def test_get_best_trial(self):
scheduler = Scheduler(
experiment=self.branin_experiment, # Has runner and metrics.
generation_strategy=self.two_sobol_steps_GS,
options=SchedulerOptions(
init_seconds_between_polls=
0.1, # Short between polls so test is fast.
),
)
self.assertIsNone(scheduler.get_best_parameters())
scheduler.run_n_trials(max_trials=1)
trial, params, _arm = scheduler.get_best_trial()
just_params, _just_arm = scheduler.get_best_parameters()
just_params_unmodeled, _just_arm_unmodled = scheduler.get_best_parameters(
use_model_predictions=False)
with self.assertRaisesRegex(NotImplementedError,
"Please use `get_best_parameters`"):
scheduler.get_pareto_optimal_parameters()
self.assertEqual(trial, 0)
self.assertIn("x1", params)
self.assertIn("x2", params)
self.assertEqual(params, just_params)
self.assertEqual(params, just_params_unmodeled)
def test_validate_early_stopping_strategy(self):
class DummyEarlyStoppingStrategy(BaseEarlyStoppingStrategy):
def should_stop_trials_early(
self,
trial_indices: Set[int],
experiment: Experiment,
**kwargs: Dict[str, Any],
) -> Set[int]:
return {}
with patch(
f"{BraninMetric.__module__}.BraninMetric.is_available_while_running",
return_value=False,
), self.assertRaises(ValueError):
Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.sobol_GPEI_GS,
options=SchedulerOptions(
early_stopping_strategy=DummyEarlyStoppingStrategy()),
)
# should not error
Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.sobol_GPEI_GS,
options=SchedulerOptions(
early_stopping_strategy=DummyEarlyStoppingStrategy()),
)
def test_run_multi_arm_generator_run_error(self, mock_gen):
scheduler = Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.sobol_GPEI_GS,
options=SchedulerOptions(total_trials=1),
)
with self.assertRaisesRegex(SchedulerInternalError,
".* only one was expected"):
scheduler.run_all_trials()
def test_base_report_results(self):
self.branin_experiment.runner = NoReportResultsRunner()
scheduler = Scheduler(
experiment=self.branin_experiment, # Has runner and metrics.
generation_strategy=self.two_sobol_steps_GS,
options=SchedulerOptions(init_seconds_between_polls=0, ),
)
self.assertEqual(scheduler.run_n_trials(max_trials=3),
OptimizationResult())
def test_retries(self):
# Check that retries will be performed for a retriable error.
self.branin_experiment.runner = BrokenRunnerRuntimeError()
scheduler = Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.two_sobol_steps_GS,
options=SchedulerOptions(total_trials=1),
)
# Should raise after 3 retries.
with self.assertRaisesRegex(RuntimeError, ".* testing .*"):
scheduler.run_all_trials()
self.assertEqual(scheduler.run_trial_call_count, 3)
def test_retries_nonretriable_error(self):
# Check that no retries will be performed for `ValueError`, since we
# exclude it from the retriable errors.
self.branin_experiment.runner = BrokenRunnerValueError()
scheduler = Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.two_sobol_steps_GS,
options=SchedulerOptions(total_trials=1),
)
# Should raise right away since ValueError is non-retriable.
with self.assertRaisesRegex(ValueError, ".* testing .*"):
scheduler.run_all_trials()
self.assertEqual(scheduler.run_trial_call_count, 1)
def test_optimization_complete(self, _):
# With runners & metrics, `Scheduler.run_all_trials` should run.
scheduler = Scheduler(
experiment=self.branin_experiment, # Has runner and metrics.
generation_strategy=self.two_sobol_steps_GS,
options=SchedulerOptions(
max_pending_trials=100,
init_seconds_between_polls=
0.1, # Short between polls so test is fast.
),
)
scheduler.run_n_trials(max_trials=1)
# no trials should run if _gen_multiple throws an OptimizationComplete error
self.assertEqual(len(scheduler.experiment.trials), 0)
def test_timeout(self):
scheduler = Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.two_sobol_steps_GS,
options=SchedulerOptions(
total_trials=8,
init_seconds_between_polls=
0, # No wait between polls so test is fast.
),
)
scheduler.run_all_trials(
timeout_hours=0) # Forcing optimization to time out.
self.assertEqual(len(scheduler.experiment.trials), 0)
self.assertIn("aborted",
scheduler.experiment._properties["run_trials_success"])
def test_set_ttl(self):
scheduler = Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.two_sobol_steps_GS,
options=SchedulerOptions(
total_trials=2,
ttl_seconds_for_trials=1,
init_seconds_between_polls=
0, # No wait between polls so test is fast.
min_seconds_before_poll=0.0,
),
)
scheduler.run_all_trials()
self.assertTrue(
all(t.ttl_seconds == 1
for t in scheduler.experiment.trials.values()))
def test_repr(self):
scheduler = Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.sobol_GPEI_GS,
options=SchedulerOptions(
total_trials=0,
tolerated_trial_failure_rate=0.2,
init_seconds_between_polls=10,
),
)
self.assertEqual(
f"{scheduler}",
("Scheduler(experiment=Experiment(branin_test_experiment), "
"generation_strategy=GenerationStrategy(name='Sobol+GPEI', "
"steps=[Sobol for 5 trials, GPEI for subsequent trials]), "
"options=SchedulerOptions(max_pending_trials=10, "
"trial_type=<class 'ax.core.trial.Trial'>, "
"total_trials=0, tolerated_trial_failure_rate=0.2, "
"min_failed_trials_for_failure_rate_check=5, log_filepath=None, "
"logging_level=20, ttl_seconds_for_trials=None, init_seconds_between_"
"polls=10, min_seconds_before_poll=1.0, seconds_between_polls_backoff_"
"factor=1.5, run_trials_in_batches=False, "
"debug_log_run_metadata=False, early_stopping_strategy=None, "
"suppress_storage_errors_after_retries=False))"),
)
def test_validate_early_stopping_strategy(self):
with patch(
f"{BraninMetric.__module__}.BraninMetric.is_available_while_running",
return_value=False,
), self.assertRaises(ValueError):
Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.sobol_GPEI_GS,
options=SchedulerOptions(
early_stopping_strategy=DummyEarlyStoppingStrategy()),
)
with patch.object(
OptimizationConfig, "is_moo_problem", return_value=True
), self.assertRaisesRegex(
UnsupportedError,
"Early stopping is not supported on multi-objective problems",
):
Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.sobol_GPEI_GS,
options=SchedulerOptions(
early_stopping_strategy=DummyEarlyStoppingStrategy()),
)
with patch(
f"{Scheduler.__module__}.len",
return_value=1,
), self.assertRaisesRegex(
UnsupportedError,
"Early stopping is not supported on problems with outcome constraints.",
):
Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.sobol_GPEI_GS,
options=SchedulerOptions(
early_stopping_strategy=DummyEarlyStoppingStrategy()),
)
# should not error
Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.sobol_GPEI_GS,
options=SchedulerOptions(
early_stopping_strategy=DummyEarlyStoppingStrategy()),
)
def test_run_preattached_trials_only(self):
# assert that pre-attached trials run when max_trials = number of
# pre-attached trials
scheduler = Scheduler(
experiment=self.branin_experiment, # Has runner and metrics.
generation_strategy=self.two_sobol_steps_GS,
options=SchedulerOptions(
init_seconds_between_polls=
0.1, # Short between polls so test is fast.
),
)
trial = scheduler.experiment.new_trial()
parameter_dict = {"x1": 5, "x2": 5}
trial.add_arm(Arm(parameters=parameter_dict))
with self.assertRaisesRegex(
UserInputError,
"number of pre-attached candidate trials .* is greater than"):
scheduler.run_n_trials(max_trials=0)
scheduler.run_n_trials(max_trials=1)
self.assertEqual(len(scheduler.experiment.trials), 1)
self.assertDictEqual(scheduler.experiment.trials[0].arm.parameters,
parameter_dict)
self.assertTrue( # Make sure all trials got to complete.
all(t.completed_successfully
for t in scheduler.experiment.trials.values()))
def test_run_all_trials_using_runner_and_metrics(self, mock_get_pending):
# With runners & metrics, `Scheduler.run_all_trials` should run.
scheduler = Scheduler(
experiment=self.branin_experiment, # Has runner and metrics.
generation_strategy=self.two_sobol_steps_GS,
options=SchedulerOptions(
total_trials=8,
init_seconds_between_polls=
0.1, # Short between polls so test is fast.
),
)
scheduler.run_all_trials()
# Check that we got pending feat. at least 8 times (1 for each new trial and
# maybe more for cases where we tried to generate trials but ran into limit on
# paralel., as polling trial statuses is randomized in Scheduler),
# so some trials might not yet have come back.
self.assertGreaterEqual(len(mock_get_pending.call_args_list), 8)
self.assertTrue( # Make sure all trials got to complete.
all(t.completed_successfully
for t in scheduler.experiment.trials.values()))
self.assertEqual(len(scheduler.experiment.trials), 8)
# Check that all the data, fetched during optimization, was attached to the
# experiment.
dat = scheduler.experiment.fetch_data().df
self.assertEqual(set(dat["trial_index"].values), set(range(8)))
self.assertEqual(
scheduler.experiment._properties[
ExperimentStatusProperties.RUN_TRIALS_STATUS],
["started", "success"],
)
self.assertEqual(
scheduler.experiment._properties[
ExperimentStatusProperties.NUM_TRIALS_RUN_PER_CALL],
[8],
)
self.assertEqual(
scheduler.experiment._properties[
ExperimentStatusProperties.RESUMED_FROM_STORAGE_TIMESTAMPS],
[],
)
def test_stop_at_MAX_SECONDS_BETWEEN_POLLS(self):
self.branin_experiment.runner = InfinitePollRunner()
scheduler = Scheduler(
experiment=self.branin_experiment, # Has runner and metrics.
generation_strategy=self.two_sobol_steps_GS,
options=SchedulerOptions(
total_trials=8,
init_seconds_between_polls=
1, # No wait between polls so test is fast.
),
)
with patch.object(scheduler,
"wait_for_completed_trials_and_report_results",
return_value=None) as mock_await_trials:
scheduler.run_all_trials(timeout_hours=1 / 60 /
15) # 4 second timeout.
# We should be calling `wait_for_completed_trials_and_report_results`
# N = total runtime / `MAX_SECONDS_BETWEEN_POLLS` times.
self.assertEqual(
len(mock_await_trials.call_args),
2, # MAX_SECONDS_BETWEEN_POLLS as patched in decorator
)
def test_logging(self):
with NamedTemporaryFile() as temp_file:
Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.sobol_GPEI_GS,
options=SchedulerOptions(
total_trials=1,
init_seconds_between_polls=
0, # No wait bw polls so test is fast.
log_filepath=temp_file.name,
),
).run_all_trials()
self.assertGreater(os.stat(temp_file.name).st_size, 0)
self.assertIn("Running trials [0]", str(temp_file.readline()))
temp_file.close()
def test_init(self):
with self.assertRaisesRegex(
UnsupportedError,
"`Scheduler` requires that experiment specifies a `Runner`.",
):
scheduler = Scheduler(
experiment=self.branin_experiment_no_impl_runner_or_metrics,
generation_strategy=self.sobol_GPEI_GS,
options=SchedulerOptions(total_trials=10),
)
self.branin_experiment_no_impl_runner_or_metrics.runner = self.runner
with self.assertRaisesRegex(
UnsupportedError,
".*Metrics {'branin'} do not implement fetching logic.",
):
scheduler = Scheduler(
experiment=self.branin_experiment_no_impl_runner_or_metrics,
generation_strategy=self.sobol_GPEI_GS,
options=SchedulerOptions(total_trials=10),
)
scheduler = Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.sobol_GPEI_GS,
options=SchedulerOptions(
total_trials=0,
tolerated_trial_failure_rate=0.2,
init_seconds_between_polls=10,
),
)
self.assertEqual(scheduler.experiment, self.branin_experiment)
self.assertEqual(scheduler.generation_strategy, self.sobol_GPEI_GS)
self.assertEqual(scheduler.options.total_trials, 0)
self.assertEqual(scheduler.options.tolerated_trial_failure_rate, 0.2)
self.assertEqual(scheduler.options.init_seconds_between_polls, 10)
self.assertIsNone(scheduler._latest_optimization_start_timestamp)
for status_prop in ExperimentStatusProperties:
self.assertEqual(
scheduler.experiment._properties[status_prop.value], [])
scheduler.run_all_trials() # Runs no trials since total trials is 0.
# `_latest_optimization_start_timestamp` should be set now.
self.assertLessEqual(
scheduler._latest_optimization_start_timestamp,
current_timestamp_in_millis(),
)
def test_get_best_trial_moo(self):
experiment = get_branin_experiment_with_multi_objective()
experiment.runner = self.runner
scheduler = Scheduler(
experiment=experiment,
generation_strategy=self.sobol_GPEI_GS,
options=SchedulerOptions(init_seconds_between_polls=0.1),
)
scheduler.run_n_trials(max_trials=1)
with self.assertRaisesRegex(
NotImplementedError,
"Please use `get_pareto_optimal_parameters`"):
scheduler.get_best_trial()
with self.assertRaisesRegex(
NotImplementedError,
"Please use `get_pareto_optimal_parameters`"):
scheduler.get_best_parameters()
self.assertIsNotNone(scheduler.get_pareto_optimal_parameters())
def test_suppress_all_storage_errors(self, mock_save_exp, _):
init_test_engine_and_session_factory(force_init=True)
config = SQAConfig()
encoder = Encoder(config=config)
decoder = Decoder(config=config)
db_settings = DBSettings(encoder=encoder, decoder=decoder)
Scheduler(
experiment=self.branin_experiment, # Has runner and metrics.
generation_strategy=self.two_sobol_steps_GS,
options=SchedulerOptions(
max_pending_trials=100,
init_seconds_between_polls=
0.1, # Short between polls so test is fast.
suppress_storage_errors_after_retries=True,
),
db_settings=db_settings,
)
self.assertEqual(mock_save_exp.call_count, 3)
def test_logging_level(self):
# We don't have any warnings yet, so warning level of logging shouldn't yield
# any logs as of now.
with NamedTemporaryFile() as temp_file:
Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.sobol_GPEI_GS,
options=SchedulerOptions(
total_trials=3,
init_seconds_between_polls=
0, # No wait bw polls so test is fast.
log_filepath=temp_file.name,
logging_level=WARNING,
),
).run_all_trials()
# Ensure that the temp file remains empty
self.assertEqual(os.stat(temp_file.name).st_size, 0)
temp_file.close()
def test_stop_trial(self):
# With runners & metrics, `Scheduler.run_all_trials` should run.
scheduler = Scheduler(
experiment=self.branin_experiment, # Has runner and metrics.
generation_strategy=self.two_sobol_steps_GS,
options=SchedulerOptions(
init_seconds_between_polls=
0.1, # Short between polls so test is fast.
),
)
with patch.object(scheduler.experiment.runner,
"stop",
return_value=None) as mock_runner_stop:
scheduler.run_n_trials(max_trials=1)
scheduler.stop_trial_runs(trials=[scheduler.experiment.trials[0]])
mock_runner_stop.assert_called_once()
def test_failure_rate(self):
options = SchedulerOptions(
total_trials=8,
tolerated_trial_failure_rate=0.5,
init_seconds_between_polls=
0, # No wait between polls so test is fast.
min_failed_trials_for_failure_rate_check=2,
)
self.branin_experiment.runner = RunnerWithFrequentFailedTrials()
scheduler = Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.sobol_GS_no_parallelism,
options=options,
)
with self.assertRaises(FailureRateExceededError):
scheduler.run_all_trials()
# Trials will have statuses: 0, 2 - FAILED, 1 - COMPLETED. Failure rate
# is 0.5, and so if 2 of the first 3 trials are failed, we can fail
# immediately.
self.assertEqual(len(scheduler.experiment.trials), 3)
# If all trials fail, we can be certain that the sweep will
# fail after only 2 trials.
num_preexisting_trials = len(scheduler.experiment.trials)
self.branin_experiment.runner = RunnerWithAllFailedTrials()
scheduler = Scheduler(
experiment=self.branin_experiment,
generation_strategy=self.sobol_GS_no_parallelism,
options=options,
)
self.assertEqual(scheduler._num_preexisting_trials,
num_preexisting_trials)
with self.assertRaises(FailureRateExceededError):
scheduler.run_all_trials()
self.assertEqual(len(scheduler.experiment.trials),
num_preexisting_trials + 2)
def test_run_n_trials(self):
# With runners & metrics, `Scheduler.run_all_trials` should run.
scheduler = Scheduler(
experiment=self.branin_experiment, # Has runner and metrics.
generation_strategy=self.two_sobol_steps_GS,
options=SchedulerOptions(
init_seconds_between_polls=
0.1, # Short between polls so test is fast.
),
)
scheduler.run_n_trials(max_trials=1)
self.assertEqual(len(scheduler.experiment.trials), 1)
scheduler.run_n_trials(max_trials=10)
self.assertTrue( # Make sure all trials got to complete.
all(t.completed_successfully
for t in scheduler.experiment.trials.values()))
# Check that all the data, fetched during optimization, was attached to the
# experiment.
dat = scheduler.experiment.fetch_data().df
self.assertEqual(set(dat["trial_index"].values), set(range(11)))
def test_sqa_storage(self):
init_test_engine_and_session_factory(force_init=True)
config = SQAConfig()
encoder = Encoder(config=config)
decoder = Decoder(config=config)
db_settings = DBSettings(encoder=encoder, decoder=decoder)
experiment = self.branin_experiment
# Scheduler currently requires that the experiment be pre-saved.
with self.assertRaisesRegex(ValueError, ".* must specify a name"):
experiment._name = None
scheduler = Scheduler(
experiment=experiment,
generation_strategy=self.two_sobol_steps_GS,
options=SchedulerOptions(total_trials=1),
db_settings=db_settings,
)
experiment._name = "test_experiment"
NUM_TRIALS = 5
scheduler = Scheduler(
experiment=experiment,
generation_strategy=self.two_sobol_steps_GS,
options=SchedulerOptions(
total_trials=NUM_TRIALS,
init_seconds_between_polls=
0, # No wait between polls so test is fast.
),
db_settings=db_settings,
)
# Check that experiment and GS were saved.
exp, gs = scheduler._load_experiment_and_generation_strategy(
experiment.name)
self.assertEqual(exp, experiment)
self.assertEqual(gs, self.two_sobol_steps_GS)
scheduler.run_all_trials()
# Check that experiment and GS were saved and test reloading with reduced state.
exp, gs = scheduler._load_experiment_and_generation_strategy(
experiment.name, reduced_state=True)
self.assertEqual(len(exp.trials), NUM_TRIALS)
self.assertEqual(len(gs._generator_runs), NUM_TRIALS)
# Test `from_stored_experiment`.
new_scheduler = Scheduler.from_stored_experiment(
experiment_name=experiment.name,
options=SchedulerOptions(
total_trials=NUM_TRIALS + 1,
init_seconds_between_polls=
0, # No wait between polls so test is fast.
),
db_settings=db_settings,
)
# Hack "resumed from storage timestamp" into `exp` to make sure all other fields
# are equal, since difference in resumed from storage timestamps is expected.
exp._properties[
ExperimentStatusProperties.
RESUMED_FROM_STORAGE_TIMESTAMPS] = new_scheduler.experiment._properties[
ExperimentStatusProperties.RESUMED_FROM_STORAGE_TIMESTAMPS]
self.assertEqual(new_scheduler.experiment, exp)
self.assertEqual(new_scheduler.generation_strategy, gs)
self.assertEqual(
len(new_scheduler.experiment._properties[
ExperimentStatusProperties.RESUMED_FROM_STORAGE_TIMESTAMPS]),
1,
)
