names is specified, will filter to just those arm whose names are given in the list. Assumes experiment has a multiobjective optimization config from which the objectives and outcome constraints will be extracted. Will generate a ParetoFrontierResults for every pair of metrics in the experiment's multiobjective optimization config. Args: experiment: The experiment. data: Data to use for computing Pareto frontier. If not provided, will lookup data from experiment. rel: Relativize results wrt experiment status quo. If None, then rel will be taken for each objective separately from its own objective threshold. `rel` must be specified if there are missing objective thresholds. arm_names: If provided, computes Pareto frontier only from among the provided list of arm names, plus status quo if set on experiment. Returns: ParetoFrontierResults that can be used with interact_pareto_frontier. """ if data is None: data = experiment.lookup_data() if experiment.optimization_config is None: raise ValueError("Experiment must have an optimization config") if arm_names is not None: if ( experiment.status_quo is not None and experiment.status_quo.name not in arm_names ): # Make sure status quo is always included, for derelativization arm_names.append(experiment.status_quo.name) data = Data(data.df[data.df["arm_name"].isin(arm_names)]) mb = get_tensor_converter_model(experiment=experiment, data=data) pareto_observations = observed_pareto_frontier(modelbridge=mb) # Convert to ParetoFrontierResults objective_metric_names = { metric.name for metric in experiment.optimization_config.objective.metrics # pyre-ignore } obj_metr_list = sorted(objective_metric_names) pfr_means = {name: [] for name in obj_metr_list} pfr_sems = {name: [] for name in obj_metr_list} for obs in pareto_observations: for i, name in enumerate(obs.data.metric_names): if name in objective_metric_names: pfr_means[name].append(obs.data.means[i]) pfr_sems[name].append(np.sqrt(obs.data.covariance[i, i])) # Get objective thresholds rel_objth = {} objective_thresholds = {} if experiment.optimization_config.objective_thresholds is not None: # pyre-ignore for objth in experiment.optimization_config.objective_thresholds: rel_objth[objth.metric.name] = objth.relative objective_thresholds[objth.metric.name] = objth.bound # Identify which metrics should be relativized if rel in [True, False]: metric_is_rel = {name: rel for name in pfr_means} else: if len(rel_objth) != len(pfr_means): raise UserInputError( "At least one objective is missing an objective threshold. " "`rel` must be specified as True or False when there are missing " "objective thresholds." ) # Default to however the threshold is specified metric_is_rel = rel_objth # Compute SQ values sq_means, sq_sems = _extract_sq_data(experiment, data) # Relativize data and thresholds as needed for name in pfr_means: if metric_is_rel[name]: pfr_means[name], pfr_sems[name] = _relativize_values( means=pfr_means[name], sq_mean=sq_means[name], sems=pfr_sems[name], sq_sem=sq_sems[name], ) if name in objective_thresholds and not rel_objth[name]: # Metric is rel but obj th is not. # Need to relativize the objective threshold objective_thresholds[name] = _relativize_values( means=[objective_thresholds[name]], sq_mean=sq_means[name], sems=[np.nan], sq_sem=np.nan, )[0][0] elif name in objective_thresholds and rel_objth[name]: # Metric is not rel but obj th is, so need to derelativize obj th objective_thresholds[name] = ( 1 + objective_thresholds[name] / 100.0 ) * sq_means[name] absolute_metrics = [name for name, val in metric_is_rel.items() if not val] # Construct ParetoFrontResults for each pair pfr_list = [] param_dicts = [obs.features.parameters for obs in pareto_observations] pfr_arm_names = [obs.arm_name for obs in pareto_observations] for metric_a, metric_b in combinations(obj_metr_list, 2): pfr_list.append( ParetoFrontierResults( param_dicts=param_dicts, means=pfr_means, sems=pfr_sems, primary_metric=metric_a, secondary_metric=metric_b, absolute_metrics=absolute_metrics, objective_thresholds=objective_thresholds, arm_names=pfr_arm_names, ) ) return pfr_list
[docs]def to_nonrobust_search_space(search_space: SearchSpace) -> SearchSpace: """Reduces a RobustSearchSpace to a SearchSpace. This is a no-op for all other search spaces. """ if isinstance(search_space, RobustSearchSpace): return SearchSpace( parameters=[p.clone() for p in search_space._parameters.values()], parameter_constraints=[ pc.clone() for pc in search_space._parameter_constraints ], ) else: return search_space
[docs]def get_tensor_converter_model(experiment: Experiment, data: Data) -> TorchModelBridge: """ Constructs a minimal model for converting things to tensors. Model fitting will instantiate all of the transforms but will not do any expensive (i.e. GP) fitting beyond that. The model will raise an error if it is used for predicting or generating. Will work for any search space regardless of types of parameters. Args: experiment: Experiment. data: Data for fitting the model. Returns: A torch modelbridge with transforms set. """ # Transforms is the minimal set that will work for converting any search # space to tensors. return TorchModelBridge( experiment=experiment, search_space=to_nonrobust_search_space(experiment.search_space), data=data, model=TorchModel(), transforms=[SearchSpaceToFloat], fit_out_of_design=True, )
[docs]def compute_posterior_pareto_frontier( experiment: Experiment, primary_objective: Metric, secondary_objective: Metric, data: Optional[Data] = None, outcome_constraints: Optional[List[OutcomeConstraint]] = None, absolute_metrics: Optional[List[str]] = None, num_points: int = 10, trial_index: Optional[int] = None, chebyshev: bool = True, ) -> ParetoFrontierResults: """Compute the Pareto frontier between two objectives. For experiments with batch trials, a trial index or data object must be provided. This is done by fitting a GP and finding the pareto front according to the GP posterior mean. Args: experiment: The experiment to compute a pareto frontier for. primary_objective: The primary objective to optimize. secondary_objective: The secondary objective against which to trade off the primary objective. outcome_constraints: Outcome constraints to be respected by the optimization. Can only contain constraints on metrics that are not primary or secondary objectives. absolute_metrics: List of outcome metrics that should NOT be relativized w.r.t. the status quo (all other outcomes will be in % relative to status_quo). num_points: The number of points to compute on the Pareto frontier. chebyshev: Whether to use augmented_chebyshev_scalarization when computing Pareto Frontier points. Returns: ParetoFrontierResults: A NamedTuple with fields listed in its definition. """ model_gen_options = { "acquisition_function_kwargs": {"chebyshev_scalarization": chebyshev} } if ( trial_index is None and data is None and any(isinstance(t, BatchTrial) for t in experiment.trials.values()) ): raise UnsupportedError( "Must specify trial index or data for experiment with batch trials" ) absolute_metrics = [] if absolute_metrics is None else absolute_metrics for metric in absolute_metrics: if metric not in experiment.metrics: raise ValueError(f"Model was not fit on metric `{metric}`") if outcome_constraints is None: outcome_constraints = [] else: # ensure we don't constrain an objective _validate_outcome_constraints( outcome_constraints=outcome_constraints, primary_objective=primary_objective, secondary_objective=secondary_objective, ) # build posterior mean model if not data: try: data = ( experiment.trials[trial_index].fetch_data() if trial_index else experiment.fetch_data() ) except Exception as e: logger.info(f"Could not fetch data from experiment or trial: {e}") # The weights here are just dummy weights that we pass in to construct the # modelbridge. We set the weight to -1 if `lower_is_better` is `True` and # 1 otherwise. This code would benefit from a serious revamp. oc = _build_new_optimization_config( weights=np.array( [ -1 if primary_objective.lower_is_better else 1, -1 if secondary_objective.lower_is_better else 1, ] ), primary_objective=primary_objective, secondary_objective=secondary_objective, outcome_constraints=outcome_constraints, ) model = Models.MOO( experiment=experiment, data=data, acqf_constructor=get_PosteriorMean, optimization_config=oc, ) status_quo = experiment.status_quo if status_quo: try: status_quo_prediction = model.predict( [ ObservationFeatures( parameters=status_quo.parameters, trial_index=trial_index, ) ] ) except ValueError as e: logger.warning(f"Could not predict OOD status_quo outcomes: {e}") status_quo = None status_quo_prediction = None else: status_quo_prediction = None param_dicts: List[TParameterization] = [] # Construct weightings with linear angular spacing. # TODO: Verify whether 0, 1 weights cause problems because of subset_model. alpha = np.linspace(0 + 0.01, np.pi / 2 - 0.01, num_points) primary_weight = (-1 if primary_objective.lower_is_better else 1) * np.cos(alpha) secondary_weight = (-1 if secondary_objective.lower_is_better else 1) * np.sin( alpha ) weights_list = np.stack([primary_weight, secondary_weight]).transpose() for weights in weights_list: outcome_constraints = outcome_constraints oc = _build_new_optimization_config( weights=weights, primary_objective=primary_objective, secondary_objective=secondary_objective, outcome_constraints=outcome_constraints, ) # TODO: (jej) T64002590 Let this serve as a starting point for optimization. # ex. Add global spacing criterion. Implement on BoTorch side. # pyre-fixme [6]: Expected different type for model_gen_options run = model.gen(1, model_gen_options=model_gen_options, optimization_config=oc) param_dicts.append(run.arms[0].parameters) # Call predict on points to get their decomposed metrics. means, cov = model.predict( [ObservationFeatures(parameters) for parameters in param_dicts] ) return _extract_pareto_frontier_results( param_dicts=param_dicts, means=means, variances=cov, primary_metric=primary_objective.name, secondary_metric=secondary_objective.name, absolute_metrics=absolute_metrics, outcome_constraints=outcome_constraints, status_quo_prediction=status_quo_prediction, )
def _extract_pareto_frontier_results( param_dicts: List[TParameterization], means: Mu, variances: Cov, primary_metric: str, secondary_metric: str, absolute_metrics: List[str], outcome_constraints: Optional[List[OutcomeConstraint]], status_quo_prediction: Optional[Tuple[Mu, Cov]], ) -> ParetoFrontierResults: """Extract prediction results into ParetoFrontierResults struture.""" metrics = list(means.keys()) means_out = {metric: m.copy() for metric, m in means.items()} sems_out = {metric: np.sqrt(v[metric]) for metric, v in variances.items()} # relativize predicted outcomes if requested primary_is_relative = primary_metric not in absolute_metrics secondary_is_relative = secondary_metric not in absolute_metrics # Relativized metrics require a status quo prediction if primary_is_relative or secondary_is_relative: if status_quo_prediction is None: raise AxError("Relativized metrics require a valid status quo prediction") sq_mean, sq_sem = status_quo_prediction for metric in metrics: if metric not in absolute_metrics and metric in sq_mean: means_out[metric], sems_out[metric] = relativize( means_t=means_out[metric], sems_t=sems_out[metric], mean_c=sq_mean[metric][0], sem_c=np.sqrt(sq_sem[metric][metric][0]), as_percent=True, ) return ParetoFrontierResults( param_dicts=param_dicts, means=means_out, sems=sems_out, primary_metric=primary_metric, secondary_metric=secondary_metric, absolute_metrics=absolute_metrics, objective_thresholds=None, arm_names=None, ) def _validate_outcome_constraints( outcome_constraints: List[OutcomeConstraint], primary_objective: Metric, secondary_objective: Metric, ) -> None: """Validate that outcome constraints don't involve objectives.""" objective_metrics = [primary_objective.name, secondary_objective.name] if outcome_constraints is not None: for oc in outcome_constraints: if oc.metric.name in objective_metrics: raise ValueError( "Metric `{metric_name}` occurs in both outcome constraints " "and objectives".format(metric_name=oc.metric.name) ) def _build_new_optimization_config( # pyre-fixme[2]: Parameter must be annotated. weights, # pyre-fixme[2]: Parameter must be annotated. primary_objective, # pyre-fixme[2]: Parameter must be annotated. secondary_objective, # pyre-fixme[2]: Parameter must be annotated. outcome_constraints=None, ) -> MultiObjectiveOptimizationConfig: obj = ScalarizedObjective( metrics=[primary_objective, secondary_objective], weights=weights, minimize=False, ) optimization_config = MultiObjectiveOptimizationConfig( objective=obj, outcome_constraints=outcome_constraints ) return optimization_config
[docs]def infer_reference_point_from_experiment( experiment: Experiment, ) -> List[ObjectiveThreshold]: """This functions is a wrapper around ``infer_reference_point`` to find the nadir point from the pareto front of an experiment. Aside from converting experiment to tensors, this wrapper transforms back and forth the objectives of the experiment so that they are appropriately used by ``infer_reference_point``. Args: experiment: The experiment for which we want to infer the reference point. Returns: A list of objective thresholds representing the reference point. """ if not experiment.is_moo_problem: raise ValueError( "This function works for MOO experiments only." f" Experiment {experiment.name} is single objective." ) # Reading experiment data. mb_reference = get_tensor_converter_model( experiment=experiment, data=experiment.fetch_data() ) obs_feats, obs_data, _ = _get_modelbridge_training_data(modelbridge=mb_reference) # Since objectives could have arbitrary orders in objective_thresholds and # further down the road `get_pareto_frontier_and_configs` arbitrarily changes the # orders of the objectives, we fix the objective orders here based on the # observation_data and maintain it throughout the flow. objective_orders = obs_data[0].metric_names # Defining a dummy reference point so that all observed points are considered # when calculating the Pareto front. Also, defining a multiplier to turn all # the objectives to be maximized. Note that the multiplier at this point # contains 0 for outcome_constraint metrics, but this will be dropped later. dummy_rp = copy.deepcopy( experiment.optimization_config.objective_thresholds # pyre-ignore ) multiplier = [0] * len(objective_orders) for ot in dummy_rp: # In the following, we find the index of the objective in # `objective_orders`. If there is an objective that does not exist # in `obs_data`, a ValueError is raised. try: objective_index = objective_orders.index(ot.metric.name) except ValueError: raise ValueError(f"Metric {ot.metric.name} does not exist in `obs_data`.") if ot.op == ComparisonOp.LEQ: ot.bound = np.inf multiplier[objective_index] = -1 else: ot.bound = -np.inf multiplier[objective_index] = 1 # Finding the pareto frontier frontier_observations, f, obj_w, _ = get_pareto_frontier_and_configs( modelbridge=mb_reference, observation_features=obs_feats, observation_data=obs_data, objective_thresholds=dummy_rp, use_model_predictions=False, ) if len(frontier_observations) == 0: opt_config = cast(OptimizationConfig, mb_reference._optimization_config) outcome_constraints = opt_config._outcome_constraints if len(outcome_constraints) == 0: raise RuntimeError( "No frontier observations found in the experiment and no constraints " "are present. Please check the data of the experiment." ) logger.warning( "No frontier observations found in the experiment. The likely cause is " "the absence of feasible arms in the experiment if a constraint is present." " Trying to find a reference point with the unconstrained objective values." ) opt_config._outcome_constraints = [] # removing the constraints # getting the unconstrained pareto frontier frontier_observations, f, obj_w, _ = get_pareto_frontier_and_configs( modelbridge=mb_reference, observation_features=obs_feats, observation_data=obs_data, objective_thresholds=dummy_rp, use_model_predictions=False, ) opt_config._outcome_constraints = outcome_constraints # restoring constraints # Need to reshuffle columns of `f` and `obj_w` to be consistent # with objective_orders. order = [ objective_orders.index(metric_name) for metric_name in frontier_observations[0].data.metric_names ] f = f[:, order] obj_w = obj_w[order] # Dropping the columns related to outcome constraints. f = f[:, obj_w.nonzero().view(-1)] multiplier_tensor = torch.tensor(multiplier, dtype=f.dtype, device=f.device) multiplier_nonzero = multiplier_tensor[obj_w.nonzero().view(-1)] # Transforming all the objectives to be maximized. f_transformed = multiplier_nonzero * f # Finding nadir point. rp_raw = infer_reference_point(f_transformed) # Un-transforming the reference point. rp = multiplier_nonzero * rp_raw # Removing the non-objective metrics form the order. objective_orders_reduced = [ x for (i, x) in enumerate(objective_orders) if multiplier[i] != 0 ] # Constructing the objective thresholds. # NOTE: This assumes that objective_thresholds is already initialized. nadir_objective_thresholds = copy.deepcopy( experiment.optimization_config.objective_thresholds ) for obj_threshold in nadir_objective_thresholds: obj_threshold.bound = rp[ objective_orders_reduced.index(obj_threshold.metric.name) ].item() return nadir_objective_thresholds