Source code for ax.early_stopping.strategies.percentile

#!/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.

from logging import Logger
from typing import Any, Dict, Iterable, List, Optional, Set, Tuple

import numpy as np
import pandas as pd
from ax.core.experiment import Experiment
from ax.early_stopping.strategies.base import BaseEarlyStoppingStrategy
from ax.early_stopping.utils import align_partial_results
from ax.exceptions.core import UnsupportedError
from ax.utils.common.logger import get_logger
from ax.utils.common.typeutils import not_none

logger: Logger = get_logger(__name__)


[docs]class PercentileEarlyStoppingStrategy(BaseEarlyStoppingStrategy): """Implements the strategy of stopping a trial if its performance falls below that of other trials at the same step.""" def __init__( self, metric_names: Optional[Iterable[str]] = None, seconds_between_polls: int = 300, percentile_threshold: float = 50.0, min_progression: Optional[float] = 10, max_progression: Optional[float] = None, min_curves: Optional[int] = 5, trial_indices_to_ignore: Optional[List[int]] = None, true_objective_metric_name: Optional[str] = None, normalize_progressions: bool = False, ) -> None: """Construct a PercentileEarlyStoppingStrategy instance. Args: metric_names: A (length-one) list of name of the metric to observe. If None will default to the objective metric on the Experiment's OptimizationConfig. seconds_between_polls: How often to poll the early stopping metric to evaluate whether or not the trial should be early stopped. percentile_threshold: Falling below this threshold compared to other trials at the same step will stop the run. Must be between 0.0 and 100.0. e.g. if percentile_threshold=25.0, the bottom 25% of trials are stopped. Note that "bottom" here is determined based on performance, not absolute values; if `minimize` is False, then "bottom" actually refers to the top trials in terms of metric value. min_progression: Only stop trials if the latest progression value (e.g. timestamp, epochs, training data used) is greater than this threshold. Prevents stopping prematurely before enough data is gathered to make a decision. max_progression: Do not stop trials that have passed `max_progression`. Useful if we prefer finishing a trial that are already near completion. min_curves: Trials will not be stopped until a number of trials `min_curves` have completed with curve data attached. That is, if `min_curves` trials are completed but their curve data was not successfully retrieved, further trials may not be early-stopped. trial_indices_to_ignore: Trial indices that should not be early stopped. true_objective_metric_name: The actual objective to be optimized; used in situations where early stopping uses a proxy objective (such as training loss instead of eval loss) for stopping decisions. normalize_progressions: Normalizes the progression column of the MapData df by dividing by the max. If the values were originally in [0, `prog_max`] (as we would expect), the transformed values will be in [0, 1]. Useful for inferring the max progression and allows `min_progression` to be specified in the transformed space. IMPORTANT: Typically, `min_curves` should be > 0 to ensure that at least one trial has completed and that we have a reliable approximation for `prog_max`. """ super().__init__( metric_names=metric_names, seconds_between_polls=seconds_between_polls, trial_indices_to_ignore=trial_indices_to_ignore, min_progression=min_progression, max_progression=max_progression, min_curves=min_curves, true_objective_metric_name=true_objective_metric_name, normalize_progressions=normalize_progressions, ) self.percentile_threshold = percentile_threshold if metric_names is not None and len(list(metric_names)) > 1: raise UnsupportedError( "PercentileEarlyStoppingStrategy only supports a single metric. Use " "LogicalEarlyStoppingStrategy to compose early stopping strategies " "with multiple metrics." )
[docs] def should_stop_trials_early( self, trial_indices: Set[int], experiment: Experiment, **kwargs: Dict[str, Any], ) -> Dict[int, Optional[str]]: """Stop a trial if its performance is in the bottom `percentile_threshold` of the trials at the same step. Args: trial_indices: Indices of candidate trials to consider for early stopping. experiment: Experiment that contains the trials and other contextual data. Returns: A dictionary mapping trial indices that should be early stopped to (optional) messages with the associated reason. An empty dictionary means no suggested updates to any trial's status. """ metric_name, minimize = self._default_objective_and_direction( experiment=experiment ) data = self._check_validity_and_get_data( experiment=experiment, metric_names=[metric_name] ) if data is None: # don't stop any trials if we don't get data back return {} map_key = next(iter(data.map_keys)) df = data.map_df # default checks on `min_progression` and `min_curves`; if not met, don't do # early stopping at all and return {} if not self.is_eligible_any( trial_indices=trial_indices, experiment=experiment, df=df, map_key=map_key ): return {} try: metric_to_aligned_means, _ = align_partial_results( df=df, progr_key=map_key, metrics=[metric_name], ) except Exception as e: logger.warning( f"Encountered exception while aligning data: {e}. " "Not early stopping any trials." ) return {} aligned_means = metric_to_aligned_means[metric_name] decisions = { trial_index: self._should_stop_trial_early( trial_index=trial_index, experiment=experiment, df=aligned_means, df_raw=df, map_key=map_key, minimize=minimize, ) for trial_index in trial_indices } return { trial_index: reason for trial_index, (should_stop, reason) in decisions.items() if should_stop }
def _should_stop_trial_early( self, trial_index: int, experiment: Experiment, df: pd.DataFrame, df_raw: pd.DataFrame, map_key: str, minimize: bool, ) -> Tuple[bool, Optional[str]]: """Stop a trial if its performance is in the bottom `percentile_threshold` of the trials at the same step. Args: trial_index: Indices of candidate trial to stop early. experiment: Experiment that contains the trials and other contextual data. df: Dataframe of partial results after applying interpolation, filtered to objective metric. df_raw: The original MapData dataframe (before interpolation). map_key: Name of the column of the dataset that indicates progression. minimize: Whether objective value is being minimized. Returns: A tuple `(should_stop, reason)`, where `should_stop` is `True` iff the trial should be stopped, and `reason` is an (optional) string providing information on why the trial should or should not be stopped. """ logger.info(f"Considering trial {trial_index} for early stopping.") stopping_eligible, reason = self.is_eligible( trial_index=trial_index, experiment=experiment, df=df_raw, map_key=map_key ) if not stopping_eligible: return False, reason # dropna() here will exclude trials that have not made it to the # last progression of the trial under consideration, and therefore # can't be included in the comparison df_trial = not_none(df[trial_index].dropna()) trial_last_prog = df_trial.index.max() data_at_last_progression = df.loc[trial_last_prog].dropna() logger.info( "Early stopping objective at last progression is:\n" f"{data_at_last_progression}." ) # check for enough number of trials with data if ( self.min_curves is not None and len(data_at_last_progression) < self.min_curves # pyre-ignore[58] ): return self._log_and_return_num_trials_with_data( logger=logger, trial_index=trial_index, trial_last_progression=trial_last_prog, num_trials_with_data=len(data_at_last_progression), min_curves=self.min_curves, # pyre-ignore[6] ) # percentile early stopping logic percentile_threshold = ( 100.0 - self.percentile_threshold if minimize else self.percentile_threshold ) percentile_value = np.percentile(data_at_last_progression, percentile_threshold) trial_objective_value = data_at_last_progression[trial_index] should_early_stop = ( trial_objective_value > percentile_value if minimize else trial_objective_value < percentile_value ) comp = "worse" if should_early_stop else "better" reason = ( f"Trial objective value {trial_objective_value} is {comp} than " f"{percentile_threshold:.1f}-th percentile ({percentile_value}) " "across comparable trials." ) logger.info( f"Early stopping decision for {trial_index}: {should_early_stop}. " f"Reason: {reason}" ) return should_early_stop, reason