#!/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 __future__ import annotations
import logging
import re
import warnings
from collections import defaultdict, OrderedDict
from datetime import datetime
from enum import Enum
from functools import partial, reduce
from typing import Any, Dict, Iterable, List, Mapping, Optional, Set, Tuple, Type
import ax.core.observation as observation
import pandas as pd
from ax.core.arm import Arm
from ax.core.base_trial import BaseTrial, DEFAULT_STATUSES_TO_WARM_START, TrialStatus
from ax.core.batch_trial import BatchTrial, LifecycleStage
from ax.core.data import Data
from ax.core.generator_run import GeneratorRun
from ax.core.map_data import MapData
from ax.core.map_metric import MapMetric
from ax.core.metric import Metric, MetricFetchE, MetricFetchResult
from ax.core.optimization_config import OptimizationConfig
from ax.core.parameter import Parameter
from ax.core.runner import Runner
from ax.core.search_space import HierarchicalSearchSpace, SearchSpace
from ax.core.trial import Trial
from ax.core.types import TParameterization
from ax.exceptions.core import UnsupportedError, UserInputError
from ax.utils.common.base import Base
from ax.utils.common.constants import EXPERIMENT_IS_TEST_WARNING, Keys
from ax.utils.common.docutils import copy_doc
from ax.utils.common.logger import _round_floats_for_logging, get_logger
from ax.utils.common.result import Err, Ok
from ax.utils.common.timeutils import current_timestamp_in_millis
from ax.utils.common.typeutils import checked_cast, not_none
logger: logging.Logger = get_logger(__name__)
[docs]class DataType(Enum):
DATA = 1
MAP_DATA = 3
# pyre-fixme[24]: Generic type `type` expects 1 type parameter, use `typing.Type` to
# avoid runtime subscripting errors.
DATA_TYPE_LOOKUP: Dict[DataType, Type] = {
DataType.DATA: Data,
DataType.MAP_DATA: MapData,
}
DEFAULT_OBJECTIVE_NAME = "objective"
ROUND_FLOATS_IN_LOGS_TO_DECIMAL_PLACES: int = 6
# pyre-fixme[5]: Global expression must be annotated.
round_floats_for_logging = partial(
_round_floats_for_logging,
decimal_places=ROUND_FLOATS_IN_LOGS_TO_DECIMAL_PLACES,
)
# pyre-fixme[13]: Attribute `_search_space` is never initialized.
[docs]class Experiment(Base):
"""Base class for defining an experiment."""
def __init__(
self,
search_space: SearchSpace,
name: Optional[str] = None,
optimization_config: Optional[OptimizationConfig] = None,
tracking_metrics: Optional[List[Metric]] = None,
runner: Optional[Runner] = None,
status_quo: Optional[Arm] = None,
description: Optional[str] = None,
is_test: bool = False,
experiment_type: Optional[str] = None,
properties: Optional[Dict[str, Any]] = None,
default_data_type: Optional[DataType] = None,
) -> None:
"""Inits Experiment.
Args:
search_space: Search space of the experiment.
name: Name of the experiment.
optimization_config: Optimization config of the experiment.
tracking_metrics: Additional tracking metrics not used for optimization.
runner: Default runner used for trials on this experiment.
status_quo: Arm representing existing "control" arm.
description: Description of the experiment.
is_test: Convenience metadata tracker for the user to mark test experiments.
experiment_type: The class of experiments this one belongs to.
properties: Dictionary of this experiment's properties.
default_data_type: Enum representing the data type this experiment uses.
"""
# appease pyre
self._search_space: SearchSpace
self._status_quo: Optional[Arm] = None
self._is_test: bool
self._name = name
self.description = description
self.runner = runner
self.is_test = is_test
self._data_by_trial: Dict[int, OrderedDict[int, Data]] = {}
self._experiment_type: Optional[str] = experiment_type
# pyre-fixme[4]: Attribute must be annotated.
self._optimization_config = None
self._tracking_metrics: Dict[str, Metric] = {}
self._time_created: datetime = datetime.now()
self._trials: Dict[int, BaseTrial] = {}
self._properties: Dict[str, Any] = properties or {}
# pyre-fixme[4]: Attribute must be annotated.
self._default_data_type = default_data_type or DataType.DATA
# Used to keep track of whether any trials on the experiment
# specify a TTL. Since trials need to be checked for their TTL's
# expiration often, having this attribute helps avoid unnecessary
# TTL checks for experiments that do not use TTL.
self._trials_have_ttl = False
# Make sure all statuses appear in this dict, to avoid key errors.
self._trial_indices_by_status: Dict[TrialStatus, Set[int]] = {
status: set() for status in TrialStatus
}
self._arms_by_signature: Dict[str, Arm] = {}
self._arms_by_name: Dict[str, Arm] = {}
self.add_tracking_metrics(tracking_metrics or [])
# call setters defined below
self.search_space = search_space
self.status_quo = status_quo
if optimization_config is not None:
self.optimization_config = optimization_config
@property
def has_name(self) -> bool:
"""Return true if experiment's name is not None."""
return self._name is not None
@property
def name(self) -> str:
"""Get experiment name. Throws if name is None."""
if self._name is None:
raise ValueError("Experiment's name is None.")
return self._name
@name.setter
def name(self, name: str) -> None:
"""Set experiment name."""
self._name = name
@property
def is_test(self) -> bool:
"""Get whether the experiment is a test."""
return self._is_test
@is_test.setter
def is_test(self, is_test: bool) -> None:
"""Set whether the experiment is a test."""
if is_test:
logger.info(EXPERIMENT_IS_TEST_WARNING)
self._is_test = is_test
@property
def time_created(self) -> datetime:
"""Creation time of the experiment."""
return self._time_created
@property
def experiment_type(self) -> Optional[str]:
"""The type of the experiment."""
return self._experiment_type
@experiment_type.setter
def experiment_type(self, experiment_type: Optional[str]) -> None:
"""Set the type of the experiment."""
self._experiment_type = experiment_type
@property
def search_space(self) -> SearchSpace:
"""The search space for this experiment.
When setting a new search space, all parameter names and types
must be preserved. However, if no trials have been created, all
modifications are allowed.
"""
# TODO: maybe return a copy here to guard against implicit changes
return self._search_space
@search_space.setter
def search_space(self, search_space: SearchSpace) -> None:
# Allow all modifications when no trials present.
if not hasattr(self, "_search_space") or len(self.trials) < 1:
self._search_space = search_space
return
# At least 1 trial is present.
if self.immutable_search_space_and_opt_config:
raise UnsupportedError(
"Modifications of search space are disabled by the "
f"`{Keys.IMMUTABLE_SEARCH_SPACE_AND_OPT_CONF.value}` "
"property that is set to `True` on this experiment."
)
if len(search_space.parameters) < len(self._search_space.parameters):
raise ValueError(
"New search_space must contain all parameters in the existing."
)
for param_name, parameter in search_space.parameters.items():
if param_name not in self._search_space.parameters:
raise ValueError(
f"Cannot add new parameter `{param_name}` because "
"it is not defined in the existing search space."
)
elif (
parameter.parameter_type
!= self._search_space.parameters[param_name].parameter_type
):
raise ValueError(
f"Expected parameter `{param_name}` to be of type "
f"{self._search_space.parameters[param_name].parameter_type}, "
f"got {parameter.parameter_type}."
)
self._search_space = search_space
@property
def status_quo(self) -> Optional[Arm]:
"""The existing arm that new arms will be compared against."""
return self._status_quo
@status_quo.setter
def status_quo(self, status_quo: Optional[Arm]) -> None:
if status_quo is not None:
self.search_space.check_types(
parameterization=status_quo.parameters, raise_error=True
)
self.search_space.check_all_parameters_present(
parameterization=status_quo.parameters, raise_error=True
)
# Compute a unique name if "status_quo" is taken
name = "status_quo"
sq_idx = 0
arms_by_name = self.arms_by_name
while name in arms_by_name:
name = f"status_quo_e{sq_idx}"
sq_idx += 1
self._name_and_store_arm_if_not_exists(arm=status_quo, proposed_name=name)
logger.warn(
"Experiment's status_quo is updated. "
"Generally the status_quo should not be changed after being set."
)
# If old status_quo not present in any trials,
# remove from _arms_by_signature
if self._status_quo is not None:
persist_old_sq = False
for trial in self._trials.values():
# pyre-fixme[16]: `Optional` has no attribute `name`.
if self._status_quo.name in trial.arms_by_name:
persist_old_sq = True
break
if not persist_old_sq:
# pyre-fixme[16]: `Optional` has no attribute `signature`.
self._arms_by_signature.pop(self._status_quo.signature)
self._arms_by_name.pop(self._status_quo.name)
self._status_quo = status_quo
@property
def parameters(self) -> Dict[str, Parameter]:
"""The parameters in the experiment's search space."""
return self.search_space.parameters
@property
def arms_by_name(self) -> Dict[str, Arm]:
"""The arms belonging to this experiment, by their name."""
return self._arms_by_name
@property
def arms_by_signature(self) -> Dict[str, Arm]:
"""The arms belonging to this experiment, by their signature."""
return self._arms_by_signature
@property
def sum_trial_sizes(self) -> int:
"""Sum of numbers of arms attached to each trial in this experiment."""
return reduce(lambda a, b: a + len(b.arms_by_name), self._trials.values(), 0)
@property
def num_abandoned_arms(self) -> int:
"""How many arms attached to this experiment are abandoned."""
abandoned = set()
for trial in self.trials.values():
for x in trial.abandoned_arms:
abandoned.add(x)
return len(abandoned)
@property
def optimization_config(self) -> Optional[OptimizationConfig]:
"""The experiment's optimization config."""
return self._optimization_config
@property
def is_moo_problem(self) -> bool:
"""Whether the experiment's optimization config contains multiple objectives."""
if self.optimization_config is None:
return False
return not_none(self.optimization_config).is_moo_problem
@optimization_config.setter
def optimization_config(self, optimization_config: OptimizationConfig) -> None:
if (
len(self.trials) > 0
and getattr(self, "_optimization_config", None) is not None
and self.immutable_search_space_and_opt_config
):
raise UnsupportedError(
"Modifications of optimization config are disabled by the "
f"`{Keys.IMMUTABLE_SEARCH_SPACE_AND_OPT_CONF.value}` "
"property that is set to `True` on this experiment."
)
for metric_name in optimization_config.metrics.keys():
if metric_name in self._tracking_metrics:
self.remove_tracking_metric(metric_name)
self._optimization_config = optimization_config
if any(
isinstance(metric, MapMetric)
for metric in optimization_config.metrics.values()
):
self._default_data_type = DataType.MAP_DATA
@property
def data_by_trial(self) -> Dict[int, OrderedDict[int, Data]]:
"""Data stored on the experiment, indexed by trial index and storage time.
First key is trial index and second key is storage time in milliseconds.
For a given trial, data is ordered by storage time, so first added data
will appear first in the list.
"""
return self._data_by_trial
@property
def immutable_search_space_and_opt_config(self) -> bool:
"""Boolean representing whether search space and metrics on this experiment
are mutable (by default they are).
NOTE: For experiments with immutable search spaces and metrics, generator
runs will not store copies of search space and metrics, which improves
storage layer performance. Not keeping copies of those on generator runs
also disables keeping track of changes to search space and metrics,
thereby necessitating that those attributes be immutable on experiment.
"""
return self._properties.get(Keys.IMMUTABLE_SEARCH_SPACE_AND_OPT_CONF, False)
@property
def tracking_metrics(self) -> List[Metric]:
return list(self._tracking_metrics.values())
[docs] def add_tracking_metric(self, metric: Metric) -> Experiment:
"""Add a new metric to the experiment.
Args:
metric: Metric to be added.
"""
if metric.name in self._tracking_metrics:
raise ValueError(
f"Metric `{metric.name}` already defined on experiment. "
"Use `update_tracking_metric` to update an existing metric definition."
)
optimization_config = self.optimization_config
if optimization_config and metric.name in optimization_config.metrics:
raise ValueError(
f"Metric `{metric.name}` already present in experiment's "
"OptimizationConfig. Set a new OptimizationConfig without this metric "
"before adding it to tracking metrics."
)
if isinstance(metric, MapMetric):
self._default_data_type = DataType.MAP_DATA
self._tracking_metrics[metric.name] = metric
return self
[docs] def add_tracking_metrics(self, metrics: List[Metric]) -> Experiment:
"""Add a list of new metrics to the experiment.
If any of the metrics are already defined on the experiment,
we raise an error and don't add any of them to the experiment
Args:
metrics: Metrics to be added.
"""
# Before setting any metrics, we validate none are already on
# the experiment
for metric in metrics:
self.add_tracking_metric(metric)
return self
[docs] def update_tracking_metric(self, metric: Metric) -> Experiment:
"""Redefine a metric that already exists on the experiment.
Args:
metric: New metric definition.
"""
if metric.name not in self._tracking_metrics:
raise ValueError(f"Metric `{metric.name}` doesn't exist on experiment.")
self._tracking_metrics[metric.name] = metric
return self
[docs] def remove_tracking_metric(self, metric_name: str) -> Experiment:
"""Remove a metric that already exists on the experiment.
Args:
metric_name: Unique name of metric to remove.
"""
if metric_name not in self._tracking_metrics:
raise ValueError(f"Metric `{metric_name}` doesn't exist on experiment.")
del self._tracking_metrics[metric_name]
return self
@property
def metrics(self) -> Dict[str, Metric]:
"""The metrics attached to the experiment."""
optimization_config_metrics: Dict[str, Metric] = {}
if self.optimization_config is not None:
optimization_config_metrics = self.optimization_config.metrics
return {**self._tracking_metrics, **optimization_config_metrics}
def _metrics_by_class(
self, metrics: Optional[List[Metric]] = None
) -> Dict[Type[Metric], List[Metric]]:
metrics_by_class: Dict[Type[Metric], List[Metric]] = defaultdict(list)
for metric in metrics or list(self.metrics.values()):
# By default, all metrics are grouped by their class for fetch;
# however, for some metrics, `fetch_trial_data_multi` of a
# superclass is used for fetch the subclassing metrics' data. In
# those cases, "fetch_multi_group_by_metric" property on metric
# will be set to a class other than its own (likely a superclass).
metrics_by_class[metric.fetch_multi_group_by_metric].append(metric)
return metrics_by_class
[docs] def fetch_data_results(
self,
metrics: Optional[List[Metric]] = None,
combine_with_last_data: bool = False,
overwrite_existing_data: bool = False,
**kwargs: Any,
) -> Dict[int, Dict[str, MetricFetchResult]]:
"""Fetches data for all trials on this experiment and for either the
specified metrics or all metrics currently on the experiment, if `metrics`
argument is not specified.
If a metric fetch fails, the Exception will be captured in the
MetricFetchResult along with a message.
NOTE: For metrics that are not available while trial is running, the data
may be retrieved from cache on the experiment. Data is cached on the experiment
via calls to `experiment.attach_data` and whether a given metric class is
available while trial is running is determined by the boolean returned from its
`is_available_while_running` class method.
Args:
metrics: If provided, fetch data for these metrics instead of the ones
defined on the experiment.
kwargs: keyword args to pass to underlying metrics' fetch data functions.
Returns:
A nested Dictionary from trial_index => metric_name => result
"""
return self._lookup_or_fetch_trials_results(
trials=list(self.trials.values()),
metrics=metrics,
combine_with_last_data=combine_with_last_data,
overwrite_existing_data=overwrite_existing_data,
**kwargs,
)
[docs] def fetch_trials_data_results(
self,
trial_indices: Iterable[int],
metrics: Optional[List[Metric]] = None,
combine_with_last_data: bool = False,
overwrite_existing_data: bool = False,
**kwargs: Any,
) -> Dict[int, Dict[str, MetricFetchResult]]:
"""Fetches data for specific trials on the experiment.
If a metric fetch fails, the Exception will be captured in the
MetricFetchResult along with a message.
NOTE: For metrics that are not available while trial is running, the data
may be retrieved from cache on the experiment. Data is cached on the experiment
via calls to `experiment.attach_data` and whether a given metric class is
available while trial is running is determined by the boolean returned from its
`is_available_while_running` class method.
Args:
trial_indices: Indices of trials, for which to fetch data.
metrics: If provided, fetch data for these metrics instead of the ones
defined on the experiment.
kwargs: keyword args to pass to underlying metrics' fetch data functions.
Returns:
A nested Dictionary from trial_index => metric_name => result
"""
return self._lookup_or_fetch_trials_results(
trials=self.get_trials_by_indices(trial_indices=trial_indices),
metrics=metrics,
combine_with_last_data=combine_with_last_data,
overwrite_existing_data=overwrite_existing_data,
**kwargs,
)
[docs] def fetch_data(
self,
metrics: Optional[List[Metric]] = None,
combine_with_last_data: bool = False,
overwrite_existing_data: bool = False,
**kwargs: Any,
) -> Data:
"""Fetches data for all trials on this experiment and for either the
specified metrics or all metrics currently on the experiment, if `metrics`
argument is not specified.
NOTE: For metrics that are not available while trial is running, the data
may be retrieved from cache on the experiment. Data is cached on the experiment
via calls to `experiment.attach_data` and whether a given metric class is
available while trial is running is determined by the boolean returned from its
`is_available_while_running` class method.
NOTE: This can be lossy (ex. a MapData could get implicitly cast to a Data and
lose rows) if Experiment.default_data_type is misconfigured!
Args:
metrics: If provided, fetch data for these metrics instead of the ones
defined on the experiment.
kwargs: keyword args to pass to underlying metrics' fetch data functions.
Returns:
Data for the experiment.
"""
results = self._lookup_or_fetch_trials_results(
trials=list(self.trials.values()),
metrics=metrics,
combine_with_last_data=combine_with_last_data,
overwrite_existing_data=overwrite_existing_data,
**kwargs,
)
base_metric_cls = (
MapMetric if self.default_data_constructor == MapData else Metric
)
return base_metric_cls._unwrap_experiment_data_multi(
results=results,
)
[docs] def fetch_trials_data(
self,
trial_indices: Iterable[int],
metrics: Optional[List[Metric]] = None,
combine_with_last_data: bool = False,
overwrite_existing_data: bool = False,
**kwargs: Any,
) -> Data:
"""Fetches data for specific trials on the experiment.
NOTE: For metrics that are not available while trial is running, the data
may be retrieved from cache on the experiment. Data is cached on the experiment
via calls to `experiment.attach_data` and whetner a given metric class is
available while trial is running is determined by the boolean returned from its
`is_available_while_running` class method.
NOTE: This can be lossy (ex. a MapData could get implicitly cast to a Data and
lose rows) if Experiment.default_data_type is misconfigured!
Args:
trial_indices: Indices of trials, for which to fetch data.
metrics: If provided, fetch data for these metrics instead of the ones
defined on the experiment.
kwargs: Keyword args to pass to underlying metrics' fetch data functions.
Returns:
Data for the specific trials on the experiment.
"""
results = self._lookup_or_fetch_trials_results(
trials=self.get_trials_by_indices(trial_indices=trial_indices),
metrics=metrics,
combine_with_last_data=combine_with_last_data,
overwrite_existing_data=overwrite_existing_data,
**kwargs,
)
base_metric_cls = (
MapMetric if self.default_data_constructor == MapData else Metric
)
return base_metric_cls._unwrap_experiment_data_multi(
results=results,
)
def _lookup_or_fetch_trials_results(
self,
trials: List[BaseTrial],
metrics: Optional[Iterable[Metric]] = None,
combine_with_last_data: bool = False,
overwrite_existing_data: bool = False,
**kwargs: Any,
) -> Dict[int, Dict[str, MetricFetchResult]]:
if not self.metrics and not metrics:
raise ValueError(
"No metrics to fetch data for, as no metrics are defined for "
"this experiment, and none were passed in to `fetch_data`."
)
if not any(t.status.expecting_data for t in trials):
logger.info("No trials are in a state expecting data.")
return {}
metrics_to_fetch = list(metrics or self.metrics.values())
metrics_by_class = self._metrics_by_class(metrics=metrics_to_fetch)
results: Dict[int, Dict[str, MetricFetchResult]] = {}
contains_new_data = False
for metric_cls, metrics in metrics_by_class.items():
first_metric_of_group = metrics[0]
(
new_fetch_results,
new_results_contains_new_data,
) = first_metric_of_group.fetch_data_prefer_lookup(
experiment=self,
metrics=metrics_by_class[metric_cls],
trials=trials,
**kwargs,
)
contains_new_data = contains_new_data or new_results_contains_new_data
# Merge in results
results = {
trial.index: {
**(
new_fetch_results[trial.index]
if trial.index in new_fetch_results
else {}
),
**(results[trial.index] if trial.index in results else {}),
}
for trial in trials
}
if contains_new_data:
try:
self.attach_fetch_results(
results=results,
combine_with_last_data=combine_with_last_data,
overwrite_existing_data=overwrite_existing_data,
)
except ValueError as e:
# TODO: Log and track these unexpected errors.
logger.error(
f"Encountered ValueError {e} while attaching results. Proceeding "
"and returning Results fetched without attaching."
)
return results
@copy_doc(BaseTrial.fetch_data)
def _fetch_trial_data(
self, trial_index: int, metrics: Optional[List[Metric]] = None, **kwargs: Any
) -> Dict[str, MetricFetchResult]:
trial = self.trials[trial_index]
trial_data = self._lookup_or_fetch_trials_results(
trials=[trial], metrics=metrics, **kwargs
)
if trial_index in trial_data:
return trial_data[trial_index]
return {}
[docs] def attach_data(
self,
data: Data,
combine_with_last_data: bool = False,
overwrite_existing_data: bool = False,
) -> int:
"""Attach data to experiment. Stores data in `experiment._data_by_trial`,
to be looked up via `experiment.lookup_data_for_trial`.
Args:
data: Data object to store.
combine_with_last_data: By default, when attaching data, it's identified
by its timestamp, and `experiment.lookup_data_for_trial` returns
data by most recent timestamp. Sometimes, however, we want to combine
the data from multiple calls to `attach_data` into one dataframe.
This might be because:
- We attached data for some metrics at one point and data for
the rest of the metrics later on.
- We attached data for some fidelity at one point and data for
another fidelity later one.
To achieve that goal, set `combine_with_last_data` to `True`.
In this case, we will take the most recent previously attached
data, append the newly attached data to it, attach a new
Data object with the merged result, and delete the old one.
Afterwards, calls to `lookup_data_for_trial` will return this
new combined data object. This operation will also validate that the
newly added data does not contain observations for metrics that
already have observations at the same fidelity in the most recent data.
overwrite_existing_data: By default, we keep around all data that has
ever been attached to the experiment. However, if we know that
the incoming data contains all the information we need for a given
trial, we can replace the existing data for that trial, thereby
reducing the amount we need to store in the database.
Returns:
Timestamp of storage in millis.
"""
if combine_with_last_data and overwrite_existing_data:
raise UnsupportedError(
"Cannot set both combine_with_last_data=True and "
"overwrite_existing_data=True. Data can either be "
"combined, or overwritten, or neither."
)
data_type = type(data)
data_init_args = data.deserialize_init_args(data.serialize_init_args(data))
if data.df.empty:
raise ValueError("Data to attach is empty.")
metrics_not_on_exp = set(data.true_df["metric_name"].values) - set(
self.metrics.keys()
)
if metrics_not_on_exp:
logger.info(
f"Attached data has some metrics ({metrics_not_on_exp}) that are "
"not among the metrics on this experiment. Note that attaching data "
"will not automatically add those metrics to the experiment. "
"For these metrics to be automatically fetched by `experiment."
"fetch_data`, add them via `experiment.add_tracking_metric` or update "
"the experiment's optimization config."
)
cur_time_millis = current_timestamp_in_millis()
for trial_index, trial_df in data.true_df.groupby(data.true_df["trial_index"]):
# Overwrite `df` so that `data` only has current trial data.
data_init_args["df"] = trial_df
current_trial_data = (
self._data_by_trial[trial_index]
if trial_index in self._data_by_trial
else OrderedDict()
)
if combine_with_last_data and len(current_trial_data) > 0:
last_data_type, last_data = self._get_last_data_without_similar_rows(
current_trial_data=current_trial_data, new_df=trial_df
)
current_trial_data.popitem()
current_trial_data[cur_time_millis] = last_data_type.from_multiple_data(
[
last_data,
last_data_type(**data_init_args),
]
)
elif overwrite_existing_data:
if len(current_trial_data) > 0:
_, last_data = list(current_trial_data.items())[-1]
last_data_metrics = set(last_data.df["metric_name"])
new_data_metrics = set(trial_df["metric_name"])
difference = last_data_metrics.difference(new_data_metrics)
if len(difference) > 0:
raise ValueError(
"overwrite_trial_data is True, but the new data contains "
"only a subset of the metrics that are present in the "
f"previous data. Missing metrics: {difference}"
)
current_trial_data = OrderedDict(
{cur_time_millis: data_type(**data_init_args)}
)
else:
current_trial_data[cur_time_millis] = data_type(**data_init_args)
self._data_by_trial[trial_index] = current_trial_data
return cur_time_millis
@staticmethod
def _get_last_data_without_similar_rows(
current_trial_data: OrderedDict[int, Data], new_df: pd.DataFrame
) -> Tuple[Type[Data], Data]:
"""Get a copy of last data with rows filtered out sharing values for
"trial_index", "metric_name", and "arm_name" with the new data so we
can cleanly combine them.
Args:
current_trial_data: The data currently attached to a trial
new_df: A DataFrame containing new data to be attached
Returns:
A tuple of two things:
- The type of the last data that was attached
- A Data object with the most recent data attached, minus the
rows that share the same values
for "trial_index", "metric_name", and "arm_name" in new_df
"""
last_ts, last_data = list(current_trial_data.items())[-1]
# Get the init args other than 'df' for last data
# in case it was a child class of `Data`
last_data_init_args = last_data.deserialize_init_args(
last_data.serialize_init_args(last_data)
)
del last_data_init_args["df"]
last_data_type = type(last_data)
merge_keys = ["trial_index", "metric_name", "arm_name"] + (
# pyre-ignore[16]
last_data.map_keys
if issubclass(last_data_type, MapData)
else []
)
# this merge is like a SQL left join on merge keys
# it will return a dataframe with the columns in merge_keys
# plus "_merge" and any other columns in last_data.true_df with _left appended
# plus any other columns in new_df with _right appended
merged = pd.merge(
last_data.true_df,
new_df,
on=merge_keys,
how="left",
indicator=True,
suffixes=("_left", "_right"),
)
# Filter out all rows that are also present in new_df
last_df = merged[merged["_merge"] == "left_only"]
# Drop the _merge column
last_df = last_df.drop(columns=["_merge"])
# Drop columns ending with "_right", which should all have null values
right_columns = [c for c in last_df.columns if re.match(r".*_right$", c)]
last_df = last_df.drop(columns=right_columns)
# Remove the "_left" suffix from the column names
last_df.columns = last_df.columns.str.replace(r"_left$", "", regex=True)
return type(last_data), last_data_type(df=last_df, **last_data_init_args)
[docs] def attach_fetch_results(
self,
results: Mapping[int, Mapping[str, MetricFetchResult]],
combine_with_last_data: bool = False,
overwrite_existing_data: bool = False,
) -> Optional[int]:
"""
UNSAFE: Prefer to use attach_data directly instead.
Attach fetched data results to the Experiment so they will not have to be
fetched again. Returns the timestamp from attachment, which is used as a
dict key for _data_by_trial.
NOTE: Any Errs in the results passed in will silently be dropped! This will
cause the Experiment to fail to find them in the _data_by_trial cache and
attempt to refetch at fetch time. If this is not your intended behavior you
MUST resolve your results first and use attach_data directly instead.
"""
flattened = [
result for sublist in results.values() for result in sublist.values()
]
oks: List[Ok[Data, MetricFetchE]] = [
result for result in flattened if isinstance(result, Ok)
]
for result in flattened:
if isinstance(result, Err):
logger.error(
"Discovered Metric fetching Err while attaching data "
f"{result.err}. "
"Ignoring for now -- will retry query on next call to fetch."
)
if len(oks) < 1:
return None
data = self.default_data_constructor.from_multiple_data(
data=[ok.ok for ok in oks]
)
return self.attach_data(
data=data,
combine_with_last_data=combine_with_last_data,
overwrite_existing_data=overwrite_existing_data,
)
[docs] def lookup_data_for_ts(self, timestamp: int) -> Data:
"""Collect data for all trials stored at this timestamp.
Useful when many trials' data was fetched and stored simultaneously
and user wants to retrieve same collection of data later.
Can also be used to lookup specific data for a single trial
when storage time is known.
Args:
timestamp: Timestamp in millis at which data was stored.
Returns:
Data object with all data stored at the timestamp.
"""
trial_datas = []
for _trial_index, ts_to_data in self._data_by_trial.items():
if timestamp in ts_to_data:
trial_datas.append(ts_to_data[timestamp])
return self.default_data_constructor.from_multiple_data(trial_datas)
[docs] def lookup_data_for_trial(
self,
trial_index: int,
) -> Tuple[Data, int]:
"""Lookup stored data for a specific trial.
Returns latest data object, and its storage timestamp, present for this trial.
Returns empty data and -1 if no data present.
Args:
trial_index: The index of the trial to lookup data for.
Returns:
The requested data object, and its storage timestamp in milliseconds.
"""
try:
trial_data_dict = self._data_by_trial[trial_index]
except KeyError:
return (self.default_data_constructor(), -1)
if len(trial_data_dict) == 0:
return (self.default_data_constructor(), -1)
storage_time = max(trial_data_dict.keys())
trial_data = trial_data_dict[storage_time]
return trial_data, storage_time
[docs] def lookup_data(
self,
trial_indices: Optional[Iterable[int]] = None,
) -> Data:
"""Lookup data for all trials on this experiment and for either the
specified metrics or all metrics currently on the experiment, if `metrics`
argument is not specified.
Args:
trial_indices: Indices of trials, for which to fetch data.
Returns:
Data for the experiment.
"""
data_by_trial = []
trial_indices = trial_indices or list(self.trials.keys())
for trial_index in trial_indices:
data_by_trial.append(
self.lookup_data_for_trial(
trial_index=trial_index,
)[0]
)
if not data_by_trial:
return self.default_data_constructor()
last_data = data_by_trial[-1]
last_data_type = type(last_data)
data = last_data_type.from_multiple_data(data_by_trial)
return data
@property
def num_trials(self) -> int:
"""How many trials are associated with this experiment."""
return len(self._trials)
@property
def trials(self) -> Dict[int, BaseTrial]:
"""The trials associated with the experiment.
NOTE: If some trials on this experiment specify their TTL, `RUNNING` trials
will be checked for whether their TTL elapsed during this call. Found past-
TTL trials will be marked as `FAILED`.
"""
self._check_TTL_on_running_trials()
return self._trials
@property
def trials_by_status(self) -> Dict[TrialStatus, List[BaseTrial]]:
"""Trials associated with the experiment, grouped by trial status."""
# Make sure all statuses appear in this dict, to avoid key errors.
return {
status: self.get_trials_by_indices(trial_indices=idcs)
for status, idcs in self.trial_indices_by_status.items()
}
@property
def trials_expecting_data(self) -> List[BaseTrial]:
"""List[BaseTrial]: the list of all trials for which data has arrived
or is expected to arrive.
"""
return [trial for trial in self.trials.values() if trial.status.expecting_data]
@property
def completed_trials(self) -> List[BaseTrial]:
"""List[BaseTrial]: the list of all trials for which data has arrived
or is expected to arrive.
"""
return self.trials_by_status[TrialStatus.COMPLETED]
@property
def trial_indices_by_status(self) -> Dict[TrialStatus, Set[int]]:
"""Indices of trials associated with the experiment, grouped by trial
status.
"""
self._check_TTL_on_running_trials() # Marks past-TTL trials as failed.
return self._trial_indices_by_status
@property
def running_trial_indices(self) -> Set[int]:
"""Indices of running trials, associated with the experiment."""
return self._trial_indices_by_status[TrialStatus.RUNNING]
@property
def default_data_type(self) -> DataType:
return self._default_data_type
@property
# pyre-fixme[24]: Generic type `type` expects 1 type parameter, use
# `typing.Type` to avoid runtime subscripting errors.
def default_data_constructor(self) -> Type:
return DATA_TYPE_LOOKUP[self.default_data_type]
[docs] def new_trial(
self,
generator_run: Optional[GeneratorRun] = None,
trial_type: Optional[str] = None,
ttl_seconds: Optional[int] = None,
) -> Trial:
"""Create a new trial associated with this experiment.
Args:
generator_run: GeneratorRun, associated with this trial.
Trial has only one arm attached to it and this generator_run
must therefore contain one arm. This arm can 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.
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.
"""
if ttl_seconds is not None:
self._trials_have_ttl = True
return Trial(
experiment=self,
trial_type=trial_type,
generator_run=generator_run,
ttl_seconds=ttl_seconds,
)
[docs] def new_batch_trial(
self,
generator_run: Optional[GeneratorRun] = None,
trial_type: Optional[str] = None,
optimize_for_power: Optional[bool] = False,
ttl_seconds: Optional[int] = None,
lifecycle_stage: Optional[LifecycleStage] = None,
) -> BatchTrial:
"""Create a new batch trial associated with this experiment.
Args:
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.
lifecycle_stage: The stage of the experiment lifecycle that this
trial represents
"""
if ttl_seconds is not None:
self._trials_have_ttl = True
return BatchTrial(
experiment=self,
trial_type=trial_type,
generator_run=generator_run,
optimize_for_power=optimize_for_power,
ttl_seconds=ttl_seconds,
lifecycle_stage=lifecycle_stage,
)
[docs] def get_trials_by_indices(self, trial_indices: Iterable[int]) -> List[BaseTrial]:
"""Grabs trials on this experiment by their indices."""
trial_indices = list(trial_indices)
try:
return [self.trials[idx] for idx in trial_indices]
except KeyError:
missing = set(trial_indices) - set(self.trials)
raise ValueError(
f"Trial indices {missing} are not associated with the experiment."
)
[docs] def reset_runners(self, runner: Runner) -> None:
"""Replace all candidate trials runners.
Args:
runner: New runner to replace with.
"""
for trial in self._trials.values():
if trial.status == TrialStatus.CANDIDATE:
trial.runner = runner
self.runner = runner
def _attach_trial(self, trial: BaseTrial, index: Optional[int] = None) -> int:
"""Attach a trial to this experiment.
Should only be called within the trial constructor.
Args:
trial: The trial to be attached.
index: If specified, the trial's index will be set accordingly.
This should generally not be specified, as the index
will be automatically determined based on the number
of existing trials. This is only used for the purpose
of loading from storage.
Returns:
The index of the trial within the experiment's trial list.
"""
if trial.experiment is not self:
raise ValueError("BatchTrial does not belong to this experiment.")
for existing_trial in self._trials.values():
if existing_trial is trial:
raise ValueError("BatchTrial already attached to experiment.")
if index is not None and index in self._trials:
logger.debug(
f"Trial index {index} already exists on the experiment. Overwriting."
)
index = (
index
if index is not None
else (0 if len(self._trials) == 0 else max(self._trials.keys()) + 1)
)
self._trials[index] = trial
return index
[docs] def validate_trials(self, trials: Iterable[BaseTrial]) -> None:
"""Raise ValueError if any of the trials in the input are not from
this experiment.
"""
for t in trials:
if t.experiment is not self:
raise ValueError(
"All trials must be from the input experiment (name: "
f"{self.name}), but trial #{t.index} is from "
f"another experiment (name: {t.experiment.name})."
)
[docs] def warm_start_from_old_experiment(
self,
old_experiment: Experiment,
copy_run_metadata_keys: Optional[List[str]] = None,
trial_statuses_to_copy: Optional[List[TrialStatus]] = None,
search_space_check_membership_raise_error: bool = True,
) -> List[Trial]:
"""Copy all completed trials with data from an old Ax expeirment to this one.
This function checks that the parameters of each trial are members of the
current experiment's search_space.
NOTE: Currently only handles experiments with 1-arm ``Trial``-s, not
``BatchTrial``-s as there has not yet been need for support of the latter.
Args:
old_experiment: The experiment from which to transfer trials and data
copy_run_metadata_keys: A list of keys denoting which items to copy over
from each trial's run_metadata. Defaults to
``old_experiment.runner.run_metadata_report_keys``.
trial_statuses_to_copy: All trials with a status in this list will be
copied. By default, copies all ``RUNNING``, ``COMPLETED``,
``ABANDONED``, and ``EARLY_STOPPED`` trials.
search_space_check_membership_raise_error: Whether to raise an exception
if the warm started trials being imported fall outside of the
defined search space.
Returns:
List of trials successfully copied from old_experiment to this one
"""
if len(self.trials) > 0:
raise ValueError(
f"Can only warm-start experiments that don't yet have trials. "
f"Experiment {self._name} has {len(self.trials)} trials."
)
if copy_run_metadata_keys is None and old_experiment.runner is not None:
copy_run_metadata_keys = old_experiment.runner.run_metadata_report_keys
old_parameter_names = set(old_experiment.search_space.parameters.keys())
parameter_names = set(self.search_space.parameters.keys())
if old_parameter_names.symmetric_difference(parameter_names):
raise ValueError(
f"Cannot warm-start experiment '{self._name}' from experiment "
f"'{old_experiment._name}' due to mismatch in search space parameters."
f"Parameters in '{self._name}' but not in '{old_experiment._name}': "
f"{old_parameter_names - parameter_names}. Vice-versa: "
f"{parameter_names - old_parameter_names}."
)
trial_statuses_to_copy = (
trial_statuses_to_copy
if trial_statuses_to_copy is not None
else DEFAULT_STATUSES_TO_WARM_START
)
warm_start_trials = [
trial
for trial in old_experiment.trials.values()
if trial.status in trial_statuses_to_copy
]
copied_trials = []
for trial in warm_start_trials:
if not isinstance(trial, Trial):
raise NotImplementedError(
"Only experiments with 1-arm trials currently supported."
)
self.search_space.check_membership(
not_none(trial.arm).parameters,
raise_error=search_space_check_membership_raise_error,
)
dat, ts = old_experiment.lookup_data_for_trial(trial_index=trial.index)
# Set trial index and arm name to their values in new trial.
new_trial = self.new_trial()
add_arm_and_prevent_naming_collision(
new_trial=new_trial,
old_trial=trial,
old_experiment_name=old_experiment._name,
)
new_trial.mark_running(no_runner_required=True)
new_trial._properties["source"] = (
f"Warm start from Experiment: `{old_experiment._name}`, "
f"trial: `{trial.index}`"
)
# Associates a generation_model_key to the new trial.
generation_model_key = trial._properties.get("generation_model_key")
if generation_model_key is None and trial.generator_run is not None:
generation_model_key = trial.generator_run._model_key or "Manual"
new_trial._properties["generation_model_key"] = generation_model_key
if copy_run_metadata_keys is not None:
for run_metadata_field in copy_run_metadata_keys:
new_trial.update_run_metadata(
{run_metadata_field: trial.run_metadata.get(run_metadata_field)}
)
# Trial has data, so we replicate it on the new experiment.
has_data = ts != -1 and not dat.df.empty
if has_data:
new_df = dat.true_df.copy()
new_df["trial_index"].replace(
{trial.index: new_trial.index}, inplace=True
)
new_df["arm_name"].replace(
{not_none(trial.arm).name: not_none(new_trial.arm).name},
inplace=True,
)
# Attach updated data to new trial on experiment.
old_data = (
old_experiment.default_data_constructor(
df=new_df,
map_key_infos=checked_cast(
MapData, old_experiment.lookup_data()
).map_key_infos,
)
if old_experiment.default_data_type == DataType.MAP_DATA
else old_experiment.default_data_constructor(df=new_df)
)
self.attach_data(data=old_data)
if trial.status == TrialStatus.ABANDONED:
new_trial.mark_abandoned(reason=trial.abandoned_reason)
elif trial.status is not TrialStatus.RUNNING:
new_trial.mark_as(trial.status)
copied_trials.append(new_trial)
if self._name is not None:
logger.info(
f"Copied {len(copied_trials)} completed trials and their data "
f"from {old_experiment._name} to {self._name}."
)
else:
logger.info(
f"Copied {len(copied_trials)} completed trials and their data "
f"from {old_experiment._name}."
)
return copied_trials
def _name_and_store_arm_if_not_exists(self, arm: Arm, proposed_name: str) -> None:
"""Tries to lookup arm with same signature, otherwise names and stores it.
- Looks up if arm already exists on experiment
- If so, name the input arm the same as the existing arm
- else name the arm with given name and store in _arms_by_signature
Args:
arm: The arm object to name.
proposed_name: The name to assign if it doesn't have one already.
"""
# If arm is identical to an existing arm, return that
# so that the names match.
if arm.signature in self.arms_by_signature:
existing_arm = self.arms_by_signature[arm.signature]
if arm.has_name:
if arm.name != existing_arm.name:
raise ValueError(
f"Arm already exists with name {existing_arm.name}, "
f"which doesn't match given arm name of {arm.name}."
)
else:
arm.name = existing_arm.name
else:
if not arm.has_name:
arm.name = proposed_name
self._register_arm(arm)
def _register_arm(self, arm: Arm) -> None:
"""Add a new arm to the experiment, updating the relevant
lookup dictionaries.
Args:
arm: Arm to add
"""
self._arms_by_signature[arm.signature] = arm
self._arms_by_name[arm.name] = arm
def _check_TTL_on_running_trials(self) -> None:
"""Checks whether any past-TTL trials are still marked as `RUNNING`
and marks them as failed if so.
NOTE: this function just calls `trial.status` for each trial, as the
computation of that property checks the TTL for trials.
"""
if not self._trials_have_ttl:
return
running = list(self._trial_indices_by_status[TrialStatus.RUNNING])
for idx in running:
self._trials[idx].status # `status` property checks TTL if applicable.
def __repr__(self) -> str:
return self.__class__.__name__ + f"({self._name})"
# --- MultiTypeExperiment convenience functions ---
#
# Certain functionalities have special behavior for multi-type experiments.
# This defines the base behavior for regular experiments that will be
# overridden in the MultiTypeExperiment class.
@property
def default_trial_type(self) -> Optional[str]:
"""Default trial type assigned to trials in this experiment.
In the base experiment class this is always None. For experiments
with multiple trial types, use the MultiTypeExperiment class.
"""
return None
[docs] def runner_for_trial(self, trial: BaseTrial) -> Optional[Runner]:
"""The default runner to use for a given trial.
In the base experiment class, this is always the default experiment runner.
For experiments with multiple trial types, use the MultiTypeExperiment class.
"""
return self.runner
[docs] def supports_trial_type(self, trial_type: Optional[str]) -> bool:
"""Whether this experiment allows trials of the given type.
The base experiment class only supports None. For experiments
with multiple trial types, use the MultiTypeExperiment class.
"""
return trial_type is None
[docs] def attach_trial(
self,
parameterizations: List[TParameterization],
arm_names: Optional[List[str]] = None,
ttl_seconds: Optional[int] = None,
run_metadata: Optional[Dict[str, Any]] = None,
optimize_for_power: bool = False,
) -> Tuple[Dict[str, TParameterization], int]:
"""Attach a new trial with the given parameterization to the experiment.
Args:
parameterizations: List of parameterization for the new trial. If
only one is provided a single-arm Trial is created. If multiple
arms are provided a BatchTrial is created.
arm_names: Names of arm(s) in the new trial.
ttl_seconds: If specified, will consider the trial failed after this
many seconds. Used to detect dead trials that were not marked
failed properly.
run_metadata: Metadata to attach to the trial.
optimize_for_power: For BatchTrial only.
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.
Returns:
Tuple of arm name to parameterization dict, and trial index from
newly created trial.
"""
# If more than one parameterization is provided,
# proceed with a batch trial
is_batch = len(parameterizations) > 1
# Validate search space membership for all parameterizations
for parameterization in parameterizations:
self.search_space.validate_membership(parameters=parameterization)
# Validate number of arm names if any arm names are provided.
named_arms = False
arm_names = arm_names or []
if len(arm_names) > 0:
named_arms = True
if len(arm_names) != len(parameterizations):
raise UserInputError(
f"Number of arm names ({len(arm_names)} "
"does not match number of parameterizations "
f"({len(parameterizations)})."
)
# Prepare arm(s) to be added to the trial created later
arms = [
Arm(
parameters=parameterization,
name=arm_names[i] if named_arms else None,
)
for i, parameterization in enumerate(parameterizations)
]
# Create the trial and add arm(s)
trial = None
if is_batch:
# TODO: HSS support for batch trials.
if self.search_space.is_hierarchical:
raise NotImplementedError(
"Support for batch trials "
"in hierarchical search space coming soon. Let "
"the Ax developers know if you have a use "
"case for it."
)
trial = self.new_batch_trial(
ttl_seconds=ttl_seconds, optimize_for_power=optimize_for_power
).add_arms_and_weights(arms=arms)
else:
# If search space is hierarchical, we need to store dummy values of
# parameters that are not in the arm (but are in flattened
# search space), as metadata, so later we are able to make the
# data for this arm "complete" in the flattened search space.
candidate_metadata = None
if self.search_space.is_hierarchical:
hss = checked_cast(HierarchicalSearchSpace, self.search_space)
candidate_metadata = hss.cast_observation_features(
observation_features=hss.flatten_observation_features(
observation_features=observation.ObservationFeatures(
parameters=parameterizations[0]
),
inject_dummy_values_to_complete_flat_parameterization=True,
)
).metadata
trial = self.new_trial(ttl_seconds=ttl_seconds).add_arm(
arms[0],
candidate_metadata=candidate_metadata,
)
trial.mark_running(no_runner_required=True)
logger.info(
"Attached custom parameterizations "
f"{round_floats_for_logging(item=parameterizations)} "
f"as trial {trial.index}."
)
if run_metadata is not None:
trial.update_run_metadata(metadata=run_metadata)
return {arm.name: arm.parameters for arm in trial.arms}, trial.index
[docs]def add_arm_and_prevent_naming_collision(
new_trial: Trial, old_trial: Trial, old_experiment_name: Optional[str] = None
) -> None:
# Add all of an old trial's arms to a new trial. Rename any arm with auto-generated
# naming format to prevent naming collisions during warm-start. If an old
# experiment name is provided, append that to the original arm name. Else, clear
# the arm name. Preserves all names not matching the automatic naming format.
# experiment is not named, clear the arm's name.
# `arm_index` is 0 since all trials are single-armed.
old_arm_name = not_none(old_trial.arm).name
has_default_name = bool(old_arm_name == old_trial._get_default_name(arm_index=0))
if has_default_name:
new_arm = not_none(old_trial.arm).clone(clear_name=True)
if old_experiment_name is not None:
new_arm.name = f"{old_arm_name}_{old_experiment_name}"
new_trial.add_arm(new_arm)
else:
try:
new_trial.add_arm(not_none(old_trial.arm).clone(clear_name=False))
except ValueError as e:
warnings.warn(
f"Attaching arm {old_trial.arm} to trial {new_trial} while preserving "
f"its name failed with error: {e}. Retrying with `clear_name=True`."
)
new_trial.add_arm(not_none(old_trial.arm).clone(clear_name=True))
