Source code for ax.core.batch_trial

#!/usr/bin/env python3
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

from collections import OrderedDict, defaultdict
from datetime import datetime
from typing import (
    TYPE_CHECKING,
    DefaultDict,
    Dict,
    List,
    MutableMapping,
    NamedTuple,
    Optional,
    Union,
)

import numpy as np
from ax.core.arm import Arm
from ax.core.base_trial import BaseTrial
from ax.core.generator_run import GeneratorRun, GeneratorRunType
from ax.core.trial import immutable_once_run
from ax.core.types import TCandidateMetadata
from ax.utils.common.docutils import copy_doc
from ax.utils.common.equality import datetime_equals, equality_typechecker
from ax.utils.common.logger import get_logger
from ax.utils.common.typeutils import checked_cast, not_none


logger = get_logger(__name__)


if TYPE_CHECKING:
    # import as module to make sphinx-autodoc-typehints happy
    from ax import core  # noqa F401  # pragma: no cover


[docs]class AbandonedArm(NamedTuple): """Tuple storing metadata of arm that has been abandoned within a BatchTrial. """ name: str time: datetime reason: Optional[str] = None @equality_typechecker def __eq__(self, other: "AbandonedArm") -> bool: return ( self.name == other.name and self.reason == other.reason and datetime_equals(self.time, other.time) )
[docs]class GeneratorRunStruct(NamedTuple): """Stores GeneratorRun object as well as the weight with which it was added.""" generator_run: GeneratorRun weight: float
[docs]class BatchTrial(BaseTrial): """Batched trial that has multiple attached arms, meant to be *deployed and evaluated together*, and possibly arm weights, which are a measure of how much of the total resources allocated to evaluating a batch should go towards evaluating the specific arm. For instance, for field experiments the weights could describe the fraction of the total experiment population assigned to the different treatment arms. Interpretation of the weights is defined in Runner. NOTE: A `BatchTrial` is not just a trial with many arms; it is a trial, for which it is important that the arms are evaluated simultaneously, e.g. in an A/B test where the evaluation results are subject to nonstationarity. For cases where multiple arms are evaluated separately and independently of each other, use multiple `Trial` objects with a single arm each. Args: experiment: Experiment, to which this trial is attached generator_run: GeneratorRun, associated with this trial. This can a also be set later through `add_arm` or `add_generator_run`, but a trial's associated generator run is immutable once set. trial_type: Type of this trial, if used in MultiTypeExperiment. optimize_for_power: Whether to optimize the weights of arms in this trial such that the experiment's power to detect effects of certain size is as high as possible. Refer to documentation of `BatchTrial.set_status_quo_and_optimize_power` for more detail. ttl_seconds: If specified, trials will be considered failed after this many seconds since the time the trial was ran, unless the trial is completed before then. Meant to be used to detect 'dead' trials, for which the evaluation process might have crashed etc., and which should be considered failed after their 'time to live' has passed. index: If specified, the trial's index will be set accordingly. This should generally not be specified, as in the index will be automatically determined based on the number of existing trials. This is only used for the purpose of loading from storage. """ def __init__( self, experiment: "core.experiment.Experiment", generator_run: Optional[GeneratorRun] = None, trial_type: Optional[str] = None, optimize_for_power: Optional[bool] = False, ttl_seconds: Optional[int] = None, index: Optional[int] = None, ) -> None: super().__init__( experiment=experiment, trial_type=trial_type, ttl_seconds=ttl_seconds, index=index, ) self._generator_run_structs: List[GeneratorRunStruct] = [] self._abandoned_arms_metadata: Dict[str, AbandonedArm] = {} self._status_quo: Optional[Arm] = None self._status_quo_weight_override: Optional[float] = None if generator_run is not None: self.add_generator_run(generator_run=generator_run) self.optimize_for_power = optimize_for_power status_quo = experiment.status_quo if optimize_for_power: if status_quo is None: raise ValueError( "Can only optimize for power if experiment has a status quo." ) self.set_status_quo_and_optimize_power(status_quo=status_quo) else: # Set the status quo for tracking purposes # It will not be included in arm_weights self._status_quo = status_quo @property def experiment(self) -> "core.experiment.Experiment": """The experiment this batch belongs to.""" return self._experiment @property def index(self) -> int: """The index of this batch within the experiment's batch list.""" return self._index @property def generator_run_structs(self) -> List[GeneratorRunStruct]: """List of generator run structs attached to this trial. Struct holds generator_run object and the weight with which it was added. """ return self._generator_run_structs @property def arm_weights(self) -> MutableMapping[Arm, float]: """The set of arms and associated weights for the trial. These are constructed by merging the arms and weights from each generator run that is attached to the trial. """ arm_weights = OrderedDict() if len(self._generator_run_structs) == 0 and self.status_quo is None: return arm_weights for struct in self._generator_run_structs: multiplier = struct.weight for arm, weight in struct.generator_run.arm_weights.items(): scaled_weight = weight * multiplier if arm in arm_weights: arm_weights[arm] += scaled_weight else: arm_weights[arm] = scaled_weight if self.status_quo is not None and self._status_quo_weight_override is not None: # If override is specified, this is the weight the status quo gets, # regardless of whether it appeared in any generator runs. # If no override is specified, status quo does not appear in arm_weights. arm_weights[self.status_quo] = self._status_quo_weight_override return arm_weights @arm_weights.setter def arm_weights(self, arm_weights: MutableMapping[Arm, float]) -> None: raise NotImplementedError("Use `trial.add_arms_and_weights`") @immutable_once_run def add_arm(self, arm: Arm, weight: float = 1.0) -> "BatchTrial": """Add a arm to the trial. Args: arm: The arm to be added. weight: The weight with which this arm should be added. Returns: The trial instance. """ return self.add_arms_and_weights(arms=[arm], weights=[weight]) @immutable_once_run def add_arms_and_weights( self, arms: List[Arm], weights: Optional[List[float]] = None, multiplier: float = 1.0, ) -> "BatchTrial": """Add arms and weights to the trial. Args: arms: The arms to be added. weights: The weights associated with the arms. multiplier: The multiplier applied to input weights before merging with the current set of arms and weights. Returns: The trial instance. """ return self.add_generator_run( generator_run=GeneratorRun( arms=arms, weights=weights, type=GeneratorRunType.MANUAL.name ), multiplier=multiplier, ) @immutable_once_run def add_generator_run( self, generator_run: GeneratorRun, multiplier: float = 1.0 ) -> "BatchTrial": """Add a generator run to the trial. The arms and weights from the generator run will be merged with the existing arms and weights on the trial, and the generator run object will be linked to the trial for tracking. Args: generator_run: The generator run to be added. multiplier: The multiplier applied to input weights before merging with the current set of arms and weights. Returns: The trial instance. """ # Copy the generator run, to preserve initial and skip mutations to arms. generator_run = generator_run.clone() # First validate generator run arms for arm in generator_run.arms: self.experiment.search_space.check_types(arm.parameters, raise_error=True) # Add names to arms # For those not yet added to this experiment, create a new name # Else, use the name of the existing arm for arm in generator_run.arms: self._check_existing_and_name_arm(arm) self._generator_run_structs.append( GeneratorRunStruct(generator_run=generator_run, weight=multiplier) ) generator_run.index = len(self._generator_run_structs) - 1 if self.status_quo is not None and self.optimize_for_power: self.set_status_quo_and_optimize_power(status_quo=not_none(self.status_quo)) self._set_generation_step_index( generation_step_index=generator_run._generation_step_index ) return self @property def status_quo(self) -> Optional[Arm]: """The control arm for this batch.""" return self._status_quo @status_quo.setter def status_quo(self, status_quo: Optional[Arm]) -> None: raise NotImplementedError( "Use `set_status_quo_with_weight` or " "`set_status_quo_and_optimize_power` " "to set the status quo arm." )
[docs] def unset_status_quo(self) -> None: """Set the status quo to None.""" self._status_quo = None
@immutable_once_run def set_status_quo_with_weight( self, status_quo: Arm, weight: float ) -> "BatchTrial": """Sets status quo arm with given weight. This weight *overrides* any weight the status quo has from generator runs attached to this batch. Thus, this function is not the same as using add_arm, which will result in the weight being additive over all generator runs. """ # Assign a name to this arm if none exists if weight is not None and weight <= 0.0: raise ValueError("Status quo weight must be positive.") if status_quo is not None: self.experiment.search_space.check_types( status_quo.parameters, raise_error=True ) self.experiment._name_and_store_arm_if_not_exists( arm=status_quo, proposed_name="status_quo_" + str(self.index) ) self._status_quo = status_quo self._status_quo_weight_override = weight return self @immutable_once_run def set_status_quo_and_optimize_power(self, status_quo: Arm) -> "BatchTrial": """Adds a status quo arm to the batch and optimizes for power. NOTE: this optimization based on the arms that are currently attached to the batch. If you add more arms later, you should re-run this function. If you want the optimization to happen automatically, set batch.optimize_for_power = True. This function will maximize power across the multiple pair-wise comparisons of existing arms against the status_quo. Specifically, this function assigns sqrt(sum_weights) weight to the status quo, where sum_weights is the sum of the weights of the existing arms, excluding the status quo. This will be optimal in terms of statistical power in the case where: 1) status quo is the only arm to compare against, 2) all other arms are of equal interest. """ if len(self.arms) == 0: # If status quo is the only arm, just set its weight to 1 # Can't use logic below, because it will choose 0 self.set_status_quo_with_weight(status_quo=status_quo, weight=1) return self # arm_weights should always have at least one arm now arm_weights = not_none(self.arm_weights) sum_weights = sum(w for arm, w in arm_weights.items() if arm != status_quo) optimal_status_quo_weight_override = np.sqrt(sum_weights) self.set_status_quo_with_weight( status_quo=status_quo, weight=optimal_status_quo_weight_override ) return self @property def arms(self) -> List[Arm]: """All arms contained in the trial.""" arm_weights = self.arm_weights return [] if arm_weights is None else list(arm_weights.keys()) @property def weights(self) -> List[float]: """Weights corresponding to arms contained in the trial.""" arm_weights = self.arm_weights return [] if arm_weights is None else list(arm_weights.values()) @property def arms_by_name(self) -> Dict[str, Arm]: """Map from arm name to object for all arms in trial.""" arms_by_name = {} for arm in self.arms: if not arm.has_name: raise ValueError( # pragma: no cover "Arms attached to a trial must have a name." ) arms_by_name[arm.name] = arm return arms_by_name @property def abandoned_arms(self) -> List[Arm]: """List of arms that have been abandoned within this trial""" return [ self.arms_by_name[arm.name] for arm in self._abandoned_arms_metadata.values() ] @copy_doc(BaseTrial.generator_runs) @property def generator_runs(self) -> List[GeneratorRun]: return [grs.generator_run for grs in self.generator_run_structs] @property def abandoned_arms_metadata(self) -> List[AbandonedArm]: return list(self._abandoned_arms_metadata.values()) @property def is_factorial(self) -> bool: """Return true if the trial's arms are a factorial design with no linked factors. """ # To match the model behavior, this should probably actually be pulled # from exp.parameters. However, that seems rather ugly when this function # intuitively should just depend on the arms. sufficient_factors = all(len(arm.parameters or []) >= 2 for arm in self.arms) if not sufficient_factors: return False param_levels: DefaultDict[str, Dict[Union[str, float], int]] = ( defaultdict(dict) ) for arm in self.arms: for param_name, param_value in arm.parameters.items(): # Expected `Union[float, str]` for 2nd anonymous parameter to call # `dict.__setitem__` but got `Optional[Union[bool, float, str]]`. # pyre-fixme[6]: Expected `Union[float, str]` for 1st param but got `... param_levels[param_name][param_value] = 1 param_cardinality = 1 for param_values in param_levels.values(): param_cardinality *= len(param_values) return len(self.arms) == param_cardinality
[docs] def run(self) -> "BatchTrial": return checked_cast(BatchTrial, super().run())
[docs] def normalized_arm_weights( self, total: float = 1, trunc_digits: Optional[int] = None ) -> MutableMapping[Arm, float]: """Returns arms with a new set of weights normalized to the given total. This method is useful for many runners where we need to normalize weights to a certain total without mutating the weights attached to a trial. Args: total: The total weight to which to normalize. Default is 1, in which case arm weights can be interpreted as probabilities. trunc_digits: The number of digits to keep. If the resulting total weight is not equal to `total`, re-allocate weight in such a way to maintain relative weights as best as possible. Returns: Mapping from arms to the new set of weights. """ weights = np.array(self.weights) if trunc_digits is not None: atomic_weight = 10 ** -trunc_digits # pyre-fixme[16]: `float` has no attribute `astype`. int_weights = ( (total / atomic_weight) * (weights / np.sum(weights)) ).astype(int) n_leftover = int(total / atomic_weight) - np.sum(int_weights) int_weights[:n_leftover] += 1 weights = int_weights * atomic_weight else: weights = weights * (total / np.sum(weights)) return OrderedDict(zip(self.arms, weights))
[docs] def mark_arm_abandoned( self, arm_name: str, reason: Optional[str] = None ) -> "BatchTrial": """Mark a arm abandoned. Usually done after deployment when one arm causes issues but user wants to continue running other arms in the batch. Args: arm_name: The name of the arm to abandon. reason: The reason for abandoning the arm. Returns: The batch instance. """ if arm_name not in self.arms_by_name: raise ValueError("Arm must be contained in batch.") abandoned_arm = AbandonedArm(name=arm_name, time=datetime.now(), reason=reason) self._abandoned_arms_metadata[arm_name] = abandoned_arm return self
[docs] def clone(self) -> "BatchTrial": """Clone the trial. Returns: A new instance of the trial. """ new_trial = self._experiment.new_batch_trial() for struct in self._generator_run_structs: new_trial.add_generator_run(struct.generator_run, struct.weight) new_trial.trial_type = self._trial_type new_trial.runner = self._runner return new_trial
def __repr__(self) -> str: return ( "BatchTrial(" f"experiment_name='{self._experiment._name}', " f"index={self._index}, " f"status={self._status})" ) def _get_candidate_metadata_from_all_generator_runs( self, ) -> Dict[str, TCandidateMetadata]: """Retrieves combined candidate metadata from all generator runs on this batch trial in the form of { arm name -> candidate metadata} mapping. NOTE: this does not handle the case of the same arm appearing in multiple generator runs in the same trial: metadata from only one of the generator runs containing the arm will be retrieved. """ cand_metadata = {} for gr_struct in self._generator_run_structs: gr = gr_struct.generator_run if gr.candidate_metadata_by_arm_signature: gr_cand_metadata = gr.candidate_metadata_by_arm_signature warn = False for arm in gr.arms: if arm.name in cand_metadata: warn = True if gr_cand_metadata: # Reformat the mapping to be by arm name, since arm signature # is not stored in Ax data. cand_metadata[arm.name] = gr_cand_metadata.get(arm.signature) if warn: logger.warning( "The same arm appears in multiple generator runs in batch " f"{self.index}. Candidate metadata will only contain metadata " "for one of those generator runs, and the candidate metadata " "for the arm from another generator run will not be propagated." ) return cand_metadata def _get_candidate_metadata(self, arm_name: str) -> TCandidateMetadata: """Retrieves candidate metadata for a specific arm.""" try: arm = self.arms_by_name[arm_name] except KeyError: raise ValueError( f"Arm by name {arm_name} is not part of trial #{self.index}." ) for gr_struct in self._generator_run_structs: gr = gr_struct.generator_run if gr and gr.candidate_metadata_by_arm_signature and arm in gr.arms: return not_none(gr.candidate_metadata_by_arm_signature).get( arm.signature ) return None