Source code for ax.modelbridge.dispatch_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 logging
import warnings
from math import ceil
from typing import Any, cast, Dict, Optional, Type, Union

import torch
from ax.core.experiment import Experiment
from ax.core.optimization_config import OptimizationConfig
from ax.core.parameter import ChoiceParameter, ParameterType, RangeParameter
from ax.core.search_space import SearchSpace
from ax.modelbridge.generation_strategy import GenerationStep, GenerationStrategy
from ax.modelbridge.registry import MODEL_KEY_TO_MODEL_SETUP, Models
from ax.modelbridge.transforms.base import Transform
from ax.modelbridge.transforms.winsorize import Winsorize
from ax.models.torch.botorch_modular.model import BoTorchModel as ModularBoTorchModel
from ax.models.types import TConfig
from ax.models.winsorization_config import WinsorizationConfig
from ax.utils.common.logger import get_logger
from ax.utils.common.typeutils import not_none


logger: logging.Logger = get_logger(__name__)


DEFAULT_BAYESIAN_PARALLELISM = 3
# `BO_MIXED` optimizes all range parameters once for each combination of choice
# parameters, then takes the optimum of those optima. The cost associated with this
# method grows with the number of combinations, and so it is only used when the
# number of enumerated discrete combinations is below some maximum value.
MAX_DISCRETE_ENUMERATIONS_MIXED = 65
MAX_DISCRETE_ENUMERATIONS_NO_CONTINUOUS_OPTIMIZATION = 1e4
MAX_ONE_HOT_ENCODINGS_CONTINUOUS_OPTIMIZATION = 33
SAASBO_INCOMPATIBLE_MESSAGE = (
    "SAASBO is incompatible with {} generation strategy. "
    "Disregarding user input `use_saasbo = True`."
)


def _make_sobol_step(
    num_trials: int = -1,
    min_trials_observed: Optional[int] = None,
    enforce_num_trials: bool = True,
    max_parallelism: Optional[int] = None,
    seed: Optional[int] = None,
    should_deduplicate: bool = False,
) -> GenerationStep:
    """Shortcut for creating a Sobol generation step."""
    return GenerationStep(
        model=Models.SOBOL,
        num_trials=num_trials,
        # NOTE: ceil(-1 / 2) = 0, so this is safe to do when num trials is -1.
        min_trials_observed=min_trials_observed or ceil(num_trials / 2),
        enforce_num_trials=enforce_num_trials,
        max_parallelism=max_parallelism,
        model_kwargs={"deduplicate": True, "seed": seed},
        should_deduplicate=should_deduplicate,
    )


def _make_botorch_step(
    num_trials: int = -1,
    min_trials_observed: Optional[int] = None,
    enforce_num_trials: bool = True,
    max_parallelism: Optional[int] = None,
    model: Models = Models.BOTORCH_MODULAR,
    model_kwargs: Optional[Dict[str, Any]] = None,
    winsorization_config: Optional[
        Union[WinsorizationConfig, Dict[str, WinsorizationConfig]]
    ] = None,
    no_winsorization: bool = False,
    should_deduplicate: bool = False,
    verbose: Optional[bool] = None,
    disable_progbar: Optional[bool] = None,
    jit_compile: Optional[bool] = None,
    derelativize_with_raw_status_quo: bool = False,
) -> GenerationStep:
    """Shortcut for creating a BayesOpt generation step."""
    model_kwargs = model_kwargs or {}

    winsorization_transform_config = _get_winsorization_transform_config(
        winsorization_config=winsorization_config,
        no_winsorization=no_winsorization,
        derelativize_with_raw_status_quo=derelativize_with_raw_status_quo,
    )

    derelativization_transform_config = {
        "use_raw_status_quo": derelativize_with_raw_status_quo
    }
    model_kwargs["transform_configs"] = model_kwargs.get("transform_configs", {})
    model_kwargs["transform_configs"][
        "Derelativize"
    ] = derelativization_transform_config

    if not no_winsorization:
        _, default_bridge_kwargs = model.view_defaults()
        default_transforms = default_bridge_kwargs["transforms"]
        transforms = model_kwargs.get("transforms", default_transforms)
        model_kwargs["transforms"] = [cast(Type[Transform], Winsorize)] + transforms
        if winsorization_transform_config is not None:
            model_kwargs["transform_configs"][
                "Winsorize"
            ] = winsorization_transform_config

    if MODEL_KEY_TO_MODEL_SETUP[model.value].model_class != ModularBoTorchModel:
        if verbose is not None:
            model_kwargs.update({"verbose": verbose})
        if disable_progbar is not None:
            model_kwargs.update({"disable_progbar": disable_progbar})
        if jit_compile is not None:
            model_kwargs.update({"jit_compile": jit_compile})
    else:
        # TODO[T164389105] Rewrite choose_generation_strategy to be MBM first
        logger.info(
            "`verbose`, `disable_progbar`, and `jit_compile` are not yet supported "
            "when using `choose_generation_strategy` with ModularBoTorchModel, "
            "dropping these arguments."
        )
    return GenerationStep(
        model=model,
        num_trials=num_trials,
        # NOTE: ceil(-1 / 2) = 0, so this is safe to do when num trials is -1.
        min_trials_observed=min_trials_observed or ceil(num_trials / 2),
        enforce_num_trials=enforce_num_trials,
        max_parallelism=max_parallelism,
        # `model_kwargs` should default to `None` if empty
        model_kwargs=model_kwargs if len(model_kwargs) > 0 else None,
        should_deduplicate=should_deduplicate,
    )


def _suggest_gp_model(
    search_space: SearchSpace,
    num_trials: Optional[int] = None,
    optimization_config: Optional[OptimizationConfig] = None,
    use_saasbo: bool = False,
) -> Union[None, Models]:
    """Suggest a model based on the search space. None means we use Sobol.

    1. We use Sobol if the number of total iterations in the optimization is
    known in advance and there are fewer distinct points in the search space
    than the known intended number of total iterations.
    2. We use ``BO_MIXED`` if there are fewer ordered parameters in the search space
    than the sum of options for the *unordered* choice parameters, and the number
    of discrete enumerations to be performed by the optimizer is less than
    ``MAX_DISCRETE_ENUMERATIONS_MIXED``, or if there are only choice parameters and
    the number of choice combinations to enumerate is less than
    ``MAX_DISCRETE_ENUMERATIONS_CHOICE_ONLY``. Note that we do not count 2-level choice
    parameters as unordered, since these do not affect the modeling choice.
    3. If there are more ordered parameters in the search space than the sum of options
    for the *unordered* choice parameters, or if there is at least one ordered
    parameter and the number of parameters needed to encode all unordered parameters
    is less than ``MAX_ONE_HOT_ENCODINGS_CONTINUOUS_OPTIMIZATION``, we use BO with
    continuous relaxation.
    * If ``optimization_config`` has multiple objectives, we use ``MOO`` if
    ``use_saasbo is False`` and ``FULLYBAYESIANMOO`` otherwise.
    * Otherwise, we use ``GPEI`` if ``use_saasbo is False`` and ``FULLYBAYESIAN``
    otherwise.
    """
    # Count tunable parameter types.
    num_ordered_parameters = 0
    num_unordered_choices = 0
    num_enumerated_combinations = 1
    num_possible_points = 1
    all_range_parameters_are_discrete = True
    all_parameters_are_enumerated = True
    for parameter in search_space.tunable_parameters.values():
        should_enumerate_param = False
        num_param_discrete_values = None
        if isinstance(parameter, ChoiceParameter):
            should_enumerate_param = True
            num_param_discrete_values = len(parameter.values)
            num_possible_points *= num_param_discrete_values
            if parameter.is_ordered is False and num_param_discrete_values > 2:
                num_unordered_choices += num_param_discrete_values
            else:
                num_ordered_parameters += 1
        elif isinstance(parameter, RangeParameter):
            num_ordered_parameters += 1
            if parameter.parameter_type == ParameterType.FLOAT:
                all_range_parameters_are_discrete = False
            else:
                num_param_discrete_values = int(parameter.upper - parameter.lower) + 1
                num_possible_points *= num_param_discrete_values

        if should_enumerate_param:
            num_enumerated_combinations *= not_none(num_param_discrete_values)
        else:
            all_parameters_are_enumerated = False

    # Use Sobol if number of trials is known and sufficient to try all possible points.
    if (
        num_trials is not None
        and all_range_parameters_are_discrete
        and num_possible_points <= num_trials
    ):
        logger.info("Using Sobol since we can enumerate the search space.")
        if use_saasbo:
            logger.warning(SAASBO_INCOMPATIBLE_MESSAGE.format("Sobol"))
        return None

    # Use mixed Bayesian optimization when appropriate. This logic is currently tied to
    # the fact that acquisition function optimization for mixed BayesOpt currently
    # enumerates all combinations of choice parameters.
    # We use continuous relaxation if there are more ordered parameters than there
    # are choices for unordered parameters.
    if (num_ordered_parameters < num_unordered_choices) and (
        num_enumerated_combinations <= MAX_DISCRETE_ENUMERATIONS_MIXED
        or (
            all_parameters_are_enumerated
            and num_enumerated_combinations
            < MAX_DISCRETE_ENUMERATIONS_NO_CONTINUOUS_OPTIMIZATION
        )
    ):
        logger.info(
            "Using Bayesian optimization with a categorical kernel for improved "
            "performance with a large number of unordered categorical parameters."
        )
        if use_saasbo:
            logger.warning(SAASBO_INCOMPATIBLE_MESSAGE.format("`BO_MIXED`"))
        return Models.BO_MIXED

    if num_ordered_parameters >= num_unordered_choices or (
        num_unordered_choices < MAX_ONE_HOT_ENCODINGS_CONTINUOUS_OPTIMIZATION
        and num_ordered_parameters > 0
    ):
        # These use one-hot encoding for unordered choice parameters, resulting in a
        # total of num_unordered_choices OHE parameters.
        # So, we do not want to use them when there are too many unordered choices.
        method = Models.SAASBO if use_saasbo else Models.BOTORCH_MODULAR
        reason = (
            "there are more ordered parameters than there are categories for the "
            "unordered categorical parameters."
            if num_ordered_parameters >= num_unordered_choices
            else "there is at least one ordered parameter and there are fewer than "
            f"{MAX_ONE_HOT_ENCODINGS_CONTINUOUS_OPTIMIZATION} choices for "
            "unordered parameters."
        )
        logger.info(f"Using {method} since {reason}")
        return method

    logger.info(
        f"Using Sobol since there are more than {MAX_DISCRETE_ENUMERATIONS_MIXED} "
        "combinations of enumerated parameters. For improved performance, make sure "
        "that all ordered `ChoiceParameter`s are encoded as such (`is_ordered=True`), "
        "and use `RangeParameter`s in place of ordered `ChoiceParameter`s where "
        "possible. Also, consider removing some or all unordered `ChoiceParameter`s."
    )
    if use_saasbo:
        logger.warning(SAASBO_INCOMPATIBLE_MESSAGE.format("Sobol"))
    return None


[docs]def calculate_num_initialization_trials( num_tunable_parameters: int, num_trials: Optional[int], use_batch_trials: bool, ) -> int: """ Applies rules from high to low priority - 1 for batch trials. - At least 5 - At most 1/5th of num_trials. - Twice the number of tunable parameters """ if use_batch_trials: # Batched trials. return 1 ret = 2 * num_tunable_parameters if num_trials is not None: ret = min(ret, not_none(num_trials) // 5) return max(ret, 5)
[docs]def choose_generation_strategy( search_space: SearchSpace, *, use_batch_trials: bool = False, enforce_sequential_optimization: bool = True, random_seed: Optional[int] = None, torch_device: Optional[torch.device] = None, no_winsorization: bool = False, winsorization_config: Optional[ Union[WinsorizationConfig, Dict[str, WinsorizationConfig]] ] = None, derelativize_with_raw_status_quo: bool = False, no_bayesian_optimization: bool = False, num_trials: Optional[int] = None, num_initialization_trials: Optional[int] = None, num_completed_initialization_trials: int = 0, max_initialization_trials: Optional[int] = None, max_parallelism_cap: Optional[int] = None, max_parallelism_override: Optional[int] = None, optimization_config: Optional[OptimizationConfig] = None, should_deduplicate: bool = False, use_saasbo: bool = False, verbose: Optional[bool] = None, disable_progbar: Optional[bool] = None, jit_compile: Optional[bool] = None, experiment: Optional[Experiment] = None, ) -> GenerationStrategy: """Select an appropriate generation strategy based on the properties of the search space and expected settings of the experiment, such as number of arms per trial, optimization algorithm settings, expected number of trials in the experiment, etc. Args: search_space: SearchSpace, based on the properties of which to select the generation strategy. use_batch_trials: Whether this generation strategy will be used to generate batched trials instead of 1-arm trials. enforce_sequential_optimization: Whether to enforce that 1) the generation strategy needs to be updated with ``min_trials_observed`` observations for a given generation step before proceeding to the next one and 2) maximum number of trials running at once (max_parallelism) if enforced for the BayesOpt step. NOTE: ``max_parallelism_override`` and ``max_parallelism_cap`` settings will still take their effect on max parallelism even if ``enforce_sequential_optimization=False``, so if those settings are specified, max parallelism will be enforced. random_seed: Fixed random seed for the Sobol generator. torch_device: The device to use for generation steps implemented in PyTorch (e.g. via BoTorch). Some generation steps (in particular EHVI-based ones for multi-objective optimization) can be sped up by running candidate generation on the GPU. If not specified, uses the default torch device (usually the CPU). no_winsorization: Whether to apply the winsorization transform prior to applying other transforms for fitting the BoTorch model. winsorization_config: Explicit winsorization settings, if winsorizing. Usually only `upper_quantile_margin` is set when minimizing, and only `lower_quantile_margin` when maximizing. derelativize_with_raw_status_quo: Whether to derelativize using the raw status quo values in any transforms. This argument is primarily to allow automatic Winsorization when relative constraints are present. Note: automatic Winsorization will fail if this is set to `False` (or unset) and there are relative constraints present. no_bayesian_optimization: If True, Bayesian optimization generation strategy will not be suggested and quasi-random strategy will be used. num_trials: Total number of trials in the optimization, if known in advance. num_initialization_trials: Specific number of initialization trials, if wanted. Typically, initialization trials are generated quasi-randomly. max_initialization_trials: If ``num_initialization_trials`` unspecified, it will be determined automatically. This arg provides a cap on that automatically determined number. num_completed_initialization_trials: The final calculated number of initialization trials is reduced by this number. This is useful when warm-starting an experiment, to specify what number of completed trials can be used to satisfy the initialization_trial requirement. max_parallelism_cap: Integer cap on parallelism in this generation strategy. If specified, ``max_parallelism`` setting in each generation step will be set to the minimum of the default setting for that step and the value of this cap. ``max_parallelism_cap`` is meant to just be a hard limit on parallelism (e.g. to avoid overloading machine(s) that evaluate the experiment trials). Specify only if not specifying ``max_parallelism_override``. max_parallelism_override: Integer, with which to override the default max parallelism setting for all steps in the generation strategy returned from this function. Each generation step has a ``max_parallelism`` value, which restricts how many trials can run simultaneously during a given generation step. By default, the parallelism setting is chosen as appropriate for the model in a given generation step. If ``max_parallelism_override`` is -1, no max parallelism will be enforced for any step of the generation strategy. Be aware that parallelism is limited to improve performance of Bayesian optimization, so only disable its limiting if necessary. optimization_config: used to infer whether to use MOO and will be passed in to ``Winsorize`` via its ``transform_config`` in order to determine default winsorization behavior when necessary. should_deduplicate: Whether to deduplicate the parameters of proposed arms against those of previous arms via rejection sampling. If this is True, the generation strategy will discard generator runs produced from the generation step that has `should_deduplicate=True` if they contain arms already present on the experiment and replace them with new generator runs. If no generator run with entirely unique arms could be produced in 5 attempts, a `GenerationStrategyRepeatedPoints` error will be raised, as we assume that the optimization converged when the model can no longer suggest unique arms. use_saasbo: Whether to use SAAS prior for any GPEI generation steps. verbose: Whether GP model should produce verbose logs. If not ``None``, its value gets added to ``model_kwargs`` during ``generation_strategy`` construction. Defaults to ``True`` for SAASBO, else ``None``. Verbose outputs are currently only available for SAASBO, so if ``verbose is not None`` for a different model type, it will be overridden to ``None`` with a warning. disable_progbar: Whether GP model should produce a progress bar. If not ``None``, its value gets added to ``model_kwargs`` during ``generation_strategy`` construction. Defaults to ``True`` for SAASBO, else ``None``. Progress bars are currently only available for SAASBO, so if ``disable_probar is not None`` for a different model type, it will be overridden to ``None`` with a warning. jit_compile: Whether to use jit compilation in Pyro when SAASBO is used. experiment: If specified, ``_experiment`` attribute of the generation strategy will be set to this experiment (useful for associating a generation strategy with a given experiment before it's first used to ``gen`` with that experiment). Can also provide `optimization_config` if it is not provided as an arg to this function. """ if experiment is not None and optimization_config is None: optimization_config = experiment.optimization_config suggested_model = _suggest_gp_model( search_space=search_space, num_trials=num_trials, optimization_config=optimization_config, use_saasbo=use_saasbo, ) # Determine max parallelism for the generation steps. if max_parallelism_override == -1: # `max_parallelism_override` of -1 means no max parallelism enforcement in # the generation strategy, which means `max_parallelism=None` in gen. steps. sobol_parallelism = bo_parallelism = None elif max_parallelism_override is not None: sobol_parallelism = bo_parallelism = max_parallelism_override elif max_parallelism_cap is not None: # Max parallelism override is None by now sobol_parallelism = max_parallelism_cap bo_parallelism = min(max_parallelism_cap, DEFAULT_BAYESIAN_PARALLELISM) elif not enforce_sequential_optimization: # If no max parallelism settings specified and not enforcing sequential # optimization, do not limit parallelism. sobol_parallelism = bo_parallelism = None else: # No additional max parallelism settings, use defaults sobol_parallelism = None # No restriction on Sobol phase bo_parallelism = DEFAULT_BAYESIAN_PARALLELISM if not no_bayesian_optimization and suggested_model is not None: if not enforce_sequential_optimization and ( max_parallelism_override or max_parallelism_cap ): logger.info( "If `enforce_sequential_optimization` is False, max parallelism is " "not enforced and other max parallelism settings will be ignored." ) if max_parallelism_override and max_parallelism_cap: raise ValueError( "If `max_parallelism_override` specified, cannot also apply " "`max_parallelism_cap`." ) # If number of initialization trials is not specified, estimate it. logger.info( "Calculating the number of remaining initialization trials based on " f"num_initialization_trials={num_initialization_trials} " f"max_initialization_trials={max_initialization_trials} " f"num_tunable_parameters={len(search_space.tunable_parameters)} " f"num_trials={num_trials} " f"use_batch_trials={use_batch_trials}" ) if num_initialization_trials is None: num_initialization_trials = calculate_num_initialization_trials( num_tunable_parameters=len(search_space.tunable_parameters), num_trials=num_trials, use_batch_trials=use_batch_trials, ) if max_initialization_trials is not None: num_initialization_trials = min( num_initialization_trials, max_initialization_trials ) logger.info( f"calculated num_initialization_trials={num_initialization_trials}" ) num_remaining_initialization_trials = max( 0, num_initialization_trials - max(0, num_completed_initialization_trials) ) logger.info( "num_completed_initialization_trials=" f"{num_completed_initialization_trials} " f"num_remaining_initialization_trials={num_remaining_initialization_trials}" ) steps = [] # `verbose` and `disable_progbar` defaults and overrides model_is_saasbo = is_saasbo(suggested_model) if verbose is None and model_is_saasbo: verbose = True elif verbose is not None and not model_is_saasbo: logger.warning( f"Overriding `verbose = {verbose}` to `None` for non-SAASBO GP step." ) verbose = None if disable_progbar is not None and not model_is_saasbo: logger.warning( f"Overriding `disable_progbar = {disable_progbar}` to `None` for " "non-SAASBO GP step." ) disable_progbar = None if jit_compile is not None and not model_is_saasbo: logger.warning( f"Overriding `jit_compile = {jit_compile}` to `None` for " "non-SAASBO GP step." ) jit_compile = None # Create `generation_strategy`, adding first Sobol step # if `num_remaining_initialization_trials` is > 0. if num_remaining_initialization_trials > 0: steps.append( _make_sobol_step( num_trials=num_remaining_initialization_trials, enforce_num_trials=enforce_sequential_optimization, seed=random_seed, max_parallelism=sobol_parallelism, should_deduplicate=should_deduplicate, ) ) steps.append( _make_botorch_step( model=suggested_model, winsorization_config=winsorization_config, derelativize_with_raw_status_quo=derelativize_with_raw_status_quo, no_winsorization=no_winsorization, max_parallelism=bo_parallelism, model_kwargs={"torch_device": torch_device}, should_deduplicate=should_deduplicate, verbose=verbose, disable_progbar=disable_progbar, jit_compile=jit_compile, ), ) gs = GenerationStrategy(steps=steps) logger.info( f"Using Bayesian Optimization generation strategy: {gs}. Iterations after" f" {num_remaining_initialization_trials} will take longer to generate due" " to model-fitting." ) else: # `no_bayesian_optimization` is True or we could not suggest BO model if verbose is not None: logger.warning( f"Ignoring `verbose = {verbose}` for `generation_strategy` " "without a GP step." ) gs = GenerationStrategy( steps=[ _make_sobol_step( seed=random_seed, should_deduplicate=should_deduplicate, max_parallelism=sobol_parallelism, ) ] ) logger.info("Using Sobol generation strategy.") if experiment: gs.experiment = experiment return gs
def _get_winsorization_transform_config( winsorization_config: Optional[ Union[WinsorizationConfig, Dict[str, WinsorizationConfig]] ], derelativize_with_raw_status_quo: bool, no_winsorization: bool, ) -> Optional[TConfig]: if no_winsorization: if winsorization_config is not None: warnings.warn( "`no_winsorization = True` but `winsorization_config` has been set. " "Not winsorizing." ) return None if winsorization_config: return {"winsorization_config": winsorization_config} return {"derelativize_with_raw_status_quo": derelativize_with_raw_status_quo}
[docs]def is_saasbo(model: Models) -> bool: return model.name in ["SAASBO", "FULLYBAYESIAN", "FULLYBAYESIANMOO"]