Source code for smac.intensification.parallel_scheduling

import typing
import warnings

import numpy as np

from smac.intensification.abstract_racer import AbstractRacer, RunInfoIntent
from smac.optimizer.epm_configuration_chooser import EPMChooser
from smac.stats.stats import Stats
from smac.configspace import Configuration
from smac.runhistory.runhistory import RunHistory, RunInfo, RunValue
from import TrajLogger

[docs]class ParallelScheduler(AbstractRacer): """Common Racer class for Intensifiers that will schedule configurations on a parallel fashion. This class instantiates intensifier objects on a need basis, that is, to prevent workers from being idle. This intensifier objects will give configurations to run Parameters ---------- stats: smac.stats.stats.Stats stats object traj_logger: TrajLogger object to log all new incumbents rng : np.random.RandomState instances : typing.List[str] list of all instance ids instance_specifics : typing.Mapping[str,np.ndarray] mapping from instance name to instance specific string cutoff : typing.Optional[int] cutoff of TA runs deterministic : bool whether the TA is deterministic or not initial_budget : typing.Optional[float] minimum budget allowed for 1 run of successive halving max_budget : typing.Optional[float] maximum budget allowed for 1 run of successive halving eta : float 'halving' factor after each iteration in a successive halving run. Defaults to 3 num_initial_challengers : typing.Optional[int] number of challengers to consider for the initial budget. If None, calculated internally run_obj_time : bool whether the run objective is runtime or not (if true, apply adaptive capping) n_seeds : typing.Optional[int] Number of seeds to use, if TA is not deterministic. Defaults to None, i.e., seed is set as 0 instance_order : typing.Optional[str] how to order instances. Can be set to: [None, shuffle_once, shuffle] * None - use as is given by the user * shuffle_once - shuffle once and use across all SH run (default) * shuffle - shuffle before every SH run adaptive_capping_slackfactor : float slack factor of adpative capping (factor * adaptive cutoff) inst_seed_pairs : typing.List[typing.Tuple[str, int]], optional Do not set this argument, it will only be used by hyperband! min_chall: int minimal number of challengers to be considered (even if time_bound is exhausted earlier). This class will raise an exception if a value larger than 1 is passed. incumbent_selection: str How to select incumbent in successive halving. Only active for real-valued budgets. Can be set to: [highest_executed_budget, highest_budget, any_budget] * highest_executed_budget - incumbent is the best in the highest budget run so far (default) * highest_budget - incumbent is selected only based on the highest budget * any_budget - incumbent is the best on any budget i.e., best performance regardless of budget """ def __init__(self, stats: Stats, traj_logger: TrajLogger, rng: np.random.RandomState, instances: typing.List[str], instance_specifics: typing.Mapping[str, np.ndarray] = None, cutoff: typing.Optional[float] = None, deterministic: bool = False, initial_budget: typing.Optional[float] = None, max_budget: typing.Optional[float] = None, eta: float = 3, num_initial_challengers: typing.Optional[int] = None, run_obj_time: bool = True, n_seeds: typing.Optional[int] = None, instance_order: typing.Optional[str] = 'shuffle_once', adaptive_capping_slackfactor: float = 1.2, inst_seed_pairs: typing.Optional[typing.List[typing.Tuple[str, int]]] = None, min_chall: int = 1, incumbent_selection: str = 'highest_executed_budget', ) -> None: super().__init__(stats=stats, traj_logger=traj_logger, rng=rng, instances=instances, instance_specifics=instance_specifics, cutoff=cutoff, deterministic=deterministic, run_obj_time=run_obj_time, adaptive_capping_slackfactor=adaptive_capping_slackfactor, min_chall=min_chall) # We have a pool of instances that yield configurations ot run self.intensifier_instances = {} # type: typing.Dict[int, AbstractRacer]
[docs] def get_next_run(self, challengers: typing.Optional[typing.List[Configuration]], incumbent: Configuration, chooser: typing.Optional[EPMChooser], run_history: RunHistory, repeat_configs: bool = False, num_workers: int = 1, ) -> typing.Tuple[RunInfoIntent, RunInfo]: """ This procedure decides from which instance to pick a config, in order to determine the next run. To prevent having idle workers, this procedure creates new instances up to the maximum number of workers available. If no new intensifier instance can be created and all intensifier objects need to wait for more data, this procedure sends a wait request to smbo. Parameters ---------- challengers : typing.List[Configuration] promising configurations incumbent: Configuration incumbent configuration chooser : smac.optimizer.epm_configuration_chooser.EPMChooser optimizer that generates next configurations to use for racing run_history : smac.runhistory.runhistory.RunHistory stores all runs we ran so far repeat_configs : bool if False, an evaluated configuration will not be generated again num_workers: int the maximum number of workers available at a given time. Returns ------- intent: RunInfoIntent Indicator of how to consume the RunInfo object run_info: RunInfo An object that encapsulates the minimum information to evaluate a configuration """ if num_workers <= 1: warnings.warn("{} is intended to be used " "with more than 1 worker but num_workers={}".format( self.__class__.__name__, num_workers )) # If repeat_configs is True, that means that not only self can repeat # configurations, but also in the context of multiprocessing, N # intensifier instances will also share configurations. The later # is not supported if repeat_configs: raise ValueError( "repeat_configs==True is not supported for parallel execution") # First get a config to run from a SH instance for i in self._sort_instances_by_stage(self.intensifier_instances): intent, run_info = self.intensifier_instances[i].get_next_run( challengers=challengers, incumbent=incumbent, chooser=chooser, run_history=run_history, repeat_configs=repeat_configs, ) # if asked to wait, the intensifier cannot come up # with a new configuration, so we continue if intent == RunInfoIntent.WAIT: continue return intent, run_info # If gotten to this point, we might look into adding a new # intensifier if self._add_new_instance(num_workers): return self.intensifier_instances[len(self.intensifier_instances) - 1].get_next_run( challengers=challengers, incumbent=incumbent, chooser=chooser, run_history=run_history, repeat_configs=repeat_configs, ) # If got to this point, no new instance can be added as # there are no idle workers and all running instances have to # wait, so we return a wait intent return RunInfoIntent.WAIT, RunInfo( config=None, instance="0", instance_specific="0", seed=0, cutoff=None, capped=False, budget=0.0, )
[docs] def process_results(self, run_info: RunInfo, incumbent: typing.Optional[Configuration], run_history: RunHistory, time_bound: float, result: RunValue, log_traj: bool = True, ) -> \ typing.Tuple[Configuration, float]: """ The intensifier stage will be updated based on the results/status of a configuration execution. To do so, this procedures redirects the result argument, to the respective intensifier object that generated the original config. Also, an incumbent will be determined. This determination is done using the complete run history, so we rely on the current intensifier choice of incumbent. That is, no need to go over each instance to get the incumbent, as there is no local runhistory Parameters ---------- run_info : RunInfo A RunInfo containing the configuration that was evaluated incumbent : typing.Optional[Configuration] Best configuration seen so far run_history : RunHistory stores all runs we ran so far if False, an evaluated configuration will not be generated again time_bound : float time in [sec] available to perform intensify result: RunValue Contain the result (status and other methadata) of exercising a challenger/incumbent. log_traj: bool Whether to log changes of incumbents in trajectory Returns ------- incumbent: Configuration current (maybe new) incumbent configuration inc_perf: float empirical performance of incumbent configuration """ return self.intensifier_instances[run_info.source_id].process_results( run_info=run_info, incumbent=incumbent, run_history=run_history, time_bound=time_bound, result=result, log_traj=log_traj, )
[docs] def _add_new_instance(self, num_workers: int) -> bool: """ Decides if it is possible to add a new intensifier instance, and adds it. If a new intensifier instance is added, True is returned, else False. Parameters: ----------- num_workers: int the maximum number of workers available at a given time. Returns ------- Whether or not a successive halving instance was added """ raise NotImplementedError()
[docs] def _get_intensifier_ranking(self, intensifier: AbstractRacer ) -> typing.Tuple[int, int]: """ Given a intensifier, returns how advance it is. This metric will be used to determine what priority to assign to the intensifier Parameters ---------- intensifier: AbstractRacer Intensifier to rank based on run progress Returns ------- ranking: int the higher this number, the faster the intensifier will get the running resources. For hyperband we can use the sh_intensifier stage, for example tie_breaker: int The configurations that have been launched to break ties. For example, in the case of Successive Halving it can be the number of configurations launched """ raise NotImplementedError()
[docs] def _sort_instances_by_stage(self, instances: typing.Dict[int, AbstractRacer] ) -> typing.List[int]: """ This procedure dictates what SH to prioritize in launching jobs. It prioritizes resource allocation to SH instances that have higher stages. In case of tie, we prioritize the SH instance with more launched configs Parameters ---------- instances: typing.Dict[int, AbstractRacer] Dict with the instances to prioritize Returns ------- List: The order in which to query for new jobs """ # This function might be called when no intensifier instances # exist (first iteration), so we return an empty list in that case if len(instances) == 0: return [] # We want to prioritize runs that are close to finishing an iteration. # In the context of successive halving for example, an iteration has stages # that are composed of # of configs and each configs has # of instance-seed pairs # so ranking will be the stage (the higher the stage, the more we want this run # to be finished earlier). Also, in case of tie (runs at same stage) we need a # tie breaker, which can be the number of configs already launched preference = [] for i, sh in instances.items(): ranking, tie_breaker = self._get_intensifier_ranking(sh) preference.append( (i, ranking, tie_breaker), ) # First we sort by config/instance/seed as the less important criteria preference.sort(key=lambda x: x[2], reverse=True) # Second by stage. The more advanced the stage is, the more we want # this intensifier to finish early preference.sort(key=lambda x: x[1], reverse=True) return [i for i, s, c in preference]