#!/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 json
import logging
import warnings
from typing import Any, Dict, List, Optional, Tuple, Union, TypeVar, Type
import ax.service.utils.best_point as best_point_utils
import numpy as np
import pandas as pd
from ax.core.abstract_data import AbstractDataFrameData
from ax.core.arm import Arm
from ax.core.base_trial import BaseTrial
from ax.core.batch_trial import BatchTrial
from ax.core.experiment import DataType, Experiment
from ax.core.generator_run import GeneratorRun
from ax.core.objective import MultiObjective, Objective
from ax.core.observation import ObservationFeatures
from ax.core.trial import Trial
from ax.core.types import (
TEvaluationOutcome,
TModelPredictArm,
TParameterization,
TParamValue,
)
from ax.exceptions.constants import CHOLESKY_ERROR_ANNOTATION
from ax.exceptions.core import OptimizationComplete
from ax.exceptions.core import UnsupportedPlotError, UnsupportedError
from ax.modelbridge.dispatch_utils import choose_generation_strategy
from ax.modelbridge.generation_strategy import GenerationStrategy
from ax.modelbridge.modelbridge_utils import (
get_pending_observation_features_based_on_trial_status,
)
from ax.plot.base import AxPlotConfig
from ax.plot.contour import plot_contour
from ax.plot.feature_importances import plot_feature_importance_by_feature
from ax.plot.helper import _format_dict, _get_in_sample_arms
from ax.plot.trace import optimization_trace_single_method
from ax.service.utils.instantiation import (
data_and_evaluations_from_raw_data,
make_experiment,
ObjectiveProperties,
build_objective_threshold,
)
from ax.service.utils.report_utils import exp_to_df
from ax.service.utils.with_db_settings_base import DBSettings, WithDBSettingsBase
from ax.storage.json_store.decoder import (
generation_strategy_from_json,
object_from_json,
)
from ax.storage.json_store.encoder import object_to_json
from ax.utils.common.executils import retry_on_exception
from ax.utils.common.logger import _round_floats_for_logging, get_logger
from ax.utils.common.typeutils import (
checked_cast,
checked_cast_complex,
checked_cast_dict,
checked_cast_optional,
not_none,
)
from botorch.utils.sampling import manual_seed
logger = get_logger(__name__)
AxClientSubclass = TypeVar("AxClientSubclass", bound="AxClient")
[docs]class AxClient(WithDBSettingsBase):
"""
Convenience handler for management of experimentation cycle through a
service-like API. External system manages scheduling of the cycle and makes
calls to this client to get next suggestion in the experiment and log back
data from the evaluation of that suggestion.
Note: `AxClient` expects to only propose 1 arm (suggestion) per trial; support
for use cases that require use of batches is coming soon.
Two custom types used in this class for convenience are `TParamValue` and
`TParameterization`. Those are shortcuts for `Union[str, bool, float, int]`
and `Dict[str, Union[str, bool, float, int]]`, respectively.
Args:
generation_strategy: Optional generation strategy. If not set, one is
intelligently chosen based on properties of search space.
db_settings: Settings for saving and reloading the underlying experiment
to a database. Expected to be of type
ax.storage.sqa_store.structs.DBSettings and require SQLAlchemy.
enforce_sequential_optimization: Whether to enforce that when it is
reasonable to switch models during the optimization (as prescribed
by `num_trials` in generation strategy), Ax will wait for enough trials
to be completed with data to proceed. Defaults to True. If set to
False, Ax will keep generating new trials from the previous model
until enough data is gathered. Use this only if necessary;
otherwise, it is more resource-efficient to
optimize sequentially, by waiting until enough data is available to
use the next model.
random_seed: Optional integer random seed, set to fix the optimization
random seed for reproducibility. Works only for Sobol quasi-random
generator and for BoTorch-powered models. For the latter models, the
trials generated from the same optimization setup with the same seed,
will be mostly similar, but the exact parameter values may still vary
and trials latter in the optimizations will diverge more and more.
This is because a degree of randomness is essential for high performance
of the Bayesian optimization models and is not controlled by the seed.
Note: In multi-threaded environments, the random seed is thread-safe,
but does not actually guarantee reproducibility. Whether the outcomes
will be exactly the same for two same operations that use the random
seed, depends on whether the threads modify the random state in the
same order across the two operations.
verbose_logging: Whether Ax should log significant optimization events,
defaults to `True`.
suppress_storage_errors: Whether to suppress SQL storage-related errors if
encounted. Only use if SQL storage is not important for the given use
case, since this will only log, but not raise, an exception if its
encountered while saving to DB or loading from it.
"""
BACH_TRIAL_RAW_DATA_FORMAT_ERROR_MESSAGE = (
"Raw data must be a dict for batched trials."
)
TRIAL_RAW_DATA_FORMAT_ERROR_MESSAGE = (
"Raw data must be data for a single arm for non batched trials."
)
def __init__(
self,
generation_strategy: Optional[GenerationStrategy] = None,
db_settings: Optional[DBSettings] = None,
enforce_sequential_optimization: bool = True,
random_seed: Optional[int] = None,
verbose_logging: bool = True,
suppress_storage_errors: bool = False,
) -> None:
super().__init__(
db_settings=db_settings, suppress_all_errors=suppress_storage_errors
)
if not verbose_logging:
logger.setLevel(logging.WARNING) # pragma: no cover
else:
logger.info(
"Starting optimization with verbose logging. To disable logging, "
"set the `verbose_logging` argument to `False`. Note that float "
"values in the logs are rounded to 2 decimal points."
)
self._generation_strategy = generation_strategy
self._experiment: Optional[Experiment] = None
self._enforce_sequential_optimization = enforce_sequential_optimization
self._random_seed = random_seed
self._suppress_storage_errors = suppress_storage_errors
if random_seed is not None:
logger.warning(
f"Random seed set to {random_seed}. Note that this setting "
"only affects the Sobol quasi-random generator "
"and BoTorch-powered Bayesian optimization models. For the latter "
"models, setting random seed to the same number for two optimizations "
"will make the generated trials similar, but not exactly the same, "
"and over time the trials will diverge more."
)
# ------------------------ Public API methods. ------------------------
[docs] def create_experiment(
self,
parameters: List[Dict[str, Union[TParamValue, List[TParamValue]]]],
name: Optional[str] = None,
objective_name: Optional[str] = None,
minimize: Optional[bool] = None,
objectives: Optional[Dict[str, ObjectiveProperties]] = None,
parameter_constraints: Optional[List[str]] = None,
outcome_constraints: Optional[List[str]] = None,
status_quo: Optional[TParameterization] = None,
overwrite_existing_experiment: bool = False,
experiment_type: Optional[str] = None,
tracking_metric_names: Optional[List[str]] = None,
choose_generation_strategy_kwargs: Optional[Dict[str, Any]] = None,
support_intermediate_data: bool = False,
immutable_search_space_and_opt_config: bool = True,
is_test: bool = False,
) -> None:
"""Create a new experiment and save it if DBSettings available.
Args:
parameters: List of dictionaries representing parameters in the
experiment search space.
Required elements in the dictionaries are:
1. "name" (name of parameter, string),
2. "type" (type of parameter: "range", "fixed", or "choice", string),
and one of the following:
3a. "bounds" for range parameters (list of two values, lower bound
first),
3b. "values" for choice parameters (list of values), or
3c. "value" for fixed parameters (single value).
Optional elements are:
1. "log_scale" (for float-valued range parameters, bool),
2. "value_type" (to specify type that values of this parameter should
take; expects "float", "int", "bool" or "str"),
3. "is_fidelity" (bool) and "target_value" (float) for fidelity
parameters,
4. "is_ordered" (bool) for choice parameters, and
5. "is_task" (bool) for task parameters.
6. "digits" (int) for float-valued range parameters.
name: Name of the experiment to be created.
objective_name[DEPRECATED]: Name of the metric used as objective
in this experiment. This metric must be present in `raw_data`
argument to `complete_trial`.
minimize[DEPRECATED]: Whether this experiment represents a minimization
problem.
objectives: Mapping from an objective name to object containing:
minimize: Whether this experiment represents a minimization problem.
threshold: The bound in the objective's threshold constraint.
parameter_constraints: List of string representation of parameter
constraints, such as "x3 >= x4" or "-x3 + 2*x4 - 3.5*x5 >= 2". For
the latter constraints, any number of arguments is accepted, and
acceptable operators are "<=" and ">=".
outcome_constraints: List of string representation of outcome
constraints of form "metric_name >= bound", like "m1 <= 3."
status_quo: Parameterization of the current state of the system.
If set, this will be added to each trial to be evaluated alongside
test configurations.
overwrite_existing_experiment: If an experiment has already been set
on this `AxClient` instance, whether to reset it to the new one.
If overwriting the experiment, generation strategy will be
re-selected for the new experiment and restarted.
To protect experiments in production, one cannot overwrite existing
experiments if the experiment is already stored in the database,
regardless of the value of `overwrite_existing_experiment`.
tracking_metric_names: Names of additional tracking metrics not used for
optimization.
choose_generation_strategy_kwargs: Keyword arguments to pass to
`choose_generation_strategy` function which determines what
generation strategy should be used when none was specified on init.
support_intermediate_data: Whether trials may report intermediate results
for trials that are still running (i.e. have not been completed via
`ax_client.complete_trial`).
immutable_search_space_and_opt_config: Whether it's possible to update the
search space and optimization config on this experiment after creation.
Defaults to True. If set to True, we won't store or load copies of the
search space and optimization config on each generator run, which will
improve storage performance.
is_test: Whether this experiment will be a test experiment (useful for
marking test experiments in storage etc). Defaults to False.
"""
objective_kwargs = {}
if (objective_name or minimize is not None) and objectives:
raise UnsupportedError(
"You may either pass an an objective object "
"or an objective_name and minimize param, but not both"
)
elif objectives and len(objectives.keys()) == 1:
objective = next(iter(objectives.keys()))
objective_kwargs["objective_name"] = objective
objective_kwargs["minimize"] = objectives[objective].minimize
elif objectives:
objective_kwargs["objectives"] = {
objective: ("minimize" if properties.minimize else "maximize")
for objective, properties in objectives.items()
}
objective_kwargs["objective_thresholds"] = [
build_objective_threshold(objective, properties)
for objective, properties in objectives.items()
if properties.threshold is not None
]
elif objective_name or minimize is not None:
objective_kwargs["objective_name"] = objective_name
objective_kwargs["minimize"] = minimize or False
warnings.warn(
"objective_name and minimize are deprecated",
category=DeprecationWarning,
)
experiment = make_experiment(
name=name,
parameters=parameters,
parameter_constraints=parameter_constraints,
outcome_constraints=outcome_constraints,
status_quo=status_quo,
experiment_type=experiment_type,
tracking_metric_names=tracking_metric_names,
support_intermediate_data=support_intermediate_data,
immutable_search_space_and_opt_config=immutable_search_space_and_opt_config,
is_test=is_test,
**objective_kwargs,
)
self._set_experiment(
experiment=experiment,
overwrite_existing_experiment=overwrite_existing_experiment,
)
self._set_generation_strategy(
choose_generation_strategy_kwargs=choose_generation_strategy_kwargs
)
self._save_generation_strategy_to_db_if_possible(
generation_strategy=self.generation_strategy,
)
[docs] @retry_on_exception(
logger=logger,
exception_types=(RuntimeError,),
suppress_all_errors=False,
wrap_error_message_in=CHOLESKY_ERROR_ANNOTATION,
)
def get_next_trial(
self, ttl_seconds: Optional[int] = None
) -> Tuple[TParameterization, int]:
"""
Generate trial with the next set of parameters to try in the iteration process.
Note: Service API currently supports only 1-arm trials.
Args:
ttl_seconds: If specified, will consider the trial failed after this
many seconds. Used to detect dead trials that were not marked
failed properly.
Returns:
Tuple of trial parameterization, trial index
"""
trial = self.experiment.new_trial(
generator_run=self._gen_new_generator_run(), ttl_seconds=ttl_seconds
)
logger.info(
f"Generated new trial {trial.index} with parameters "
f"{_round_floats_for_logging(item=not_none(trial.arm).parameters)}."
)
trial.mark_running(no_runner_required=True)
self._save_or_update_trial_in_db_if_possible(
experiment=self.experiment,
trial=trial,
)
# TODO[T79183560]: Ensure correct handling of generator run when using
# foreign keys.
self._update_generation_strategy_in_db_if_possible(
generation_strategy=self.generation_strategy,
new_generator_runs=[self.generation_strategy._generator_runs[-1]],
)
return not_none(trial.arm).parameters, trial.index
[docs] def get_current_trial_generation_limit(self) -> Tuple[int, bool]:
"""How many trials this ``AxClient`` instance can currently produce via
calls to ``get_next_trial``, before more trials are completed, and whether
the optimization is complete.
NOTE: If return value of this function is ``(0, False)``, no more trials
can currently be procuded by this ``AxClient`` instance, but optimization
is not completed; once more trials are completed with data, more new
trials can be generated.
Returns: a two-item tuple of:
- the number of trials that can currently be produced, with -1
meaning unlimited trials,
- whether no more trials can be produced by this ``AxClient``
instance at any point (e.g. if the search space is exhausted or
generation strategy is completed.
"""
# Ensure that experiment is set on the generation strategy.
if self.generation_strategy._experiment is None:
self.generation_strategy.experiment = self.experiment
return self.generation_strategy.current_generator_run_limit()
[docs] def get_next_trials(
self, max_trials: int, ttl_seconds: Optional[int] = None
) -> Tuple[Dict[int, TParameterization], bool]:
"""Generate as many trials as currently possible.
NOTE: Useful for running multiple trials in parallel: produces multiple trials,
with their number limited by:
- parallelism limit on current generation step,
- number of trials in current generation step,
- number of trials required to complete before moving to next generation step,
if applicable,
- and ``max_trials`` argument to this method.
Args:
max_trials: Limit on how many trials the call to this method should produce.
ttl_seconds: If specified, will consider the trial failed after this
many seconds. Used to detect dead trials that were not marked
failed properly.
Returns: two-item tuple of:
- mapping from trial indices to parameterizations in those trials,
- boolean indicator of whether optimization is completed and no more
trials can be generated going forward.
"""
gen_limit, optimization_complete = self.get_current_trial_generation_limit()
if optimization_complete:
return {}, True
# Trial generation limit of -1 indicates that unlimited trials can be
# generated, so we only want to limit `max_trials` if `trial_generation_
# limit` is non-negative.
if gen_limit >= 0:
max_trials = min(gen_limit, max_trials)
trials_dict = {}
for _ in range(max_trials):
try:
params, trial_index = self.get_next_trial(ttl_seconds=ttl_seconds)
trials_dict[trial_index] = params
except OptimizationComplete as err:
logger.info(
f"Encountered exception indicating optimization completion: {err}"
)
return trials_dict, True
# Check whether optimization is complete now that we generated a batch
# of trials.
_, optimization_complete = self.get_current_trial_generation_limit()
return trials_dict, optimization_complete
[docs] def abandon_trial(self, trial_index: int, reason: Optional[str] = None) -> None:
"""Abandons a trial and adds optional metadata to it.
Args:
trial_index: Index of trial within the experiment.
"""
trial = self._get_trial(trial_index=trial_index)
trial.mark_abandoned(reason=reason)
[docs] def complete_trial(
self,
trial_index: int,
raw_data: TEvaluationOutcome,
metadata: Optional[Dict[str, Union[str, int]]] = None,
sample_size: Optional[int] = None,
) -> None:
"""
Completes the trial with given metric values and adds optional metadata
to it.
NOTE: When ``raw_data`` does not specify SEM for a given metric, Ax
will default to the assumption that the data is noisy (specifically,
corrupted by additive zero-mean Gaussian noise) and that the
level of noise should be inferred by the optimization model. To
indicate that the data is noiseless, set SEM to 0.0, for example:
.. code-block:: python
ax_client.complete_trial(
trial_index=0,
raw_data={"my_objective": (objective_mean_value, 0.0)}
)
Args:
trial_index: Index of trial within the experiment.
raw_data: Evaluation data for the trial. Can be a mapping from
metric name to a tuple of mean and SEM, just a tuple of mean and
SEM if only one metric in optimization, or just the mean if SEM is
unknown (then Ax will infer observation noise level).
Can also be a list of (fidelities, mapping from
metric name to a tuple of mean and SEM).
metadata: Additional metadata to track about this run.
sample_size: Number of samples collected for the underlying arm,
optional.
"""
# Validate that trial can be completed.
trial = self._get_trial(trial_index=trial_index)
self._validate_can_complete_trial(trial=trial)
if not isinstance(trial_index, int): # pragma: no cover
raise ValueError(f"Trial index must be an int, got: {trial_index}.")
data_update_repr = self._update_trial_with_raw_data(
trial_index=trial_index,
raw_data=raw_data,
metadata=metadata,
sample_size=sample_size,
complete_trial=True,
combine_with_last_data=True,
)
logger.info(f"Completed trial {trial_index} with data: " f"{data_update_repr}.")
[docs] def update_trial_data(
self,
trial_index: int,
raw_data: TEvaluationOutcome,
metadata: Optional[Dict[str, Union[str, int]]] = None,
sample_size: Optional[int] = None,
) -> None:
"""
Attaches additional data for completed trial (for example, if trial was
completed with data for only one of the required metrics and more data
needs to be attached).
Args:
trial_index: Index of trial within the experiment.
raw_data: Evaluation data for the trial. Can be a mapping from
metric name to a tuple of mean and SEM, just a tuple of mean and
SEM if only one metric in optimization, or just the mean if there
is no SEM. Can also be a list of (fidelities, mapping from
metric name to a tuple of mean and SEM).
metadata: Additional metadata to track about this run.
sample_size: Number of samples collected for the underlying arm,
optional.
"""
if not isinstance(trial_index, int): # pragma: no cover
raise ValueError(f"Trial index must be an int, got: {trial_index}.")
trial = self._get_trial(trial_index=trial_index)
if not trial.status.is_completed:
raise ValueError(
f"Trial {trial.index} has not yet been completed with data."
"To complete it, use `ax_client.complete_trial`."
)
data_update_repr = self._update_trial_with_raw_data(
trial_index=trial_index,
raw_data=raw_data,
metadata=metadata,
sample_size=sample_size,
combine_with_last_data=True,
)
# Registering trial data update is needed for generation strategies that
# leverage the `update` functionality of model and bridge setup and therefore
# need to be aware of new data added to experiment. Usually this happends
# seamlessly, by looking at newly completed trials, but in this case trial
# status does not change, so we manually register the new data.
# Currently this call will only result in a `NotImplementedError` if generation
# strategy uses `update` (`GenerationStep.use_update` is False by default).
self.generation_strategy._register_trial_data_update(trial=trial)
logger.info(f"Added data: {data_update_repr} to trial {trial.index}.")
[docs] def log_trial_failure(
self, trial_index: int, metadata: Optional[Dict[str, str]] = None
) -> None:
"""Mark that the given trial has failed while running.
Args:
trial_index: Index of trial within the experiment.
metadata: Additional metadata to track about this run.
"""
trial = self.experiment.trials[trial_index]
trial.mark_failed()
logger.info(f"Registered failure of trial {trial_index}.")
if metadata is not None:
trial._run_metadata = metadata
self._save_experiment_to_db_if_possible(
experiment=self.experiment,
)
[docs] def attach_trial(
self, parameters: TParameterization, ttl_seconds: Optional[int] = None
) -> Tuple[TParameterization, int]:
"""Attach a new trial with the given parameterization to the experiment.
Args:
parameters: Parameterization of 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.
Returns:
Tuple of parameterization and trial index from newly created trial.
"""
self._validate_search_space_membership(parameters=parameters)
trial = self.experiment.new_trial(ttl_seconds=ttl_seconds).add_arm(
Arm(parameters=parameters)
)
trial.mark_running(no_runner_required=True)
logger.info(
"Attached custom parameterization "
f"{_round_floats_for_logging(item=parameters)} as trial {trial.index}."
)
self._save_or_update_trial_in_db_if_possible(
experiment=self.experiment,
trial=trial,
)
return not_none(trial.arm).parameters, trial.index
[docs] def get_trial_parameters(self, trial_index: int) -> TParameterization:
"""Retrieve the parameterization of the trial by the given index."""
return not_none(self._get_trial(trial_index).arm).parameters
[docs] def get_best_parameters(
self, use_model_predictions: bool = True
) -> Optional[Tuple[TParameterization, Optional[TModelPredictArm]]]:
"""Identifies the best parameterization tried in the experiment so far.
First attempts to do so with the model used in optimization and
its corresponding predictions if available. Falls back to the best raw
objective based on the data fetched from the experiment.
NOTE: ``TModelPredictArm`` is of the form:
({metric_name: mean}, {metric_name_1: {metric_name_2: cov_1_2}})
Args:
use_model_predictions: Whether to extract the best point using
model predictions or directly observed values. If ``True``,
the metric means and covariances in this method's output will
also be based on model predictions and may differ from the
observed values.
Returns:
Tuple of parameterization and model predictions for it.
"""
if not_none(self.experiment.optimization_config).is_moo_problem:
raise NotImplementedError( # pragma: no cover
"Please use `get_pareto_optimal_parameters` for multi-objective "
"problems."
)
# TODO[drfreund]: Find a way to include data for last trial in the
# calculation of best parameters.
return best_point_utils.get_best_parameters(
experiment=self.experiment, use_model_predictions=use_model_predictions
)
[docs] def get_pareto_optimal_parameters(
self, use_model_predictions: bool = True
) -> Optional[Dict[int, Tuple[TParameterization, TModelPredictArm]]]:
"""Identifies the best parameterizations tried in the experiment so far,
using model predictions if ``use_model_predictions`` is true and using
observed values from the experiment otherwise. By default, uses model
predictions to account for observation noise.
NOTE: The format of this method's output is as follows:
{ trial_index --> (parameterization, (means, covariances) }, where means
are a dictionary of form { metric_name --> metric_mean } and covariances
are a nested dictionary of form
{ one_metric_name --> { another_metric_name: covariance } }.
Args:
use_model_predictions: Whether to extract the Pareto frontier using
model predictions or directly observed values. If ``True``,
the metric means and covariances in this method's output will
also be based on model predictions and may differ from the
observed values.
Returns:
``None`` if it was not possible to extract the Pareto frontier,
otherwise a mapping from trial index to the tuple of:
- the parameterization of the arm in that trial,
- two-item tuple of metric means dictionary and covariance matrix
(model-predicted if ``use_model_predictions=True`` and observed
otherwise).
"""
if not not_none(self.experiment.optimization_config).is_moo_problem:
raise NotImplementedError( # pragma: no cover
"Please use `get_best_parameters` for single-objective problems."
)
return best_point_utils.get_pareto_optimal_parameters(
experiment=self.experiment,
generation_strategy=self.generation_strategy,
use_model_predictions=use_model_predictions,
)
[docs] def get_trials_data_frame(self) -> pd.DataFrame:
return exp_to_df(exp=self.experiment)
[docs] def get_max_parallelism(self) -> List[Tuple[int, int]]:
"""Retrieves maximum number of trials that can be scheduled in parallel
at different stages of optimization.
Some optimization algorithms profit significantly from sequential
optimization (i.e. suggest a few points, get updated with data for them,
repeat, see https://ax.dev/docs/bayesopt.html).
Parallelism setting indicates how many trials should be running simulteneously
(generated, but not yet completed with data).
The output of this method is mapping of form
{num_trials -> max_parallelism_setting}, where the max_parallelism_setting
is used for num_trials trials. If max_parallelism_setting is -1, as
many of the trials can be ran in parallel, as necessary. If num_trials
in a tuple is -1, then the corresponding max_parallelism_setting
should be used for all subsequent trials.
For example, if the returned list is [(5, -1), (12, 6), (-1, 3)],
the schedule could be: run 5 trials with any parallelism, run 6 trials in
parallel twice, run 3 trials in parallel for as long as needed. Here,
'running' a trial means obtaining a next trial from `AxClient` through
get_next_trials and completing it with data when available.
Returns:
Mapping of form {num_trials -> max_parallelism_setting}.
"""
parallelism_settings = []
for step in self.generation_strategy._steps:
parallelism_settings.append(
(step.num_trials, step.max_parallelism or step.num_trials)
)
return parallelism_settings
[docs] def get_optimization_trace(
self, objective_optimum: Optional[float] = None
) -> AxPlotConfig:
"""Retrieves the plot configuration for optimization trace, which shows
the evolution of the objective mean over iterations.
Args:
objective_optimum: Optimal objective, if known, for display in the
visualization.
"""
if not self.experiment.trials:
raise ValueError("Cannot generate plot as there are no trials.")
objective = self.objective
if isinstance(objective, MultiObjective):
raise UnsupportedError(
"`get_optimization_trace` is not supported "
"for multi-objective experiments"
)
objective_name = self.objective_name
best_objectives = np.array(
[
[
checked_cast(Trial, trial).objective_mean
for trial in self.experiment.trials.values()
if trial.status.is_completed
]
]
)
hover_labels = [
_format_dict(not_none(checked_cast(Trial, trial).arm).parameters)
for trial in self.experiment.trials.values()
if trial.status.is_completed
]
return optimization_trace_single_method(
y=(
np.minimum.accumulate(best_objectives, axis=1)
if objective.minimize
else np.maximum.accumulate(best_objectives, axis=1)
),
optimum=objective_optimum,
title="Model performance vs. # of iterations",
ylabel=objective_name.capitalize(),
hover_labels=hover_labels,
model_transitions=self.generation_strategy.model_transitions,
)
[docs] def get_contour_plot(
self,
param_x: Optional[str] = None,
param_y: Optional[str] = None,
metric_name: Optional[str] = None,
) -> AxPlotConfig:
"""Retrieves a plot configuration for a contour plot of the response
surface. For response surfaces with more than two parameters,
selected two parameters will appear on the axes, and remaining parameters
will be affixed to the middle of their range. If contour params arguments
are not provided, the first two parameters in the search space will be
used. If contour metrics are not provided, objective will be used.
Args:
param_x: name of parameters to use on x-axis for
the contour response surface plots.
param_y: name of parameters to use on y-axis for
the contour response surface plots.
metric_name: Name of the metric, for which to plot the response
surface.
"""
if not self.experiment.trials:
raise ValueError("Cannot generate plot as there are no trials.")
if len(self.experiment.parameters) < 2:
raise ValueError(
"Cannot create a contour plot as experiment has less than 2 "
"parameters, but a contour-related argument was provided."
)
if (param_x or param_y) and not (param_x and param_y):
raise ValueError(
"If `param_x` is provided, `param_y` is "
"required as well, and vice-versa."
)
if not metric_name:
if isinstance(self.objective, MultiObjective):
raise UnsupportedError(
"`get_contour_plot` requires a `metric_name` "
"for multi-objective experiments"
)
metric_name = self.objective_name
if not param_x or not param_y:
parameter_names = list(self.experiment.parameters.keys())
param_x = parameter_names[0]
param_y = parameter_names[1]
if param_x not in self.experiment.parameters:
raise ValueError(
f'Parameter "{param_x}" not found in the optimization search space.'
)
if param_y not in self.experiment.parameters:
raise ValueError(
f'Parameter "{param_y}" not found in the optimization search space.'
)
if metric_name not in self.experiment.metrics:
raise ValueError(
f'Metric "{metric_name}" is not associated with this optimization.'
)
if self.generation_strategy.model is not None:
try:
logger.info(
f"Retrieving contour plot with parameter '{param_x}' on X-axis "
f"and '{param_y}' on Y-axis, for metric '{metric_name}'. "
"Remaining parameters are affixed to the middle of their range."
)
return plot_contour(
model=not_none(self.generation_strategy.model),
param_x=param_x,
param_y=param_y,
metric_name=metric_name,
)
except NotImplementedError:
# Some models don't implement '_predict', which is needed
# for the contour plots.
logger.info(
f"Model {self.generation_strategy.model} does not implement "
"`predict`, so it cannot be used to generate a response "
"surface plot."
)
raise UnsupportedPlotError(
f'Could not obtain contour plot of "{metric_name}" for parameters '
f'"{param_x}" and "{param_y}", as a model with predictive ability, '
"such as a Gaussian Process, has not yet been trained in the course "
"of this optimization."
)
[docs] def get_feature_importances(self, relative: bool = True) -> AxPlotConfig:
"""
Get a bar chart showing feature_importances for a metric.
A drop-down controls the metric for which the importances are displayed.
Args:
relative: Whether the values are displayed as percentiles or
as raw importance metrics.
"""
if not self.experiment.trials:
raise ValueError("Cannot generate plot as there are no trials.")
cur_model = self.generation_strategy.model
if cur_model is not None:
try:
return plot_feature_importance_by_feature(cur_model, relative=relative)
except NotImplementedError:
logger.info(
f"Model {self.generation_strategy.model} does not implement "
"`feature_importances`, so it cannot be used to generate "
"this plot. Only certain models, specifically GPEI, implement "
"feature importances."
)
raise ValueError(
"Could not obtain feature_importances for any metrics "
" as a model that can produce feature importances, such as a "
"Gaussian Process, has not yet been trained in the course "
"of this optimization."
)
[docs] def load_experiment_from_database(
self,
experiment_name: str,
choose_generation_strategy_kwargs: Optional[Dict[str, Any]] = None,
) -> None:
"""Load an existing experiment from database using the `DBSettings`
passed to this `AxClient` on instantiation.
Args:
experiment_name: Name of the experiment.
Returns:
Experiment object.
"""
experiment, generation_strategy = self._load_experiment_and_generation_strategy(
experiment_name=experiment_name
)
if experiment is None: # pragma: no cover
raise ValueError(f"Experiment by name '{experiment_name}' not found.")
self._experiment = experiment
logger.info(f"Loaded {experiment}.")
if generation_strategy is None: # pragma: no cover
self._set_generation_strategy(
choose_generation_strategy_kwargs=choose_generation_strategy_kwargs
)
self._save_generation_strategy_to_db_if_possible(
generation_strategy=self.generation_strategy,
)
else:
self._generation_strategy = generation_strategy
logger.info(
f"Using generation strategy associated with the loaded experiment:"
f" {generation_strategy}."
)
[docs] def get_model_predictions(
self, metric_names: Optional[List[str]] = None
) -> Dict[int, Dict[str, Tuple[float, float]]]:
"""Retrieve model-estimated means and covariances for all metrics.
Note: this function retrieves the predictions for the 'in-sample' arms,
which means that the return mapping on this function will only contain
predictions for trials that have been completed with data.
Args:
metric_names: Names of the metrics, for which to retrieve predictions.
All metrics on experiment will be retrieved if this argument was
not specified.
Returns:
A mapping from trial index to a mapping of metric names to tuples
of predicted metric mean and SEM, of form:
{ trial_index -> { metric_name: ( mean, SEM ) } }.
"""
if self.generation_strategy.model is None: # pragma: no cover
raise ValueError("No model has been instantiated yet.")
if metric_names is None and self.experiment.metrics is None:
raise ValueError( # pragma: no cover
"No metrics to retrieve specified on the experiment or as "
"argument to `get_model_predictions`."
)
arm_info, _, _ = _get_in_sample_arms(
model=not_none(self.generation_strategy.model),
metric_names=set(metric_names)
if metric_names is not None
else set(not_none(self.experiment.metrics).keys()),
)
trials = checked_cast_dict(int, Trial, self.experiment.trials)
return {
trial_index: {
m: (
arm_info[not_none(trials[trial_index].arm).name].y_hat[m],
arm_info[not_none(trials[trial_index].arm).name].se_hat[m],
)
for m in arm_info[not_none(trials[trial_index].arm).name].y_hat
}
for trial_index in trials
if not_none(trials[trial_index].arm).name in arm_info
}
[docs] def verify_trial_parameterization(
self, trial_index: int, parameterization: TParameterization
) -> bool:
"""Whether the given parameterization matches that of the arm in the trial
specified in the trial index.
"""
return (
not_none(self._get_trial(trial_index=trial_index).arm).parameters
== parameterization
)
# ------------------ JSON serialization & storage methods. -----------------
[docs] def save_to_json_file(self, filepath: str = "ax_client_snapshot.json") -> None:
"""Save a JSON-serialized snapshot of this `AxClient`'s settings and state
to a .json file by the given path.
"""
with open(filepath, "w+") as file: # pragma: no cover
file.write(json.dumps(self.to_json_snapshot()))
logger.info(f"Saved JSON-serialized state of optimization to `{filepath}`.")
[docs] @classmethod
def load_from_json_file(
cls: Type[AxClientSubclass], filepath: str = "ax_client_snapshot.json", **kwargs
) -> AxClientSubclass:
"""Restore an `AxClient` and its state from a JSON-serialized snapshot,
residing in a .json file by the given path.
"""
with open(filepath, "r") as file: # pragma: no cover
serialized = json.loads(file.read())
return cls.from_json_snapshot(serialized=serialized, **kwargs)
[docs] def to_json_snapshot(self) -> Dict[str, Any]:
"""Serialize this `AxClient` to JSON to be able to interrupt and restart
optimization and save it to file by the provided path.
Returns:
A JSON-safe dict representation of this `AxClient`.
"""
return {
"_type": self.__class__.__name__,
"experiment": object_to_json(self._experiment),
"generation_strategy": object_to_json(self._generation_strategy),
"_enforce_sequential_optimization": self._enforce_sequential_optimization,
}
[docs] @classmethod
def from_json_snapshot(
cls: Type[AxClientSubclass], serialized: Dict[str, Any], **kwargs
) -> AxClientSubclass:
"""Recreate an `AxClient` from a JSON snapshot."""
experiment = object_from_json(serialized.pop("experiment"))
serialized_generation_strategy = serialized.pop("generation_strategy")
ax_client = cls(
generation_strategy=generation_strategy_from_json(
generation_strategy_json=serialized_generation_strategy
)
if serialized_generation_strategy is not None
else None,
enforce_sequential_optimization=serialized.pop(
"_enforce_sequential_optimization"
),
**kwargs,
)
ax_client._experiment = experiment
return ax_client
# ---------------------- Private helper methods. ---------------------
@property
def experiment(self) -> Experiment:
"""Returns the experiment set on this Ax client."""
if self._experiment is None:
raise ValueError(
"Experiment not set on Ax client. Must first "
"call load_experiment or create_experiment to use handler functions."
)
return not_none(self._experiment)
@property
def generation_strategy(self) -> GenerationStrategy:
"""Returns the generation strategy, set on this experiment."""
if self._generation_strategy is None:
raise ValueError(
"No generation strategy has been set on this optimization yet."
)
return not_none(self._generation_strategy)
@property
def objective(self) -> Objective:
return not_none(self.experiment.optimization_config).objective
@property
def objective_name(self) -> str:
"""Returns the name of the objective in this optimization."""
objective = self.objective
if isinstance(objective, MultiObjective):
raise UnsupportedError(
"Multi-objective experiments contain multiple objectives"
)
return objective.metric.name
@property
def objective_names(self) -> List[str]:
"""Returns the name of the objective in this optimization."""
objective = self.objective
return [m.name for m in objective.metrics]
def _update_trial_with_raw_data(
self,
trial_index: int,
raw_data: TEvaluationOutcome,
metadata: Optional[Dict[str, Union[str, int]]] = None,
sample_size: Optional[int] = None,
complete_trial: bool = False,
combine_with_last_data: bool = False,
) -> str:
"""Helper method attaches data to a trial, returns a str of update."""
# Format the data to save.
trial = self._get_trial(trial_index=trial_index)
sample_sizes = {not_none(trial.arm).name: sample_size} if sample_size else {}
evaluations, data = self._make_evaluations_and_data(
trial=trial, raw_data=raw_data, metadata=metadata, sample_sizes=sample_sizes
)
metadata = metadata or {}
self._validate_trial_data(trial=trial, data=data)
trial.update_run_metadata(metadata=metadata)
self.experiment.attach_data(
data=data, combine_with_last_data=combine_with_last_data
)
if complete_trial:
trial.mark_completed()
self._save_or_update_trial_in_db_if_possible(
experiment=self.experiment,
trial=trial,
)
return str(
_round_floats_for_logging(item=evaluations[next(iter(evaluations.keys()))])
)
def _set_experiment(
self,
experiment: Experiment,
overwrite_existing_experiment: bool = False,
) -> None:
"""Sets the ``_experiment`` attribute on this `AxClient`` instance and saves the
experiment if this instance uses SQL storage.
NOTE: This setter **should not be used outside of this file in production**.
It can be leveraged in development, but all checked-in code that uses the
Service API should leverage ``AxClient.create_experiment`` instead and extend it
as needed. If using ``create_experiment`` is impossible and this setter is
required, please raise your use case to the AE team or on our Github.
"""
name = experiment._name
if self.db_settings_set and not name:
raise ValueError( # pragma: no cover
"Must give the experiment a name if `db_settings` is not None."
)
if self.db_settings_set:
experiment_id, _ = self._get_experiment_and_generation_strategy_db_id(
experiment_name=not_none(name)
)
if experiment_id:
raise ValueError(
f"Experiment {name} already exists in the database. "
"To protect experiments that are running in production, "
"overwriting stored experiments is not allowed. To "
"start a new experiment and store it, change the "
"experiment's name."
)
if self._experiment is not None:
if overwrite_existing_experiment:
exp_name = self.experiment._name or "untitled"
new_exp_name = name or "untitled"
logger.info(
f"Overwriting existing experiment ({exp_name}) on this client "
f"with new experiment ({new_exp_name}) and restarting the "
"generation strategy."
)
self._generation_strategy = None
else:
raise ValueError(
"Experiment already created for this client instance. "
"Set the `overwrite_existing_experiment` to `True` to overwrite "
"with new experiment."
)
self._experiment = experiment
try:
self._save_experiment_to_db_if_possible(
experiment=self.experiment,
)
except Exception:
# Unset the experiment on this `AxClient` instance if encountered and
# raising an error from saving the experiment, to avoid a case where
# overall `create_experiment` call fails with a storage error, but
# `self._experiment` is still set and user has to specify the
# `overwrite_existing_experiment` kwarg to re-attempt exp. creation.
self._experiment = None
raise
def _set_generation_strategy(
self, choose_generation_strategy_kwargs: Optional[Dict[str, Any]] = None
) -> None:
"""Selects the generation strategy and applies specified dispatch kwargs,
if any.
"""
choose_generation_strategy_kwargs = choose_generation_strategy_kwargs or {}
random_seed = choose_generation_strategy_kwargs.pop(
"random_seed", self._random_seed
)
enforce_sequential_optimization = choose_generation_strategy_kwargs.pop(
"enforce_sequential_optimization", self._enforce_sequential_optimization
)
if self._generation_strategy is None:
self._generation_strategy = choose_generation_strategy(
search_space=self.experiment.search_space,
optimization_config=self.experiment.optimization_config,
enforce_sequential_optimization=enforce_sequential_optimization,
random_seed=random_seed,
experiment=self.experiment,
**choose_generation_strategy_kwargs,
)
def _gen_new_generator_run(self, n: int = 1) -> GeneratorRun:
"""Generate new generator run for this experiment.
Args:
n: Number of arms to generate.
"""
# If random seed is not set for this optimization, context manager does
# nothing; otherwise, it sets the random seed for torch, but only for the
# scope of this call. This is important because torch seed is set globally,
# so if we just set the seed without the context manager, it can have
# serious negative impact on the performance of the models that employ
# stochasticity.
with manual_seed(seed=self._random_seed) and warnings.catch_warnings():
# Filter out GPYTorch warnings to avoid confusing users.
warnings.simplefilter("ignore")
return not_none(self.generation_strategy).gen(
experiment=self.experiment,
n=n,
pending_observations=self._get_pending_observation_features(
experiment=self.experiment
),
)
def _get_trial(self, trial_index: int) -> Trial:
"""Gets trial by given index or raises an error if it does not exist."""
if trial_index in self.experiment.trials:
trial = self.experiment.trials.get(trial_index)
if not isinstance(trial, Trial):
raise NotImplementedError(
"`AxClient` only supports `Trial`, not `BatchTrial`."
)
return trial
raise ValueError(f"Trial {trial_index} does not yet exist.")
def _find_last_trial_with_parameterization(
self, parameterization: TParameterization
) -> int:
"""Given a parameterization, find the last trial in the experiment that
contains an arm with that parameterization.
"""
for trial_idx in sorted(self.experiment.trials.keys(), reverse=True):
if not_none(self._get_trial(trial_idx).arm).parameters == parameterization:
return trial_idx
raise ValueError(
f"No trial on experiment matches parameterization {parameterization}."
)
@classmethod
def _get_pending_observation_features(
cls, experiment
) -> Optional[Dict[str, List[ObservationFeatures]]]:
"""Extract pending points for the given experiment.
NOTE: With one-arm `Trial`-s, we use a more performant
``get_pending_observation_features_based_on_trial_status`` utility instead
of ``get_pending_observation_features``, since we can determine whether a point
is pending based on the status of the corresponding trial.
"""
return get_pending_observation_features_based_on_trial_status(
experiment=experiment
)
@classmethod
def _raw_data_by_arm(
cls,
trial: BaseTrial,
raw_data: Union[TEvaluationOutcome, Dict[str, TEvaluationOutcome]],
) -> Dict[str, TEvaluationOutcome]:
raw_data_by_arm: Dict[str, TEvaluationOutcome]
if isinstance(trial, BatchTrial): # pragma: no cover
raw_data_by_arm = checked_cast_complex(
Dict[str, TEvaluationOutcome],
raw_data,
message=cls.BACH_TRIAL_RAW_DATA_FORMAT_ERROR_MESSAGE,
)
elif isinstance(trial, Trial):
arm_name = not_none(trial.arm).name
raw_data_by_arm = {
arm_name: checked_cast_complex(
TEvaluationOutcome,
raw_data,
message=cls.TRIAL_RAW_DATA_FORMAT_ERROR_MESSAGE,
)
}
else: # pragma: no cover
raise ValueError(f"Unexpected trial type: {type(trial)}.")
not_trial_arm_names = set(raw_data_by_arm.keys()) - set(
trial.arms_by_name.keys()
)
if not_trial_arm_names:
raise ValueError( # pragma: no cover
f"Arms {not_trial_arm_names} are not part of trial #{trial.index}."
)
return raw_data_by_arm
def _make_evaluations_and_data(
self,
trial: BaseTrial,
raw_data: Union[TEvaluationOutcome, Dict[str, TEvaluationOutcome]],
metadata: Optional[Dict[str, Union[str, int]]],
sample_sizes: Optional[Dict[str, int]] = None,
) -> Tuple[Dict[str, TEvaluationOutcome], AbstractDataFrameData]:
"""Formats given raw data as Ax evaluations and `Data`.
Args:
trial: Trial within the experiment.
raw_data: Metric outcomes for 1-arm trials, map from arm name to
metric outcomes for batched trials.
sample_size: Integer sample size for 1-arm trials, dict from arm
name to sample size for batched trials. Optional.
metadata: Additional metadata to track about this run.
data_is_for_batched_trials: Whether making evaluations and data for
a batched trial or a 1-arm trial.
"""
raw_data_by_arm = self._raw_data_by_arm(trial=trial, raw_data=raw_data)
evaluations, data = data_and_evaluations_from_raw_data(
raw_data=raw_data_by_arm,
metric_names=self.objective_names,
trial_index=trial.index,
sample_sizes=sample_sizes or {},
start_time=(
checked_cast_optional(int, metadata.get("start_time"))
if metadata is not None
else None
),
end_time=(
checked_cast_optional(int, metadata.get("end_time"))
if metadata is not None
else None
),
)
return evaluations, data
# ------------------------------ Validators. -------------------------------
@staticmethod
def _validate_can_complete_trial(trial: BaseTrial) -> None:
if trial.status.is_completed:
raise ValueError(
f"Trial {trial.index} has already been completed with data."
"To add more data to it (for example, for a different metric), "
"use `ax_client.update_trial_data`."
)
if trial.status.is_abandoned or trial.status.is_failed:
raise ValueError(
f"Trial {trial.index} has been marked {trial.status.name}, so it "
"no longer expects data."
)
def _validate_search_space_membership(self, parameters: TParameterization) -> None:
self.experiment.search_space.check_membership(
parameterization=parameters, raise_error=True
)
# `check_membership` uses int and float interchangeably, which we don't
# want here.
for p_name, parameter in self.experiment.search_space.parameters.items():
if not isinstance(parameters[p_name], parameter.python_type):
typ = type(parameters[p_name])
raise ValueError(
f"Value for parameter {p_name} is of type {typ}, expected "
f"{parameter.python_type}. If the intention was to have the "
f"parameter on experiment be of type {typ}, set `value_type` "
f"on experiment creation for {p_name}."
)
def _validate_trial_data(self, trial: Trial, data: AbstractDataFrameData) -> None:
for metric_name in data.df["metric_name"].values:
if metric_name not in self.experiment.metrics:
logger.info(
f"Data was logged for metric {metric_name} that was not yet "
"tracked on the experiment. Please specify `tracking_metric_"
"names` argument in AxClient.create_experiment to add tracking "
"metrics to the experiment. Without those, all data users "
"specify is still attached to the experiment, but will not be "
"fetched in `experiment.fetch_data()`, but you can still use "
"`experiment.lookup_data_for_trial` to get all attached data."
)
# -------- Backward-compatibility with old save / load method names. -------
[docs] @staticmethod
def get_recommended_max_parallelism() -> None:
raise NotImplementedError(
"Use `get_max_parallelism` instead; parallelism levels are now "
"enforced in generation strategy, so max parallelism is no longer "
"just recommended."
)
[docs] @staticmethod
def load_experiment(experiment_name: str) -> None:
raise NotImplementedError(
"Use `load_experiment_from_database` to load from SQL database or "
"`load_from_json_file` to load optimization state from .json file."
)
[docs] @staticmethod
def load(filepath: Optional[str] = None) -> None:
raise NotImplementedError(
"Use `load_experiment_from_database` to load from SQL database or "
"`load_from_json_file` to load optimization state from .json file."
)
[docs] @staticmethod
def save(filepath: Optional[str] = None) -> None:
raise NotImplementedError(
"Use `save_to_json_file` to save optimization state to .json file."
)
