#!/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 __future__ import annotations
from typing import TYPE_CHECKING, Any, Dict, Iterable, Optional, Type
from ax.core.data import Data
from ax.utils.common.base import SortableBase
from ax.utils.common.serialization import extract_init_args, serialize_init_args
from ax.utils.common.typeutils import checked_cast
if TYPE_CHECKING: # pragma: no cover
# import as module to make sphinx-autodoc-typehints happy
from ax import core # noqa F401
[docs]class Metric(SortableBase):
"""Base class for representing metrics.
The `fetch_trial_data` method is the essential method to override when
subclassing, which specifies how to retrieve a Metric, for a given trial.
A Metric must return a Data object, which requires (at minimum) the following:
https://ax.dev/api/_modules/ax/core/data.html#Data.required_columns
Attributes:
lower_is_better: Flag for metrics which should be minimized.
properties: Properties specific to a particular metric.
"""
data_constructor: Type[Data] = Data
def __init__(
self,
name: str,
lower_is_better: Optional[bool] = None,
properties: Optional[Dict[str, Any]] = None,
) -> None:
"""Inits Metric.
Args:
name: Name of metric.
lower_is_better: Flag for metrics which should be minimized.
properties: Dictionary of this metric's properties
"""
self._name = name
self.lower_is_better = lower_is_better
self.properties = properties or {}
[docs] @classmethod
def is_available_while_running(cls) -> bool:
"""Whether metrics of this class are available while the trial is running.
Metrics that are not available while the trial is running are assumed to be
available only upon trial completion. For such metrics, data is assumed to
never change once the trial is completed.
NOTE: If this method returns `False`, data-fetching via `experiment.fetch_data`
will return the data cached on the experiment (for the metrics of the given
class) whenever its available. Data is cached on experiment when attached
via `experiment.attach_data`.
"""
return False
[docs] @classmethod
def overwrite_existing_data(cls) -> bool:
"""Indicates whether, when attaching data, we should overwrite all previously
attached data with the new dataframe.
"""
return False
[docs] @classmethod
def combine_with_last_data(cls) -> bool:
"""Indicates whether, when attaching data, we should merge the new dataframe
into the most recently attached dataframe.
"""
return False
@property
def name(self) -> str:
"""Get name of metric."""
return self._name
@property
def fetch_multi_group_by_metric(self) -> Type[Metric]:
"""Metric class, with which to group this metric in `Experiment._metrics_by_class`,
which is used to combine metrics on experiment into groups and then fetch their
data via `Metric.fetch_trial_data_multi` for each group.
NOTE: By default, this property will just return the class on which it is
defined; however, in some cases it is useful to group metrics by their
superclass, in which case this property should return that superclass.
"""
return self.__class__
[docs] @classmethod
def serialize_init_args(cls, metric: Metric) -> Dict[str, Any]:
"""Serialize the properties needed to initialize the metric.
Used for storage.
"""
return serialize_init_args(object=metric)
[docs] @classmethod
def deserialize_init_args(cls, args: Dict[str, Any]) -> Dict[str, Any]:
"""Given a dictionary, extract the properties needed to initialize the metric.
Used for storage.
"""
return extract_init_args(args=args, class_=cls)
[docs] def fetch_trial_data(self, trial: core.base_trial.BaseTrial, **kwargs: Any) -> Data:
"""Fetch data for one trial."""
raise NotImplementedError(
f"Metric {self.name} does not implement data-fetching logic."
) # pragma: no cover
[docs] def fetch_experiment_data(
self, experiment: core.experiment.Experiment, **kwargs: Any
) -> Data:
"""Fetch this metric's data for an experiment.
Default behavior is to fetch data from all trials expecting data
and concatenate the results.
"""
return self.data_constructor.from_multiple_data(
[
checked_cast(
self.data_constructor, self.fetch_trial_data(trial, **kwargs)
)
if trial.status.expecting_data
else self.data_constructor()
for trial in experiment.trials.values()
],
)
[docs] @classmethod
def fetch_trial_data_multi(
cls, trial: core.base_trial.BaseTrial, metrics: Iterable[Metric], **kwargs: Any
) -> Data:
"""Fetch multiple metrics data for one trial.
Default behavior calls `fetch_trial_data` for each metric.
Subclasses should override this to trial data computation for multiple metrics.
"""
dat = cls.data_constructor.from_multiple_data(
[metric.fetch_trial_data(trial, **kwargs) for metric in metrics]
)
return dat
[docs] @classmethod
def fetch_experiment_data_multi(
cls,
experiment: core.experiment.Experiment,
metrics: Iterable[Metric],
trials: Optional[Iterable[core.base_trial.BaseTrial]] = None,
**kwargs: Any,
) -> Data:
"""Fetch multiple metrics data for an experiment.
Default behavior calls `fetch_trial_data_multi` for each trial.
Subclasses should override to batch data computation across trials + metrics.
"""
return cls.data_constructor.from_multiple_data(
[
cls.fetch_trial_data_multi(trial, metrics, **kwargs)
if trial.status.expecting_data
else cls.data_constructor()
for trial in (experiment.trials.values() if trials is None else trials)
]
)
[docs] @classmethod
def lookup_or_fetch_experiment_data_multi(
cls,
experiment: core.experiment.Experiment,
metrics: Iterable[Metric],
trials: Optional[Iterable[core.base_trial.BaseTrial]] = None,
**kwargs: Any,
) -> Data:
"""Fetch or lookup (with fallback to fetching) data for given metrics,
depending on whether they are available while running.
If metric is available while running, its data can change (and therefore
we should always re-fetch it). If metric is available only upon trial
completion, its data does not change, so we can look up that data on
the experiment and only fetch the data that is not already attached to
the experiment.
NOTE: If fetching data for a metrics class that is only available upon
trial completion, data fetched in this function (data that was not yet
available on experiment) will be attached to experiment.
"""
# If this metric is available while trial is running, just default to
# `fetch_experiment_data_multi`.
if cls.is_available_while_running():
cached_data = (
experiment.lookup_data(
trial_indices=[trial.index for trial in trials],
)
if trials
else None
)
fetched_data = cls.fetch_experiment_data_multi(
experiment=experiment, metrics=metrics, trials=trials, **kwargs
)
# If there is cached data, consider combining it with fetched data.
# That way, if this function is being called from within a loop over
# multiple metric classes (as is currently the case in lookup_data),
# we'll combine data from all metric classes into a single dataframe
# before attaching it to the experiment.
if cached_data:
cached_metric_names = cached_data.metric_names
# if there is a collision (ie fetched = A cached = AB), just use
# the recently fetched. That way, if we call `fetch_data` twice
# in a row (not within a for loop), we don't end up with
# duplicate data
if len(cached_metric_names.intersection({m.name for m in metrics})) > 0:
final_data = fetched_data
else:
final_data = cls.data_constructor.from_multiple_data(
[cached_data, fetched_data]
)
else:
final_data = fetched_data
if not fetched_data.df.empty:
experiment.attach_data(
final_data,
overwrite_existing_data=cls.overwrite_existing_data(),
combine_with_last_data=cls.combine_with_last_data(),
)
return fetched_data
# If this metric is available only upon trial completion, look up data
# on experiment and only fetch data that is not already cached.
if trials is None:
completed_trials = experiment.trials_by_status[
core.base_trial.TrialStatus.COMPLETED
]
else:
completed_trials = [t for t in trials if t.status.is_completed]
if not completed_trials:
return cls.data_constructor()
trials_data = []
for trial in completed_trials:
cached_trial_data = experiment.lookup_data_for_trial(
trial_index=trial.index,
)[0]
cached_metric_names = cached_trial_data.metric_names
metrics_to_fetch = [m for m in metrics if m.name not in cached_metric_names]
if not metrics_to_fetch:
# If all needed data fetched from cache, no need to fetch any other data
# for trial.
trials_data.append(cached_trial_data)
continue
try:
fetched_trial_data = cls.fetch_experiment_data_multi(
experiment=experiment,
metrics=metrics_to_fetch,
trials=[trial],
**kwargs,
)
except NotImplementedError:
# Metric does not implement fetching logic and only uses lookup.
fetched_trial_data = cls.data_constructor()
final_data = cls.data_constructor.from_multiple_data(
[cached_trial_data, fetched_trial_data]
)
if not final_data.df.empty:
experiment.attach_data(
final_data,
overwrite_existing_data=cls.overwrite_existing_data(),
combine_with_last_data=cls.combine_with_last_data(),
)
trials_data.append(final_data)
return cls.data_constructor.from_multiple_data(
trials_data, subset_metrics=[m.name for m in metrics]
)
[docs] def clone(self) -> "Metric":
"""Create a copy of this Metric."""
cls = type(self)
return cls(
**serialize_init_args(self),
)
def __repr__(self) -> str:
return "{class_name}('{metric_name}')".format(
class_name=self.__class__.__name__, metric_name=self.name
)
@property
def _unique_id(self) -> str:
return str(self)
