#!/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.
"""
Module for benchmarking Ax algorithms.
Key terms used:
* Trial –– usual Ax `Trial` or `BatchTral`, one execution of a given arm or
group of arms.
* Replication –– one run of an optimization loop; 1 method + problem combination.
* Test –– multiple replications, ran for statistical significance.
* Full run –– multiple tests: run all methods with all problems.
* Method –– (one of) the algorithm(s) being benchmarked.
* Problem –– a synthetic function, a surrogate surface, or an ML model, on which
to assess the performance of algorithms.
"""
import time
from types import FunctionType
from typing import Callable, Dict, List, Optional, Tuple, Union
import numpy as np
from ax.benchmark import utils
from ax.benchmark.benchmark_problem import BenchmarkProblem, SimpleBenchmarkProblem
from ax.core.data import Data
from ax.core.experiment import Experiment
from ax.core.generator_run import GeneratorRun
from ax.core.observation import ObservationFeatures
from ax.core.parameter import RangeParameter
from ax.modelbridge.base import gen_arms
from ax.modelbridge.generation_strategy import GenerationStrategy
from ax.runners.synthetic import SyntheticRunner
from ax.service.ax_client import AxClient
from ax.utils.common.logger import get_logger
from ax.utils.common.typeutils import not_none
from ax.utils.measurement.synthetic_functions import SyntheticFunction
logger = get_logger(__name__)
# To bypass catching of exceptions during benchmarking, since all other exceptions
# will be caught and recorded, but will not necessarily terminate the benchmarking
# run.
[docs]class NonRetryableBenchmarkingError(ValueError):
"""Error that indicates an issue with the benchmarking setup (e.g. unexpected
problem setup, a benchmarking function called incorrectly, etc.) –– something
that prevents the benchmarking suite itself from running, rather than an error
that occurs during the runs of the benchmarking trials, replications, or tests.
"""
pass
[docs]def benchmark_trial(
parameterization: Optional[np.ndarray] = None,
evaluation_function: Optional[Union[SyntheticFunction, FunctionType]] = None,
experiment: Optional[Experiment] = None,
trial_index: Optional[int] = None,
) -> Union[Tuple[float, float], Data]: # Mean and SEM or a Data object.
"""Evaluates one trial from benchmarking replication (an Ax trial or batched
trial). Evaluation requires either the `parameterization` and `evalution_
function` parameters or the `experiment` and `trial_index` parameters.
Note: evaluation function relies on the ordering of items in the
parameterization nd-array.
Args:
parameterization: The parameterization to evaluate.
evaluation_function: The evaluation function for the benchmark objective.
experiment: Experiment, for a trial on which to fetch data.
trial_index: Index of the trial, for which to fetch data.
"""
use_Service_API = parameterization is not None and evaluation_function is not None
use_Dev_API = experiment is not None and trial_index is not None
if not use_Service_API ^ use_Dev_API:
raise NonRetryableBenchmarkingError( # TODO[T53975770]: test
"A parameterization and an evaluation function required for Service-"
"API-style trial evaluation and an experiment and trial index are "
"required for Dev API trial evalution via fetching metric data."
)
if use_Service_API:
sem = 0.0 if isinstance(evaluation_function, SyntheticFunction) else None
# pyre-fixme[7]: Expected `Union[Tuple[float, float], Data]` but got
# `Tuple[typing.Any, Optional[float]]`.
return evaluation_function(parameterization), sem # pyre-ignore[29]: call err.
else:
trial_index = not_none(trial_index)
return not_none(not_none(experiment).trials.get(trial_index)).fetch_data()
[docs]def benchmark_replication( # One optimization loop.
problem: BenchmarkProblem,
method: GenerationStrategy,
num_trials: int,
replication_index: Optional[int] = None,
batch_size: int = 1,
raise_all_exceptions: bool = False,
benchmark_trial: FunctionType = benchmark_trial,
verbose_logging: bool = True,
# Number of trials that need to fail for a replication to be considered failed.
failed_trials_tolerated: int = 5,
) -> Experiment:
"""Runs one benchmarking replication (equivalent to one optimization loop).
Args:
problem: Problem to benchmark on.
method: Method to benchmark, represented as generation strategies.
num_trials: Number of trials in each test experiment.
batch_size: Batch size for this replication, defaults to 1.
raise_all_exceptions: If set to True, any encountered exception will be
raised; alternatively, failure tolerance thresholds are used and a few
number of trials `failed_trials_tolerated` can fail before a replication
is considered failed.
benchmark_trial: Function that runs a single trial. Defaults
to `benchmark_trial` in this module and must have the same signature.
verbose_logging: Whether logging level should be set to `INFO`.
failed_trials_tolerated: How many trials can fail before a replication is
considered failed and aborted. Defaults to 5.
"""
trial_exceptions = []
experiment_name = f"{method.name}_on_{problem.name}"
if replication_index is not None:
experiment_name += f"__v{replication_index}"
# Make sure the generation strategy starts from the beginning.
method = method.clone_reset()
# Choose whether to run replication via Service or Developer API, based on
# whether the problem was set up using Ax classes like `SearchSpace` and
# `OptimizationConfig` or using "RESTful" Service API-like constructs like
# dict parameter representations and `SyntheticFunction`-s or custom callables
# for evaluation function.
replication_runner = (
_benchmark_replication_Service_API
if isinstance(problem, SimpleBenchmarkProblem)
else _benchmark_replication_Dev_API
)
experiment, exceptions = replication_runner(
problem=problem, # pyre-ignore[6]
method=method,
num_trials=num_trials,
experiment_name=experiment_name,
batch_size=batch_size,
raise_all_exceptions=raise_all_exceptions,
benchmark_trial=benchmark_trial,
verbose_logging=verbose_logging,
failed_trials_tolerated=failed_trials_tolerated,
)
trial_exceptions.extend(exceptions)
return experiment
[docs]def benchmark_test( # One test, multiple replications.
problem: BenchmarkProblem,
method: GenerationStrategy,
num_trials: int,
num_replications: int = 20,
batch_size: int = 1,
raise_all_exceptions: bool = False,
benchmark_replication: FunctionType = benchmark_replication,
benchmark_trial: FunctionType = benchmark_trial,
verbose_logging: bool = True,
# Number of trials that need to fail for a replication to be considered failed.
failed_trials_tolerated: int = 5,
# Number of replications that need to fail for a test to be considered failed.
failed_replications_tolerated: int = 3,
) -> List[Experiment]:
"""Runs one benchmarking test (equivalent to one problem-method combination),
translates into `num_replication` replications, ran for statistical
significance of the results.
Args:
problem: Problem to benchmark on.
method: Method to benchmark, represented as generation strategies.
num_replications: Number of times to run each test (each problem-method
combination), for an aggregated result.
num_trials: Number of trials in each test experiment, defaults to 20.
batch_size: Batch size for this test, defaults to 1.
raise_all_exceptions: If set to True, any encountered exception will be
raised; alternatively, failure tolerance thresholds are used and a few
number of trials `failed_trials_tolerated` can fail before a replication
is considered failed, as well some replications
`failed_replications_tolerated` can fail before a benchmarking test
is considered failed.
benchmark_replication: Function that runs a single benchmarking replication.
Defaults to `benchmark_replication` in this module and must have the
same signature.
benchmark_trial: Function that runs a single trial. Defaults
to `benchmark_trial` in this module and must have the same signature.
verbose_logging: Whether logging level should be set to `INFO`.
failed_trials_tolerated: How many trials can fail before a replication is
considered failed and aborted. Defaults to 5.
failed_replications_tolerated: How many replications can fail before a
test is considered failed and aborted. Defaults to 3.
"""
replication_exceptions = []
test_replications = []
for replication_idx in range(num_replications):
try:
test_replications.append(
benchmark_replication(
problem=problem,
method=method,
replication_index=replication_idx,
num_trials=num_trials,
batch_size=batch_size,
raise_all_exceptions=raise_all_exceptions,
verbose_logging=verbose_logging,
failed_trials_tolerated=failed_trials_tolerated,
)
)
except Exception as err:
if raise_all_exceptions:
raise
replication_exceptions.append(err) # TODO[T53975770]: test
if len(replication_exceptions) > failed_replications_tolerated:
raise RuntimeError( # TODO[T53975770]: test
f"More than {failed_replications_tolerated} failed for "
"{method.name}_on_{problem.name}."
)
return test_replications
[docs]def full_benchmark_run( # Full run, multiple tests.
problem_groups: (
Optional[Dict[str, Union[List[BenchmarkProblem], List[str]]]]
) = None,
method_groups: (
Optional[Dict[str, Union[List[GenerationStrategy], List[str]]]]
) = None,
num_trials: Union[int, List[List[int]]] = 20,
num_replications: int = 20,
batch_size: Union[int, List[List[int]]] = 1,
raise_all_exceptions: bool = False,
benchmark_test: FunctionType = benchmark_test,
benchmark_replication: FunctionType = benchmark_replication,
benchmark_trial: FunctionType = benchmark_trial,
verbose_logging: bool = True,
# Number of trials that need to fail for a replication to be considered failed.
failed_trials_tolerated: int = 5,
# Number of replications that need to fail for a test to be considered failed.
failed_replications_tolerated: int = 3,
) -> Dict[str, Dict[str, List[Experiment]]]:
"""Full run of the benchmarking suite. To make benchmarking distrubuted at
a level of a test, a replication, or a trial (or any combination of those),
by passing in a wrapped (in some scheduling logic) version of a corresponding
function from this module.
Here, `problem_groups` and `method_groups` are dictionaries that have the same
keys such that we can run a specific subset of problems with a corresponding
subset of methods.
Example:
::
problem_groups = {
"single_fidelity": [ackley, branin],
"multi_fidelity": [augmented_hartmann],
}
method_groups = {
"single_fidelity": [single_task_GP_and_NEI_strategy],
"multi_fidelity": [fixed_noise_MFGP_and_MFKG_strategy],
}
Here, `ackley` and `branin` will be run against `single_task_GP_and_NEI_strategy`
and `augmented_hartmann` against `fixed_noise_MFGP_and_MFKG_strategy`.
Args:
problem_groups: Problems to benchmark on, represented as a dictionary from
category string to List of BenchmarkProblem-s or string keys (must be
in standard BOProblems). More on `problem_groups` below.
method_groups: Methods to benchmark on, represented as a dictionary from
category string to List of generation strategies or string keys (must
be in standard BOMethods). More on `method_groups` below.
num_replications: Number of times to run each test (each problem-method
combination), for an aggregated result.
num_trials: Number of trials in each test experiment.
raise_all_exceptions: If set to True, any encountered exception will be
raised; alternatively, failure tolerance thresholds are used and a few
number of trials `failed_trials_tolerated` can fail before a replication
is considered failed, as well some replications
`failed_replications_tolerated` can fail before a benchmarking test
is considered failed.
benchmark_test: Function that runs a single benchmarking test. Defaults
to `benchmark_test` in this module and must have the same signature.
benchmark_replication: Function that runs a single benchmarking replication.
Defaults to `benchmark_replication` in this module and must have the
same signature.
benchmark_trial: Function that runs a single trial. Defaults
to `benchmark_trial` in this module and must have the same signature.
verbose_logging: Whether logging level should be set to `INFO`.
failed_trials_tolerated: How many trials can fail before a replication is
considered failed and aborted. Defaults to 5.
failed_replications_tolerated: How many replications can fail before a
test is considered failed and aborted. Defaults to 3.
"""
problem_groups = problem_groups or {}
method_groups = method_groups or {}
_validate_groups(problem_groups, method_groups)
exceptions = []
tests: Dict[str, Dict[str, List[Experiment]]] = {}
for group_name in problem_groups:
problems, methods = utils.get_problems_and_methods(
problems=problem_groups.get(group_name),
methods=method_groups.get(group_name),
)
for problem_idx, problem in enumerate(problems):
tests[problem.name] = {}
for method_idx, method in enumerate(methods):
tests[problem.name][method.name] = []
try:
tests[problem.name][method.name] = benchmark_test(
problem=problem,
method=method,
num_replications=num_replications,
# For arguments passed as either numbers, or matrices,
# xtract corresponding values for the given combination.
num_trials=utils.get_corresponding(
num_trials, problem_idx, method_idx
),
batch_size=utils.get_corresponding(
batch_size, problem_idx, method_idx
),
benchmark_replication=benchmark_replication,
benchmark_trial=benchmark_trial,
raise_all_exceptions=raise_all_exceptions,
verbose_logging=verbose_logging,
failed_replications_tolerated=failed_replications_tolerated,
failed_trials_tolerated=failed_trials_tolerated,
)
except Exception as err:
if raise_all_exceptions:
raise
exceptions.append(err) # TODO[T53975770]: test
logger.info(f"Obtained benchmarking test experiments: {tests}")
return tests
def _benchmark_replication_Service_API(
problem: SimpleBenchmarkProblem,
method: GenerationStrategy,
num_trials: int,
experiment_name: str,
batch_size: int = 1,
raise_all_exceptions: bool = False,
benchmark_trial: FunctionType = benchmark_trial,
verbose_logging: bool = True,
# Number of trials that need to fail for a replication to be considered failed.
failed_trials_tolerated: int = 5,
) -> Tuple[Experiment, List[Exception]]:
"""Run a benchmark replication via the Service API because the problem was
set up in a simplified way, without the use of Ax classes like `OptimizationConfig`
or `SearchSpace`.
"""
exceptions = []
if batch_size == 1:
ax_client = AxClient(
generation_strategy=method, verbose_logging=verbose_logging
)
else: # pragma: no cover, TODO[T53975770]
assert batch_size > 1, "Batch size of 1 or greater is expected."
raise NotImplementedError(
"Batched benchmarking on `SimpleBenchmarkProblem`-s not yet implemented."
)
ax_client.create_experiment(
name=experiment_name,
parameters=problem.domain_as_ax_client_parameters(),
minimize=problem.minimize,
objective_name=problem.name,
)
parameter_names = list(ax_client.experiment.search_space.parameters.keys())
assert num_trials > 0
for _ in range(num_trials):
parameterization, idx = ax_client.get_next_trial()
param_values = np.array([parameterization.get(x) for x in parameter_names])
try:
mean, sem = benchmark_trial(
parameterization=param_values, evaluation_function=problem.f
)
# If problem indicates a noise level and is using a synthetic callable,
# add normal noise to the measurement of the mean.
if problem.uses_synthetic_function and problem.noise_sd != 0.0:
noise = np.random.randn() * problem.noise_sd
sem = (sem or 0.0) + problem.noise_sd
logger.info(
f"Adding noise of {noise} to the measurement mean ({mean})."
f"Problem noise SD setting: {problem.noise_sd}."
)
mean = mean + noise
ax_client.complete_trial(trial_index=idx, raw_data=(mean, sem))
except Exception as err: # TODO[T53975770]: test
if raise_all_exceptions:
raise
exceptions.append(err)
if len(exceptions) > failed_trials_tolerated:
raise RuntimeError( # TODO[T53975770]: test
f"More than {failed_trials_tolerated} failed for {experiment_name}."
)
return ax_client.experiment, exceptions
def _benchmark_replication_Dev_API(
problem: BenchmarkProblem,
method: GenerationStrategy,
num_trials: int,
experiment_name: str,
batch_size: int = 1,
raise_all_exceptions: bool = False,
benchmark_trial: FunctionType = benchmark_trial,
verbose_logging: bool = True,
# Number of trials that need to fail for a replication to be considered failed.
failed_trials_tolerated: int = 5,
) -> Tuple[Experiment, List[Exception]]:
"""Run a benchmark replication via the Developer API because the problem was
set up with Ax classes (likely to allow for additional complexity like
adding constraints or non-range parameters).
"""
exceptions = []
experiment = Experiment(
name=experiment_name,
search_space=problem.search_space,
optimization_config=problem.optimization_config,
runner=SyntheticRunner(),
)
for trial_index in range(num_trials):
try:
gr = method.gen(experiment=experiment, n=batch_size)
if batch_size == 1:
trial = experiment.new_trial(generator_run=gr)
else:
assert batch_size > 1
trial = experiment.new_batch_trial(generator_run=gr)
trial.run()
benchmark_trial(experiment=experiment, trial_index=trial_index)
trial.mark_completed()
except Exception as err: # TODO[T53975770]: test
if raise_all_exceptions:
raise
exceptions.append(err)
if len(exceptions) > failed_trials_tolerated:
raise RuntimeError( # TODO[T53975770]: test
f"More than {failed_trials_tolerated} failed for {experiment_name}."
)
return experiment, exceptions
[docs]def benchmark_minimize_callable(
problem: BenchmarkProblem,
num_trials: int,
method_name: str,
replication_index: Optional[int] = None,
) -> Tuple[Experiment, Callable[[List[float]], float]]:
"""
An interface for evaluating external methods on Ax benchmark problems. The
arms run and performance will be tracked by Ax, so the external method can
be evaluated alongside Ax methods.
It is designed around methods that implement an interface like
scipy.optimize.minimize. This function will return a callable evaluation
function that takes in an array of parameter values and returns a float
objective value. The evaluation function should always be minimized: if the
benchmark problem is a maximization problem, then the value returned by
the evaluation function will be negated so it can be used directly by
methods that minimize. This callable can be given to an external
minimization function, and Ax will track all of the calls made to it and
the arms that were evaluated.
This will also return an Experiment object that will track the arms
evaluated by the external method in the same way as done for Ax
internal benchmarks. This function should thus be used for each benchmark
replication.
Args:
problem: The Ax benchmark problem to be used to construct the
evalutaion function.
num_trials: The maximum number of trials for a benchmark run.
method_name: Name of the method being tested.
replication_index: Replicate number, if multiple replicates are being
run.
"""
# Some validation
if isinstance(problem, SimpleBenchmarkProblem):
raise NonRetryableBenchmarkingError("`SimpleBenchmarkProblem` not supported.")
if not all(
isinstance(p, RangeParameter) for p in problem.search_space.parameters.values()
):
raise NonRetryableBenchmarkingError("Only continuous search spaces supported.")
if any(
p.log_scale for p in problem.search_space.parameters.values() # pyre-ignore
):
raise NonRetryableBenchmarkingError("Log-scale parameters not supported.")
# Create Ax experiment
experiment_name = f"{method_name}_on_{problem.name}"
if replication_index is not None:
experiment_name += f"__v{replication_index}"
experiment = Experiment(
name=experiment_name,
search_space=problem.search_space,
optimization_config=problem.optimization_config,
runner=SyntheticRunner(),
)
max_trials = num_trials # to be used below
# Construct the evaluation function
def evaluation_function(x: List[float]) -> float:
# Check if we have exhuasted the evaluation budget
if len(experiment.trials) >= max_trials:
raise ValueError(f"Evaluation budget ({max_trials} trials) exhuasted.")
# Create an ObservationFeatures
param_dict = {
pname: x[i]
for i, pname in enumerate(problem.search_space.parameters.keys())
}
obsf = ObservationFeatures(parameters=param_dict) # pyre-ignore
# Get the time since last call
num_trials = len(experiment.trials)
if num_trials == 0:
gen_time = None
else:
previous_ts = experiment.trials[num_trials - 1].time_created.timestamp()
gen_time = time.time() - previous_ts
# Create a GR
arms, candidate_metadata_by_arm_signature = gen_arms(
observation_features=[obsf], arms_by_signature=experiment.arms_by_signature
)
gr = GeneratorRun(
arms=arms,
gen_time=gen_time,
candidate_metadata_by_arm_signature=candidate_metadata_by_arm_signature,
)
# Add it as a trial
trial = experiment.new_trial().add_generator_run(gr).run()
# Evaluate function
df = trial.fetch_data().df
if len(df) > 1:
raise Exception("Does not support multiple outcomes") # pragma: no cover
obj = float(df["mean"].values[0])
if not problem.optimization_config.objective.minimize:
obj = -obj
return obj
return experiment, evaluation_function
def _validate_groups(
problem_groups: Dict[str, Union[List[BenchmarkProblem], List[str]]],
method_groups: Dict[str, Union[List[GenerationStrategy], List[str]]],
) -> None:
# Check for dict with lists as values.
problem_groups_is_dict_of_lists = isinstance(problem_groups, dict) and all(
isinstance(problems, list) for problems in problem_groups.values()
)
if not problem_groups_is_dict_of_lists:
raise ValueError(
"`problem_groups` does not match the expected type of "
"Dict[str, List[BenchmarkProblem]]. "
"Example: problem_groups = {'single_fidelity': [problem1, problem2]}"
)
method_groups_is_dict_of_lists = isinstance(method_groups, dict) and all(
isinstance(problems, list) for problems in method_groups.values()
)
if not method_groups_is_dict_of_lists:
raise ValueError(
"`method_groups` does not match the expected type of "
"Dict[str, List[GenerationStrategy]]. "
"Example: method_groups = {'single_fidelity': [strategy1, strategy2]}"
)
# Check that `problem_groups` and `method_groups` have the same keys.
if problem_groups.keys() != method_groups.keys():
raise ValueError(
"`problem_groups` and `method_groups` should have the same keys."
)
