Source code for ax.core.generator_run

#!/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.

import copy
from collections import OrderedDict
from datetime import datetime
from enum import Enum
from typing import Any, Dict, List, MutableMapping, NamedTuple, Optional, Set, Tuple

import pandas as pd
from ax.core.arm import Arm
from ax.core.optimization_config import OptimizationConfig
from ax.core.search_space import SearchSpace
from ax.core.types import (
    TCandidateMetadata,
    TGenMetadata,
    TModelPredict,
    TModelPredictArm,
)
from ax.utils.common.equality import Base
from ax.utils.common.typeutils import not_none


[docs]class GeneratorRunType(Enum): """Class for enumerating generator run types.""" STATUS_QUO = 0 MANUAL = 1
[docs]class ArmWeight(NamedTuple): """NamedTuple for tying together arms and weights.""" arm: Arm weight: float
[docs]def extract_arm_predictions( model_predictions: TModelPredict, arm_idx: int ) -> TModelPredictArm: """Extract a particular arm from model_predictions. Args: model_predictions: Mean and Cov for all arms. arm_idx: Index of arm in prediction list. Returns: (mean, cov) for specified arm. """ means = model_predictions[0] covariances = model_predictions[1] means_per_arm = {metric: means[metric][arm_idx] for metric in means.keys()} covar_per_arm = { metric: { other_metric: covariances[metric][other_metric][arm_idx] for other_metric in covariances[metric].keys() } for metric in covariances.keys() } return (means_per_arm, covar_per_arm)
[docs]class GeneratorRun(Base): """An object that represents a single run of a generator. This object is created each time the ``gen`` method of a generator is called. It stores the arms and (optionally) weights that were generated by the run. When we add a generator run to a trial, its arms and weights will be merged with those from previous generator runs that were already attached to the trial. """ def __init__( self, arms: List[Arm], weights: Optional[List[float]] = None, optimization_config: Optional[OptimizationConfig] = None, search_space: Optional[SearchSpace] = None, model_predictions: Optional[TModelPredict] = None, best_arm_predictions: Optional[Tuple[Arm, Optional[TModelPredictArm]]] = None, type: Optional[str] = None, fit_time: Optional[float] = None, gen_time: Optional[float] = None, model_key: Optional[str] = None, model_kwargs: Optional[Dict[str, Any]] = None, bridge_kwargs: Optional[Dict[str, Any]] = None, gen_metadata: Optional[TGenMetadata] = None, model_state_after_gen: Optional[Dict[str, Any]] = None, generation_step_index: Optional[int] = None, candidate_metadata_by_arm_signature: Optional[ Dict[str, TCandidateMetadata] ] = None, ) -> None: """ Inits GeneratorRun. Args: arms: The list of arms generated by this run. weights: An optional list of weights to associate with the arms. optimization_config: The optimization config used during generation of this run. search_space: The search used during generation of this run. model_predictions: Means and covariances for the arms in this run recorded at the time the run was executed. best_arm_predictions: Optional tuple of best arm in this run (according to the optimization config) and its optional respective model predictions. type: Optional type of the run. fit_time: Optional number of seconds it took to fit the model that produced this generator run. For models with multiple invocations of gen, this is typically the fitting time since the last call to gen. gen_time: Optional number of seconds generation took. model_key: Optional name of the model that was used to produce this generator run. model_kwargs: Optional dictionary of keyword arguments to the model that was used to produce this generator run. bridge_kwargs: Optional dictionary of keyword arguments to the model bridge that was used to produce this generator run. gen_metadata: Optional dictionary of metadata generated by alongside the generator_run. model_state_after_gen: Optional dictionary of model state attributes to those attributes' values, to use when reinstantiating the model from the generator run. Note that this is the state of the model after generation, so these settings should only be applied to the model when reinstantiating it to continue generation from it, rather than to reproduce the conditions, in which this generator run was created. generation_step_index: Optional index of the generation step that produced this generator run. Applicable only if the genetator run was created via a generation strategy. candidate_metadata_by_arm_signature: Optional dictionary of arm signatures to model-produced candidate metadata that corresponds to that arm in this generator run. """ self._arm_weight_table: OrderedDict[str, ArmWeight] = OrderedDict() if weights is None: weights = [1.0 for i in range(len(arms))] if len(arms) != len(weights): raise ValueError("Weights and arms must have the same length.") if bridge_kwargs is not None or model_kwargs is not None: if model_key is None: raise ValueError( "Model key is required if model or bridge kwargs are provided." ) if bridge_kwargs is None or model_kwargs is None: raise ValueError( "Both model kwargs and bridge kwargs are required if either " "one is provided." ) for arm, weight in zip(arms, weights): existing_cw = self._arm_weight_table.get(arm.signature) if existing_cw: self._arm_weight_table[arm.signature] = ArmWeight( arm=arm, weight=existing_cw.weight + weight ) else: self._arm_weight_table[arm.signature] = ArmWeight( arm=arm, weight=weight ) self._generator_run_type: Optional[str] = type self._time_created: datetime = datetime.now() self._optimization_config = optimization_config self._search_space = search_space self._model_predictions = model_predictions self._best_arm_predictions = best_arm_predictions self._index: Optional[int] = None self._fit_time = fit_time self._gen_time = gen_time self._model_key = model_key self._model_kwargs = model_kwargs self._bridge_kwargs = bridge_kwargs self._gen_metadata = gen_metadata self._model_state_after_gen = model_state_after_gen # If candidate metadata is not None and not empty, check that all arm # signatures in it correspond to arms in this generator run. if candidate_metadata_by_arm_signature: unknown_arms_in_cand_metadata = ( set(candidate_metadata_by_arm_signature.keys()) - self.arm_signatures ) if unknown_arms_in_cand_metadata: raise ValueError( f"Arms with signatures {unknown_arms_in_cand_metadata} appear in " "candidate metadata, but not among the arms on this GeneratorRun." ) self._candidate_metadata_by_arm_signature = candidate_metadata_by_arm_signature # Validate that generation step index is non-negative. assert generation_step_index is None or generation_step_index >= 0 self._generation_step_index = generation_step_index @property def arms(self) -> List[Arm]: """Returns arms generated by this run.""" return [cw.arm for cw in self._arm_weight_table.values()] @property def arm_signatures(self) -> Set[str]: """Returns signatures of arms generated by this run.""" return {cw.arm.signature for cw in self._arm_weight_table.values()} @property def weights(self) -> List[float]: """Returns weights associated with arms generated by this run.""" return [cw.weight for cw in self._arm_weight_table.values()] @property def arm_weights(self) -> MutableMapping[Arm, float]: """Mapping from arms to weights (order matches order in `arms` property). """ return OrderedDict(zip(self.arms, self.weights)) @property def generator_run_type(self) -> Optional[str]: """The type of the generator run.""" return self._generator_run_type @property def time_created(self) -> datetime: """Creation time of the batch.""" return self._time_created @property def index(self) -> Optional[int]: """The index of this generator run within a trial's list of generator run structs. This field is set when the generator run is added to a trial. """ return self._index @index.setter def index(self, index: int) -> None: if self._index is not None and self._index != index: raise ValueError("Cannot change the index of a generator run once set.") self._index = index @property def optimization_config(self) -> Optional[OptimizationConfig]: """The optimization config used during generation of this run.""" return self._optimization_config @property def search_space(self) -> Optional[SearchSpace]: """The search used during generation of this run.""" return self._search_space @property def model_predictions(self) -> Optional[TModelPredict]: return self._model_predictions @property def fit_time(self) -> Optional[float]: return self._fit_time @property def gen_time(self) -> Optional[float]: return self._gen_time @property def model_predictions_by_arm(self) -> Optional[Dict[str, TModelPredictArm]]: if self._model_predictions is None: return None predictions: Dict[str, TModelPredictArm] = {} for idx, cond in enumerate(self.arms): predictions[cond.signature] = extract_arm_predictions( model_predictions=not_none(self._model_predictions), arm_idx=idx ) return predictions @property def best_arm_predictions(self) -> Optional[Tuple[Arm, Optional[TModelPredictArm]]]: return self._best_arm_predictions @property def gen_metadata(self) -> Optional[TGenMetadata]: """Returns metadata generated by this run.""" return self._gen_metadata @property def candidate_metadata_by_arm_signature( self, ) -> Optional[Dict[str, TCandidateMetadata]]: """Retrieves model-produced candidate metadata as a mapping from arm name (for the arm the candidate became when added to experiment) to the metadata dict. """ return self._candidate_metadata_by_arm_signature @property def param_df(self) -> pd.DataFrame: """ Constructs a Pandas dataframe with the parameter values for each arm. Useful for inspecting the contents of a generator run. Returns: pd.DataFrame: a dataframe with the generator run's arms. """ return pd.DataFrame.from_dict( {a.name_or_short_signature: a.parameters for a in self.arms}, orient="index" )
[docs] def clone(self) -> "GeneratorRun": """Return a deep copy of a GeneratorRun.""" return self._clone_with_arms(arms=self.arms, weights=self.weights)
[docs] def split_by_arm(self, populate_all_fields: bool = False) -> List["GeneratorRun"]: """Return a list of generator runs, each with all the metadata of generator run, but only with one of its arms. Useful when splitting a single generator run into multiple 1-arm trials. Args: populate_all_fields: By default, `split_by_arm` only sets some fields on the new, 'split' generator runs, in order to avoid creating multiple large objects and increasing the size of an experiment object. To force-populate all fields of the 'split' generator runs, set 'populate_all_fields' to True. """ return [ self._clone_with_arms( arms=[a], weights=[w], cherrypick_fields_to_clone=not populate_all_fields, ) for a, w in self.arm_weights.items() ]
def _clone_with_arms( self, arms: List[Arm], weights: Optional[List[float]], cherrypick_fields_to_clone: bool = False, ) -> "GeneratorRun": """Return a deep copy of a GeneratorRun with specified arms and weights. Args: arms: Arms to put into the clone generator run. weights: Weights to put into the clone generator run. cherrypick_fields_to_clone: If specified, only some properties of the generator run by the given names will be cloned. The 'cherrypicked' properties are: type, model key, and the generation step index. """ if cherrypick_fields_to_clone: return GeneratorRun( arms=[a.clone() for a in arms], weights=weights[:] if weights is not None else None, type=self.generator_run_type, model_key=self._model_key, generation_step_index=self._generation_step_index, ) # Cloning all fields of the generator run. cand_metadata = self.candidate_metadata_by_arm_signature generator_run = GeneratorRun( arms=[a.clone() for a in arms], weights=weights[:] if weights is not None else None, # pyre-fixme[16]: `Optional` has no attribute `clone`. optimization_config=self.optimization_config.clone() if self.optimization_config is not None else None, # pyre-fixme[16]: `Optional` has no attribute `clone`. search_space=self.search_space.clone() if self.search_space is not None else None, model_predictions=copy.deepcopy(self.model_predictions), best_arm_predictions=copy.deepcopy(self.best_arm_predictions), type=self.generator_run_type, fit_time=self.fit_time, gen_time=self.gen_time, model_key=self._model_key, model_kwargs=self._model_kwargs, bridge_kwargs=self._bridge_kwargs, gen_metadata=self._gen_metadata, model_state_after_gen=self._model_state_after_gen, generation_step_index=self._generation_step_index, candidate_metadata_by_arm_signature=cand_metadata, ) generator_run._time_created = self._time_created generator_run._index = self._index generator_run._model_key = self._model_key generator_run._model_kwargs = ( # pyre-fixme[16]: `Optional` has no attribute `copy`. self._model_kwargs.copy() if self._model_kwargs is not None else None ) generator_run._bridge_kwargs = ( self._bridge_kwargs.copy() if self._bridge_kwargs is not None else None ) generator_run._model_state_after_gen = ( self._model_state_after_gen.copy() if self._model_state_after_gen is not None else None ) return generator_run def __repr__(self) -> str: class_name = self.__class__.__name__ num_arms = len(self.arms) total_weight = sum(self.weights) return f"{class_name}({num_arms} arms, total weight {total_weight})"