Source code for ax.telemetry.scheduler
# 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 __future__ import annotations
from dataclasses import asdict, dataclass
from typing import Any, Dict, Optional
from warnings import warn
import numpy as np
from ax.service.scheduler import get_fitted_model_bridge, Scheduler
from ax.telemetry.common import _get_max_transformed_dimensionality
from ax.telemetry.experiment import ExperimentCompletedRecord, ExperimentCreatedRecord
from ax.telemetry.generation_strategy import GenerationStrategyCreatedRecord
[docs]@dataclass(frozen=True)
class SchedulerCreatedRecord:
"""
Record of the Scheduler creation event. This can be used for telemetry in settings
where many Schedulers are being created either manually or programatically. In
order to facilitate easy serialization only include simple types: numbers, strings,
bools, and None.
"""
experiment_created_record: ExperimentCreatedRecord
generation_strategy_created_record: GenerationStrategyCreatedRecord
# SchedulerOptions info
scheduler_total_trials: Optional[int]
scheduler_max_pending_trials: int
arms_per_trial: int
early_stopping_strategy_cls: Optional[str]
global_stopping_strategy_cls: Optional[str]
# Dimensionality of transformed SearchSpace can often be much higher due to one-hot
# encoding of unordered ChoiceParameters
transformed_dimensionality: int
[docs] @classmethod
def from_scheduler(cls, scheduler: Scheduler) -> SchedulerCreatedRecord:
return cls(
experiment_created_record=ExperimentCreatedRecord.from_experiment(
experiment=scheduler.experiment
),
generation_strategy_created_record=(
GenerationStrategyCreatedRecord.from_generation_strategy(
generation_strategy=scheduler.generation_strategy
)
),
scheduler_total_trials=scheduler.options.total_trials,
scheduler_max_pending_trials=scheduler.options.max_pending_trials,
# If batch_size is None then we are using single-Arm trials
arms_per_trial=scheduler.options.batch_size or 1,
early_stopping_strategy_cls=(
None
if scheduler.options.early_stopping_strategy is None
else scheduler.options.early_stopping_strategy.__class__.__name__
),
global_stopping_strategy_cls=(
None
if scheduler.options.global_stopping_strategy is None
else scheduler.options.global_stopping_strategy.__class__.__name__
),
transformed_dimensionality=_get_max_transformed_dimensionality(
search_space=scheduler.experiment.search_space,
generation_strategy=scheduler.generation_strategy,
),
)
[docs] def flatten(self) -> Dict[str, Any]:
"""
Flatten into an appropriate format for logging to a tabular database.
"""
self_dict = asdict(self)
experiment_created_record_dict = self_dict.pop("experiment_created_record")
generation_strategy_created_record_dict = self_dict.pop(
"generation_strategy_created_record"
)
return {
**self_dict,
**experiment_created_record_dict,
**generation_strategy_created_record_dict,
}
[docs]@dataclass(frozen=True)
class SchedulerCompletedRecord:
"""
Record of the Scheduler completion event. This will have information only available
after the optimization has completed.
"""
experiment_completed_record: ExperimentCompletedRecord
best_point_quality: float
model_fit_quality: float
model_std_quality: float
model_fit_generalization: float
model_std_generalization: float
num_metric_fetch_e_encountered: int
num_trials_bad_due_to_err: int
[docs] @classmethod
def from_scheduler(cls, scheduler: Scheduler) -> SchedulerCompletedRecord:
try:
model_bridge = get_fitted_model_bridge(scheduler)
model_fit_dict = model_bridge.compute_model_fit_metrics(
experiment=scheduler.experiment
)
# We'd ideally log the entire `model_fit_dict` as a single model fit metric
# can't capture the nuances of multiple experimental metrics, but this might
# lead to database performance issues. So instead, we take the worst
# coefficient of determination as model fit quality and store the full data
# in Manifold (TODO).
model_fit_quality = min(
model_fit_dict["coefficient_of_determination"].values()
)
# similar for uncertainty quantification, but distance from 1 matters
std = list(model_fit_dict["std_of_the_standardized_error"].values())
model_std_quality = _model_std_quality(np.array(std))
except Exception as e:
warn("Encountered exception in computing model fit quality: " + str(e))
model_fit_quality = float("-inf")
model_std_quality = float("-inf")
return cls(
experiment_completed_record=ExperimentCompletedRecord.from_experiment(
experiment=scheduler.experiment
),
best_point_quality=float("-inf"), # TODO[T147907632]
model_fit_quality=model_fit_quality,
model_std_quality=model_std_quality,
model_fit_generalization=float("-inf"), # TODO by cross_validate_by_trial
model_std_generalization=float("-inf"),
num_metric_fetch_e_encountered=scheduler._num_metric_fetch_e_encountered,
num_trials_bad_due_to_err=scheduler._num_trials_bad_due_to_err,
)
[docs] def flatten(self) -> Dict[str, Any]:
"""
Flatten into an appropriate format for logging to a tabular database.
"""
self_dict = asdict(self)
experiment_completed_record_dict = self_dict.pop("experiment_completed_record")
return {
**self_dict,
**experiment_completed_record_dict,
}
def _model_std_quality(std: np.ndarray) -> float:
"""Quantifies quality of the model uncertainty. A value of one means the
uncertainty is perfectly predictive of the true standard deviation of the error.
Values larger than one indicate over-estimation and negative values indicate
under-estimation of the true standard deviation of the error. In particular, a value
of 2 (resp. 1 / 2) represents an over-estimation (resp. under-estimation) of the
true standard deviation of the error by a factor of 2.
Args:
std: The standard deviation of the standardized error.
Returns:
The factor corresponding to the worst over- or under-estimation factor of the
standard deviation of the error among all experimentally observed metrics.
"""
max_std, min_std = np.max(std), np.min(std)
# comparing worst over-estimation factor with worst under-estimation factor
inv_model_std_quality = max_std if max_std > 1 / min_std else min_std
# reciprocal so that values greater than one indicate over-estimation and
# values smaller than indicate underestimation of the uncertainty.
return 1 / inv_model_std_quality
