Source code for smac.intensifier.successive_halving

from __future__ import annotations

from typing import Any, Iterator

import math
from collections import defaultdict

import numpy as np
from ConfigSpace import Configuration

from smac.constants import MAXINT
from smac.intensifier.abstract_intensifier import AbstractIntensifier
from smac.runhistory import TrialInfo
from smac.runhistory.dataclasses import InstanceSeedBudgetKey
from smac.runhistory.errors import NotEvaluatedError
from smac.scenario import Scenario
from smac.utils.configspace import get_config_hash
from smac.utils.data_structures import batch
from smac.utils.logging import get_logger
from smac.utils.pareto_front import calculate_pareto_front, sort_by_crowding_distance

__copyright__ = "Copyright 2022, automl.org"
__license__ = "3-clause BSD"

logger = get_logger(__name__)


[docs] class SuccessiveHalving(AbstractIntensifier): """ Implementation of Succesive Halving supporting multi-fidelity, multi-objective, and multi-processing. Internally, a tracker keeps track of configurations and their bracket and stage. The behaviour of this intensifier is as follows: - First, adds configurations from the runhistory to the tracker. The first stage is always filled-up. For example, the user provided 4 configs with the tell-method but the first stage requires 8 configs: 4 new configs are sampled and added together with the provided configs as a group to the tracker. - While loop: - If a trial in the tracker has not been yielded yet, yield it. - If we are running out of trials, we simply add a new batch of configurations to the first stage. Note ---- The implementation natively supports brackets from Hyperband. However, in the case of Successive Halving, only one bracket is used. Parameters ---------- eta : int, defaults to 3 Input that controls the proportion of configurations discarded in each round of Successive Halving. n_seeds : int, defaults to 1 How many seeds to use for each instance. instance_seed_order : str, defaults to "shuffle_once" How to order the instance-seed pairs. Can be set to: - `None`: No shuffling at all and use the instance-seed order provided by the user. - `shuffle_once`: Shuffle the instance-seed keys once and use the same order across all runs. - `shuffle`: Shuffles the instance-seed keys for each bracket individually. incumbent_selection : str, defaults to "highest_observed_budget" How to select the incumbent when using budgets. Can be set to: - `any_budget`: Incumbent is the best on any budget i.e., best performance regardless of budget. - `highest_observed_budget`: Incumbent is the best in the highest budget run so far. - `highest_budget`: Incumbent is selected only based on the highest budget. max_incumbents : int, defaults to 10 How many incumbents to keep track of in the case of multi-objective. seed : int, defaults to None Internal seed used for random events like shuffle seeds. """ def __init__( self, scenario: Scenario, eta: int = 3, n_seeds: int = 1, instance_seed_order: str | None = "shuffle_once", max_incumbents: int = 10, incumbent_selection: str = "highest_observed_budget", seed: int | None = None, ): super().__init__( scenario=scenario, n_seeds=n_seeds, max_incumbents=max_incumbents, seed=seed, ) self._eta = eta self._instance_seed_order = instance_seed_order self._incumbent_selection = incumbent_selection self._highest_observed_budget_only = False if incumbent_selection == "any_budget" else True # Global variables derived from scenario self._min_budget = self._scenario.min_budget self._max_budget = self._scenario.max_budget @property def meta(self) -> dict[str, Any]: # noqa: D102 meta = super().meta meta.update( { "eta": self._eta, "instance_seed_order": self._instance_seed_order, "incumbent_selection": self._incumbent_selection, } ) return meta
[docs] def reset(self) -> None: """Reset the internal variables of the intensifier including the tracker.""" super().reset() # States # dict[tuple[bracket, stage], list[tuple[seed to shuffle instance-seed keys, list[config_id]]] self._tracker: dict[tuple[int, int], list[tuple[int | None, list[Configuration]]]] = defaultdict(list)
[docs] def __post_init__(self) -> None: """Post initialization steps after the runhistory has been set.""" super().__post_init__() # We generate our instance seed pairs once is_keys = self.get_instance_seed_keys_of_interest() # Budgets, followed by lots of sanity-checking eta = self._eta min_budget = self._min_budget max_budget = self._max_budget if max_budget is not None and min_budget is not None and max_budget < min_budget: raise ValueError("Max budget has to be larger than min budget.") if self.uses_instances: if isinstance(min_budget, float) or isinstance(max_budget, float): raise ValueError("Successive Halving requires integer budgets when using instances.") min_budget = min_budget if min_budget is not None else 1 max_budget = max_budget if max_budget is not None else len(is_keys) if max_budget > len(is_keys): raise ValueError( f"Max budget of {max_budget} can not be greater than the number of instance-seed " f"keys ({len(is_keys)})." ) if max_budget < len(is_keys): logger.warning( f"Max budget {max_budget} does not include all instance seed " f"pairs ({len(is_keys)})." ) else: if min_budget is None or max_budget is None: raise ValueError( "Successive Halving requires the parameters min_budget and max_budget defined in the scenario." ) if len(is_keys) != 1: raise ValueError("Successive Halving supports only one seed when using budgets.") if min_budget is None or min_budget <= 0: raise ValueError("Min budget has to be larger than 0.") budget_type = "INSTANCES" if self.uses_instances else "BUDGETS" logger.info( f"Successive Halving uses budget type {budget_type} with eta {eta}, " f"min budget {min_budget}, and max budget {max_budget}." ) # Pre-computing Successive Halving variables max_iter = self._get_max_iterations(eta, max_budget, min_budget) budgets, n_configs = self._compute_configs_and_budgets_for_stages(eta, max_budget, max_iter) # Global variables self._min_budget = min_budget self._max_budget = max_budget # Stage variables, depending on the bracket (0 is the bracket here since SH only has one bracket) self._max_iterations: dict[int, int] = {0: max_iter + 1} self._n_configs_in_stage: dict[int, list] = {0: n_configs} self._budgets_in_stage: dict[int, list] = {0: budgets}
@staticmethod def _get_max_iterations(eta: int, max_budget: float | int, min_budget: float | int) -> int: return int(np.floor(np.log(max_budget / min_budget) / np.log(eta))) @staticmethod def _compute_configs_and_budgets_for_stages( eta: int, max_budget: float | int, max_iter: int, s_max: int | None = None ) -> tuple[list[int], list[int]]: if s_max is None: s_max = max_iter n_initial_challengers = math.ceil((eta**max_iter) * (s_max + 1) / (max_iter + 1)) # How many configs in each stage lin_space = -np.linspace(0, max_iter, max_iter + 1) n_configs_ = np.floor(n_initial_challengers * np.power(eta, lin_space)) n_configs = np.array(np.round(n_configs_), dtype=int).tolist() # How many budgets in each stage lin_space = -np.linspace(max_iter, 0, max_iter + 1) budgets = (max_budget * np.power(eta, lin_space)).tolist() return budgets, n_configs
[docs] def get_state(self) -> dict[str, Any]: # noqa: D102 # Replace config by dict tracker: dict[str, list[tuple[int | None, list[dict]]]] = defaultdict(list) for key in list(self._tracker.keys()): for seed, configs in self._tracker[key]: # We have to make key serializable new_key = f"{key[0]},{key[1]}" tracker[new_key].append((seed, [config.get_dictionary() for config in configs])) return {"tracker": tracker}
[docs] def set_state(self, state: dict[str, Any]) -> None: # noqa: D102 self._tracker = defaultdict(list) tracker = state["tracker"] for old_key in list(tracker.keys()): keys = [k for k in old_key.split(",")] new_key = (int(keys[0]), int(keys[1])) for seed, config_dicts in tracker[old_key]: seed = None if seed is None else int(seed) self._tracker[new_key].append( ( seed, [Configuration(self._scenario.configspace, config_dict) for config_dict in config_dicts], ) )
@property def uses_seeds(self) -> bool: # noqa: D102 return True @property def uses_budgets(self) -> bool: # noqa: D102 if self._scenario.instances is None: return True return False @property def uses_instances(self) -> bool: # noqa: D102 if self._scenario.instances is None: return False return True
[docs] def print_tracker(self) -> None: """Prints the number of configurations in each bracket/stage.""" messages = [] for (bracket, stage), others in self._tracker.items(): counter = 0 for _, config_ids in others: counter += len(config_ids) if counter > 0: messages.append(f"--- Bracket {bracket} / Stage {stage}: {counter} configs") if len(messages) > 0: logger.debug(f"{self.__class__.__name__} statistics:") for message in messages: logger.debug(message)
[docs] def get_trials_of_interest( self, config: Configuration, *, validate: bool = False, seed: int | None = None, ) -> list[TrialInfo]: # noqa: D102 is_keys = self.get_instance_seed_keys_of_interest(validate=validate, seed=seed) budget = None # When we use budgets, we always evaluated on the highest budget only if self.uses_budgets: budget = self._max_budget trials = [] for key in is_keys: trials.append(TrialInfo(config=config, instance=key.instance, seed=key.seed, budget=budget)) return trials
[docs] def get_instance_seed_budget_keys( self, config: Configuration, compare: bool = False ) -> list[InstanceSeedBudgetKey]: """Returns the instance-seed-budget keys for a given configuration. This method supports ``highest_budget``, which only returns the instance-seed-budget keys for the highest budget (if specified). In this case, the incumbents in ``update_incumbents`` are only changed if the costs on the highest budget are lower. Parameters ---------- config: Configuration The Configuration to be queried compare : bool, defaults to False Get rid of the budget information for comparing if the configuration was evaluated on the same instance-seed keys. """ isb_keys = self.runhistory.get_instance_seed_budget_keys( config, highest_observed_budget_only=self._highest_observed_budget_only ) # If incumbent should only be changed on the highest budget, we have to kick out all budgets below the highest if self.uses_budgets and self._incumbent_selection == "highest_budget": isb_keys = [key for key in isb_keys if key.budget == self._max_budget] if compare: # Get rid of duplicates isb_keys = list( set([InstanceSeedBudgetKey(instance=key.instance, seed=key.seed, budget=None) for key in isb_keys]) ) return isb_keys
def __iter__(self) -> Iterator[TrialInfo]: # noqa: D102 self.__post_init__() # Log brackets/stages logger.info("Number of configs in stage:") for bracket, n in self._n_configs_in_stage.items(): logger.info(f"--- Bracket {bracket}: {n}") logger.info("Budgets in stage:") for bracket, budgets in self._budgets_in_stage.items(): logger.info(f"--- Bracket {bracket}: {budgets}") rh = self.runhistory # We have to add already existing trials from the runhistory # Idea: We simply add existing configs to the tracker (first stage) but assign a random instance shuffle seed. # In the best case, trials (added from the users) are included in the seed and it has not re-computed again. # Note: If the intensifier was restored, we don't want to go in here if len(self._tracker) == 0: bracket = 0 stage = 0 # Print ignored budgets ignored_budgets = [] for k in rh.keys(): if k.budget not in self._budgets_in_stage[0] and k.budget not in ignored_budgets: ignored_budgets.append(k.budget) if len(ignored_budgets) > 0: logger.warning( f"Trials with budgets {ignored_budgets} will been ignored. Consider adding trials with budgets " f"{self._budgets_in_stage[0]}." ) # We batch the configs because we need n_configs in each iteration # If we don't have n_configs, we sample new ones # We take the number of configs from the first bracket and the first stage n_configs = self._n_configs_in_stage[bracket][stage] for configs in batch(rh.get_configs(), n_configs): n_rh_configs = len(configs) if len(configs) < n_configs: try: config = next(self.config_generator) configs.append(config) except StopIteration: # We stop if we don't find any configuration anymore return seed = self._get_next_order_seed() self._tracker[(bracket, stage)].append((seed, configs)) logger.info( f"Added {n_rh_configs} configs from runhistory and {n_configs - n_rh_configs} new configs to " f"Successive Halving's first bracket and first stage with order seed {seed}." ) while True: # If we don't yield trials anymore, we have to update # Otherwise, we can just keep yielding trials from the tracker update = False # We iterate over the tracker to do two things: # 1) Yield trials of configs that are not yet evaluated/running # 2) Update tracker and move better configs to the next stage # We start in reverse order to complete higher stages first logger.debug("Updating tracker:") # TODO: Process stages ascending or descending? for bracket, stage in list(self._tracker.keys()): pairs = self._tracker[(bracket, stage)].copy() for seed, configs in pairs: isb_keys = self._get_instance_seed_budget_keys_by_stage(bracket=bracket, stage=stage, seed=seed) # We iterate over the configs and yield trials which are not running/evaluated yet for config in configs: config_hash = get_config_hash(config) trials = self._get_next_trials(config, from_keys=isb_keys) logger.debug( f"--- Yielding {len(trials)}/{len(isb_keys)} for config {config_hash} in " f"stage {stage} with seed {seed}..." ) for trial in trials: yield trial update = True # If all configs were evaluated on ``n_configs_required``, we finally can compare try: successful_configs = self._get_best_configs(configs, bracket, stage, isb_keys) except NotEvaluatedError: # We can't compare anything, so we just continue with the next pairs logger.debug("--- Could not compare configs because not all trials have been evaluated yet.") continue # Update tracker # Remove current shuffle index / config pair self._tracker[(bracket, stage)].remove((seed, configs)) # Add successful to the next stage if stage < self._max_iterations[bracket] - 1: config_ids = [rh.get_config_id(config) for config in successful_configs] self._tracker[(bracket, stage + 1)].append((seed, successful_configs)) logger.debug( f"--- Promoted {len(config_ids)} configs from stage {stage} to stage {stage + 1} in " f"bracket {bracket}." ) else: logger.debug( f"--- Removed {len(successful_configs)} configs to last stage in bracket {bracket}." ) # Log how many configs are in each stage self.print_tracker() # Since we yielded something before, we want to go back as long as we do not find any trials anymore if update: continue # TODO: Aggressive progressing without knowing how well trials performed # Idea: Don't add constantly new batches (see ASHA) # If we are running out of trials, we want to add configs to the first stage # We simply add as many configs to the stage as required (_n_configs_in_stage[0]) configs = [] next_bracket = self._get_next_bracket() for _ in range(self._n_configs_in_stage[next_bracket][0]): try: config = next(self.config_generator) configs.append(config) except StopIteration: # We stop if we don't find any configuration anymore return # We keep track of the seed so we always evaluate on the same instances next_seed = self._get_next_order_seed() self._tracker[(next_bracket, 0)].append((next_seed, configs)) logger.debug( f"Added {len(configs)} new configs to bracket {next_bracket} stage 0 with shuffle seed {next_seed}." ) def _get_instance_seed_budget_keys_by_stage( self, bracket: int, stage: int, seed: int | None = None, ) -> list[InstanceSeedBudgetKey]: """Returns all instance-seed-budget keys (isb keys) for the given stage. Each stage is associated with a budget (N). Two possible options: 1) Instance based: We return N isb keys. If a seed is specified, we shuffle the keys before returning the first N instances. The budget is set to None here. 2) Budget based: We return one isb only but the budget is set to N. """ budget: float | int | None = None is_keys = self.get_instance_seed_keys_of_interest() # We have to differentiate between budgets and instances based here # If we are budget based, we always have one instance seed pair only # If we are in the instance setting, we have to return a specific number of instance seed pairs if self.uses_instances: # Shuffle instance seed pairs group-based if seed is not None: is_keys = self._reorder_instance_seed_keys(is_keys, seed=seed) # We only return the first N instances N = int(self._budgets_in_stage[bracket][stage]) is_keys = is_keys[:N] else: assert len(is_keys) == 1 # The stage defines which budget should be used (in real-valued setting) # No shuffle is needed here because we only have on instance seed pair budget = self._budgets_in_stage[bracket][stage] isbk = [] for isk in is_keys: isbk.append(InstanceSeedBudgetKey(instance=isk.instance, seed=isk.seed, budget=budget)) return isbk def _get_next_trials( self, config: Configuration, from_keys: list[InstanceSeedBudgetKey], ) -> list[TrialInfo]: """Returns trials for a given config from a list of instances (instance-seed-budget keys). The returned trials have not run or evaluated yet. """ rh = self.runhistory evaluated_trials = rh.get_trials(config, highest_observed_budget_only=False) running_trials = rh.get_running_trials(config) next_trials: list[TrialInfo] = [] for instance in from_keys: trial = TrialInfo(config=config, instance=instance.instance, seed=instance.seed, budget=instance.budget) if trial in evaluated_trials or trial in running_trials: continue next_trials.append(trial) return next_trials def _get_best_configs( self, configs: list[Configuration], bracket: int, stage: int, from_keys: list[InstanceSeedBudgetKey], ) -> list[Configuration]: """Returns the best configurations. The number of configurations is depending on the stage. Raises ``NotEvaluatedError`` if not all trials have been evaluated. """ try: n_configs = self._n_configs_in_stage[bracket][stage + 1] except IndexError: return [] rh = self.runhistory configs = configs.copy() for config in configs: isb_keys = rh.get_instance_seed_budget_keys(config) if not all(isb_key in isb_keys for isb_key in from_keys): raise NotEvaluatedError selected_configs: list[Configuration] = [] while len(selected_configs) < n_configs: # We calculate the pareto front for the given configs # We use the same isb keys for all the configs all_keys = [from_keys for _ in configs] incumbents = calculate_pareto_front(rh, configs, all_keys) # Idea: We recursively calculate the pareto front in every iteration for incumbent in incumbents: configs.remove(incumbent) selected_configs.append(incumbent) # If we have more selected configs, we remove the ones with the smallest crowding distance if len(selected_configs) > n_configs: all_keys = [from_keys for _ in selected_configs] selected_configs = sort_by_crowding_distance(rh, selected_configs, all_keys)[:n_configs] logger.debug("Found more configs than required. Removed configs with smallest crowding distance.") return selected_configs def _get_next_order_seed(self) -> int | None: """Next instances shuffle seed to use.""" # Here we have the option to shuffle the trials when specified by the user if self._instance_seed_order == "shuffle": seed = self._rng.randint(0, MAXINT) elif self._instance_seed_order == "shuffle_once": seed = 0 else: seed = None return seed def _get_next_bracket(self) -> int: """Successive Halving only uses one bracket. Therefore, we always return 0 here.""" return 0