def test_tracer(self):
global_values.declare_singletons()
global_values.tracer = Tracer(actor_id='1', thread_id='0')
try:
fib_number = self.fibonacci(n=3)
except:
pass
trace_events = global_values.tracer.trace_events
reformatted_events = Tracer.reformat_events(trace_events)
assert len(trace_events) > 0
print(json.dumps(reformatted_events, indent=2))
def setUpClass(cls) -> None:
mlos.global_values.declare_singletons()
mlos.global_values.tracer = Tracer(actor_id=cls.__name__, thread_id=0)
cls.temp_dir = os.path.join(os.getcwd(), "temp")
if not os.path.exists(cls.temp_dir):
os.mkdir(cls.temp_dir)
def setUpClass(cls):
"""Sets up all the singletons needed to test the BayesianOptimizer.
"""
warnings.simplefilter("error")
global_values.declare_singletons()
global_values.tracer = Tracer(actor_id=cls.__name__, thread_id=0)
cls.logger = create_logger(logger_name=cls.__name__)
def setup_class(cls) -> None:
mlos.global_values.declare_singletons()
mlos.global_values.tracer = Tracer(actor_id=cls.__name__, thread_id=0)
cls.temp_dir = os.path.join(os.getcwd(), "temp")
if not os.path.exists(cls.temp_dir):
os.mkdir(cls.temp_dir)
cls.logger = create_logger("TestOptimizerEvaluator")
def setup_class(cls):
""" Set's up all the objects needed to test the RandomSearchOptimizer
To test the RandomSearchOptimizer we need to first construct:
* an optimization problem
* a utility function
To construct a utility function we need the same set up as in the TestConfidenceBoundUtilityFunction
test.
:return:
"""
global_values.declare_singletons()
global_values.tracer = Tracer(actor_id=cls.__name__, thread_id=0)
objective_function_config = objective_function_config_store.get_config_by_name(
'2d_quadratic_concave_up')
objective_function = ObjectiveFunctionFactory.create_objective_function(
objective_function_config=objective_function_config)
cls.input_space = objective_function.parameter_space
cls.output_space = objective_function.output_space
cls.input_values_dataframe = objective_function.parameter_space.random_dataframe(
num_samples=2500)
cls.output_values_dataframe = objective_function.evaluate_dataframe(
cls.input_values_dataframe)
cls.model_config = homogeneous_random_forest_config_store.default
print(cls.model_config)
cls.model = MultiObjectiveHomogeneousRandomForest(
model_config=cls.model_config,
input_space=cls.input_space,
output_space=cls.output_space)
cls.model.fit(cls.input_values_dataframe,
cls.output_values_dataframe,
iteration_number=len(cls.input_values_dataframe.index))
cls.utility_function_config = Point(
utility_function_name="upper_confidence_bound_on_improvement",
alpha=0.05)
cls.optimization_problem = OptimizationProblem(
parameter_space=cls.input_space,
objective_space=cls.output_space,
objectives=[Objective(name='y', minimize=True)])
cls.utility_function = ConfidenceBoundUtilityFunction(
function_config=cls.utility_function_config,
surrogate_model=cls.model,
minimize=cls.optimization_problem.objectives[0].minimize)
def setup_class(cls):
""" Set's up all the objects needed to test the UtilityFunctionOptimizers
To test the UtilityFunctionOptimizers we need to first construct:
* an objective function for the model to approximate and its corresponding parameter and output spaces
* an optimization problem
* a regression model, then train it on some random parameters to the objective function
* a utility function that utilizes the model
* a pareto frontier over the random parameters
And only then do we get to test our utility function optimizers. This is a lot of work and a somewhat cleaner approach
would be to simply create an instance of the BayesianOptimizer to do all of the above for us, but then we might not be able
to test the utility function optimizers as thoroughly as we need to.
:return:
"""
global_values.declare_singletons()
global_values.tracer = Tracer(actor_id=cls.__name__, thread_id=0)
cls.logger = create_logger("TestUtilityFunctionOptimizers")
cls.model_config = multi_objective_pass_through_model_config_store.default
cls.model = MultiObjectivePassThroughModelForTesting(
model_config=cls.model_config,
logger=cls.logger
)
cls.objective_function = cls.model.objective_function
cls.parameter_space = cls.objective_function.parameter_space
cls.objective_space = cls.objective_function.output_space
cls.optimization_problem = cls.objective_function.default_optimization_problem
cls.utility_function_config = Point(utility_function_name="upper_confidence_bound_on_improvement", alpha=0.05)
cls.utility_function = ConfidenceBoundUtilityFunction(
function_config=cls.utility_function_config,
surrogate_model=cls.model,
minimize=cls.optimization_problem.objectives[0].minimize,
logger=cls.logger
)
# To make the pareto frontier we have to generate some random points.
#
cls.parameters_df = cls.objective_function.parameter_space.random_dataframe(1000)
cls.objectives_df = cls.objective_function.evaluate_dataframe(cls.parameters_df)
cls.pareto_frontier = ParetoFrontier(
optimization_problem=cls.optimization_problem,
objectives_df=cls.objectives_df,
parameters_df=cls.parameters_df
)
def setup_class(cls):
""" Sets up all the singletons needed to test the BayesianOptimizer.
"""
warnings.simplefilter("error")
global_values.declare_singletons()
global_values.tracer = Tracer(actor_id=cls.__name__, thread_id=0)
cls.logger = create_logger(logger_name=cls.__name__)
cls.logger.setLevel(logging.DEBUG)
cls.port = None
# Start up the gRPC service. Try a bunch of ports, before giving up so we can do several in parallel.
#
max_num_tries = 100
num_tries = 0
for port in range(50051, 50051 + max_num_tries):
num_tries += 1
try:
cls.server = OptimizerMicroserviceServer(port=port,
num_threads=10,
logger=cls.logger)
cls.server.start()
cls.port = port
break
except:
cls.logger.info(
f"Failed to create OptimizerMicroserviceServer on port {port}."
)
if num_tries == max_num_tries:
raise
cls.optimizer_service_channel = grpc.insecure_channel(
f'localhost:{cls.port}')
cls.bayesian_optimizer_factory = BayesianOptimizerFactory(
grpc_channel=cls.optimizer_service_channel, logger=cls.logger)
cls.temp_dir = os.path.join(os.getcwd(), "temp")
if not os.path.exists(cls.temp_dir):
os.mkdir(cls.temp_dir)
cls.trace_output_path = os.path.join(
cls.temp_dir, "TestBayesianOptimizerTrace.json")
try:
os.remove(cls.trace_output_path)
except OSError:
pass
def setUpClass(cls):
""" Sets up all the singletons needed to test the BayesianOptimizer.
"""
warnings.simplefilter("error")
global_values.declare_singletons()
global_values.tracer = Tracer(actor_id=cls.__name__, thread_id=0)
cls.logger = create_logger(logger_name=cls.__name__)
# Start up the gRPC service.
#
cls.server = OptimizerMicroserviceServer(port=50051, num_threads=10)
cls.server.start()
cls.optimizer_service_channel = grpc.insecure_channel('localhost:50051')
cls.bayesian_optimizer_factory = BayesianOptimizerFactory(grpc_channel=cls.optimizer_service_channel, logger=cls.logger)
def setUpClass(cls):
""" Set's up all the objects needed to test the RandomSearchOptimizer
To test the RandomSearchOptimizer we need to first construct:
* an optimization problem
* a utility function
To construct a utility function we need the same set up as in the TestConfidenceBoundUtilityFunction
test.
:return:
"""
cls.temp_dir = os.path.join(os.getcwd(), "temp")
if not os.path.exists(cls.temp_dir):
os.mkdir(cls.temp_dir)
global_values.declare_singletons()
global_values.tracer = Tracer(actor_id=cls.__name__, thread_id=0)
cls.logger = create_logger(logger_name=cls.__name__)
def setup_class(cls) -> None:
global_values.declare_singletons()
global_values.tracer = Tracer(actor_id=cls.__name__, thread_id=0)
def evaluate_optimizer(self) -> OptimizerEvaluationReport: # pylint: disable=too-many-statements,too-many-branches
evaluation_report = OptimizerEvaluationReport(
optimizer_configuration=self.optimizer_config,
objective_function_configuration=self.objective_function_config,
num_optimization_iterations=self.optimizer_evaluator_config.num_iterations,
evaluation_frequency=self.optimizer_evaluator_config.evaluation_frequency
)
if self.optimizer_evaluator_config.include_execution_trace_in_report:
mlos.global_values.declare_singletons()
if mlos.global_values.tracer is None:
mlos.global_values.tracer = Tracer()
mlos.global_values.tracer.clear_events()
if self.optimizer_evaluator_config.include_pickled_objective_function_in_report:
evaluation_report.pickled_objective_function_initial_state = pickle.dumps(self.objective_function)
if self.optimizer_evaluator_config.include_pickled_optimizer_in_report:
evaluation_report.pickled_optimizer_initial_state = pickle.dumps(self.optimizer)
multi_objective_regression_model_fit_state = MultiObjectiveRegressionModelFitState(objective_names=self.optimizer.optimization_problem.objective_names)
for objective_name in self.optimizer.optimization_problem.objective_names:
multi_objective_regression_model_fit_state[objective_name] = RegressionModelFitState()
optima_over_time = {}
optima_over_time[OptimumDefinition.BEST_OBSERVATION.value] = OptimumOverTime(
optimization_problem=self.optimizer.optimization_problem,
optimum_definition=OptimumDefinition.BEST_OBSERVATION
)
optima_over_time[OptimumDefinition.PREDICTED_VALUE_FOR_OBSERVED_CONFIG.value] = OptimumOverTime(
optimization_problem=self.optimizer.optimization_problem,
optimum_definition=OptimumDefinition.PREDICTED_VALUE_FOR_OBSERVED_CONFIG
)
optima_over_time[f"{OptimumDefinition.UPPER_CONFIDENCE_BOUND_FOR_OBSERVED_CONFIG.value}_99"] = OptimumOverTime(
optimization_problem=self.optimizer.optimization_problem,
optimum_definition=OptimumDefinition.UPPER_CONFIDENCE_BOUND_FOR_OBSERVED_CONFIG,
alpha=0.01
)
optima_over_time[f"{OptimumDefinition.LOWER_CONFIDENCE_BOUND_FOR_OBSERVED_CONFIG.value}_99"] = OptimumOverTime(
optimization_problem=self.optimizer.optimization_problem,
optimum_definition=OptimumDefinition.LOWER_CONFIDENCE_BOUND_FOR_OBSERVED_CONFIG,
alpha=0.01
)
#####################################################################################################
evaluation_report.start_time = datetime.utcnow()
i = 0
try:
with traced(scope_name="optimization_loop"):
for i in range(self.optimizer_evaluator_config.num_iterations):
parameters = self.optimizer.suggest()
objectives = self.objective_function.evaluate_point(parameters)
self.optimizer.register(parameters.to_dataframe(), objectives.to_dataframe())
if i % self.optimizer_evaluator_config.evaluation_frequency == 0:
self.logger.info(f"[{i + 1}/{self.optimizer_evaluator_config.num_iterations}]")
with traced(scope_name="evaluating_optimizer"):
if self.optimizer_evaluator_config.include_pickled_optimizer_in_report:
evaluation_report.add_pickled_optimizer(iteration=i, pickled_optimizer=pickle.dumps(self.optimizer))
if self.optimizer.trained:
multi_objective_gof_metrics = self.optimizer.compute_surrogate_model_goodness_of_fit()
for objective_name, gof_metrics in multi_objective_gof_metrics:
multi_objective_regression_model_fit_state[objective_name].set_gof_metrics(
data_set_type=DataSetType.TRAIN,
gof_metrics=gof_metrics
)
for optimum_name, optimum_over_time in optima_over_time.items():
try:
config, value = self.optimizer.optimum(
optimum_definition=optimum_over_time.optimum_definition,
alpha=optimum_over_time.alpha
)
optima_over_time[optimum_name].add_optimum_at_iteration(
iteration=i,
optimum_config=config,
optimum_value=value
)
except ValueError as e:
self.logger.info(f"Failed to get {optimum_name} optimum.", exc_info=True)
if self.optimizer_evaluator_config.report_pareto_over_time:
evaluation_report.pareto_over_time[i] = copy.deepcopy(self.optimizer.optimization_problem)
if self.optimizer_evaluator_config.report_pareto_volume_over_time:
volume_estimator = self.optimizer.pareto_frontier.approximate_pareto_volume()
ci99_on_volume = volume_estimator.get_two_sided_confidence_interval_on_pareto_volume(alpha=0.01)
evaluation_report.pareto_volume_over_time[i] = ci99_on_volume
evaluation_report.success = True
except Exception as e:
evaluation_report.success = False
evaluation_report.exception = e
evaluation_report.exception_traceback = traceback.format_exc()
evaluation_report.end_time = datetime.utcnow()
with traced(scope_name="evaluating_optimizer"):
# Once the optimization is done, we perform a final evaluation of the optimizer.
if self.optimizer.trained:
multi_objective_gof_metrics = self.optimizer.compute_surrogate_model_goodness_of_fit()
for objective_name, gof_metrics in multi_objective_gof_metrics:
multi_objective_regression_model_fit_state[objective_name].set_gof_metrics(data_set_type=DataSetType.TRAIN, gof_metrics=gof_metrics)
for optimum_name, optimum_over_time in optima_over_time.items():
try:
config, value = self.optimizer.optimum(optimum_definition=optimum_over_time.optimum_definition, alpha=optimum_over_time.alpha)
optima_over_time[optimum_name].add_optimum_at_iteration(
iteration=self.optimizer_evaluator_config.num_iterations,
optimum_config=config,
optimum_value=value
)
except Exception as e:
self.logger.info(f"Failed to get {optimum_name} optimum.", exc_info=True)
if self.optimizer_evaluator_config.report_pareto_over_time:
evaluation_report.pareto_over_time[i] = copy.deepcopy(self.optimizer.optimization_problem)
if self.optimizer_evaluator_config.report_pareto_volume_over_time:
volume_estimator = self.optimizer.pareto_frontier.approximate_pareto_volume()
ci99_on_volume = volume_estimator.get_two_sided_confidence_interval_on_pareto_volume(alpha=0.01)
evaluation_report.pareto_volume_over_time[i] = ci99_on_volume
if self.optimizer_evaluator_config.include_execution_trace_in_report:
evaluation_report.execution_trace = mlos.global_values.tracer.trace_events
mlos.global_values.tracer.clear_events()
if self.optimizer_evaluator_config.include_pickled_optimizer_in_report:
evaluation_report.add_pickled_optimizer(iteration=i, pickled_optimizer=pickle.dumps(self.optimizer))
if self.optimizer_evaluator_config.include_pickled_objective_function_in_report:
evaluation_report.pickled_objective_function_final_state = pickle.dumps(self.objective_function)
if self.optimizer_evaluator_config.report_regression_model_goodness_of_fit:
evaluation_report.regression_model_fit_state = multi_objective_regression_model_fit_state
if self.optimizer_evaluator_config.report_optima_over_time:
evaluation_report.optima_over_time = optima_over_time
return evaluation_report
def write_to_disk(self, target_folder):
"""Writes the report to disk.
The layout on disk is as follows:
- optimizer_config.json
- objective_function_config.json
- goodness_of_fit.pickle
- objective_function_initial_state.pickle
- objective_function_final_state.pickle
- execution_trace.json
- execution_info.json
- pickled_optimizers:
- {iteration_number}.pickle
"""
optimizer_config_file = os.path.join(target_folder,
"optimizer_config.json")
with open(optimizer_config_file, 'w') as out_file:
out_file.write(self.optimizer_configuration.to_json(indent=2))
objective_function_config_file = os.path.join(
target_folder, "objective_function_config.json")
with open(objective_function_config_file, 'w') as out_file:
out_file.write(
self.objective_function_configuration.to_json(indent=2))
if len(self.pickled_optimizers_over_time) > 0:
pickled_optimizers_dir = os.path.join(target_folder,
"pickled_optimizers")
if not os.path.exists(pickled_optimizers_dir):
os.mkdir(pickled_optimizers_dir)
for iteration, pickled_optimizer in self.pickled_optimizers_over_time.items(
):
with open(
os.path.join(pickled_optimizers_dir,
f"{iteration}.pickle"), 'wb') as out_file:
out_file.write(pickled_optimizer)
if self.pickled_objective_function_initial_state is not None:
with open(
os.path.join(target_folder,
"objective_function_initial_state.pickle"),
"wb") as out_file:
out_file.write(self.pickled_objective_function_initial_state)
if self.pickled_objective_function_final_state is not None:
with open(
os.path.join(target_folder,
"objective_function_final_state.pickle"),
"wb") as out_file:
out_file.write(self.pickled_objective_function_final_state)
if self.regression_model_fit_state is not None:
with open(
os.path.join(
target_folder,
"regression_model_goodness_of_fit_state.pickle"),
"wb") as out_file:
pickle.dump(self.regression_model_fit_state, out_file)
if self.optima_over_time is not None:
with open(os.path.join(target_folder, "optima_over_time.pickle"),
"wb") as out_file:
pickle.dump(self.optima_over_time, out_file)
if len(self.pareto_over_time) > 0:
with open(os.path.join(target_folder, "pareto_over_time.pickle"),
"wb") as out_file:
pickle.dump(self.pareto_over_time, out_file)
if len(self.pareto_volume_over_time) > 0:
with open(
os.path.join(target_folder,
"pareto_volume_over_time.json"),
"w") as out_file:
json.dump(self.pareto_volume_over_time, out_file, indent=2)
if self.execution_trace is not None:
tracer = Tracer()
tracer.trace_events = self.execution_trace
tracer.dump_trace_to_file(output_file_path=os.path.join(
target_folder, "execution_trace.json"))
with open(os.path.join(target_folder, "execution_info.json"),
'w') as out_file:
execution_info_dict = {
'success':
self.success,
'num_optimization_iterations':
self.num_optimization_iterations,
'evaluation_frequency':
self.evaluation_frequency,
'exception':
str(self.exception) if self.exception is not None else None,
'exception_stack_trace':
self.exception_traceback,
'start_time':
self.start_time.strftime(self.DATETIME_FORMAT),
'end_time':
self.end_time.strftime(self.DATETIME_FORMAT)
}
json.dump(execution_info_dict, out_file, indent=2)
def evaluate_optimizer(self) -> OptimizerEvaluationReport: # pylint: disable=too-many-statements
evaluation_report = OptimizerEvaluationReport(
optimizer_configuration=self.optimizer_config,
objective_function_configuration=self.objective_function_config,
num_optimization_iterations=self.optimizer_evaluator_config.num_iterations,
evaluation_frequency=self.optimizer_evaluator_config.evaluation_frequency,
)
if self.optimizer_evaluator_config.include_execution_trace_in_report:
mlos.global_values.declare_singletons()
if mlos.global_values.tracer is None:
mlos.global_values.tracer = Tracer()
mlos.global_values.tracer.clear_events()
if self.optimizer_evaluator_config.include_pickled_objective_function_in_report:
evaluation_report.pickled_objective_function_initial_state = pickle.dumps(self.objective_function)
if self.optimizer_evaluator_config.include_pickled_optimizer_in_report:
evaluation_report.pickled_optimizer_initial_state = pickle.dumps(self.optimizer)
regression_model_fit_state = RegressionModelFitState()
optima_over_time = {}
optima_over_time[OptimumDefinition.BEST_OBSERVATION.value] = OptimumOverTime(
optimization_problem=self.optimizer.optimization_problem,
optimum_definition=OptimumDefinition.BEST_OBSERVATION
)
optima_over_time[OptimumDefinition.PREDICTED_VALUE_FOR_OBSERVED_CONFIG.value] = OptimumOverTime(
optimization_problem=self.optimizer.optimization_problem,
optimum_definition=OptimumDefinition.PREDICTED_VALUE_FOR_OBSERVED_CONFIG
)
optima_over_time[f"{OptimumDefinition.UPPER_CONFIDENCE_BOUND_FOR_OBSERVED_CONFIG.value}_99"] = OptimumOverTime(
optimization_problem=self.optimizer.optimization_problem,
optimum_definition=OptimumDefinition.UPPER_CONFIDENCE_BOUND_FOR_OBSERVED_CONFIG,
alpha=0.01
)
optima_over_time[f"{OptimumDefinition.LOWER_CONFIDENCE_BOUND_FOR_OBSERVED_CONFIG.value}_99"] = OptimumOverTime(
optimization_problem=self.optimizer.optimization_problem,
optimum_definition=OptimumDefinition.LOWER_CONFIDENCE_BOUND_FOR_OBSERVED_CONFIG,
alpha=0.01
)
#####################################################################################################
i = 0
try:
with traced(scope_name="optimization_loop"):
for i in range(self.optimizer_evaluator_config.num_iterations):
parameters = self.optimizer.suggest()
objectives = self.objective_function.evaluate_point(parameters)
self.optimizer.register(parameters.to_dataframe(), objectives.to_dataframe())
if i % self.optimizer_evaluator_config.evaluation_frequency == 0:
print(f"[{i + 1}/{self.optimizer_evaluator_config.num_iterations}]")
with traced(scope_name="evaluating_optimizer"):
if self.optimizer_evaluator_config.include_pickled_optimizer_in_report:
evaluation_report.add_pickled_optimizer(iteration=i, pickled_optimizer=pickle.dumps(self.optimizer))
if self.optimizer.trained:
gof_metrics = self.optimizer.compute_surrogate_model_goodness_of_fit()
regression_model_fit_state.set_gof_metrics(data_set_type=DataSetType.TRAIN, gof_metrics=gof_metrics)
for optimum_name, optimum_over_time in optima_over_time.items():
try:
config, value = self.optimizer.optimum(
optimum_definition=optimum_over_time.optimum_definition,
alpha=optimum_over_time.alpha
)
optima_over_time[optimum_name].add_optimum_at_iteration(
iteration=i,
optimum_config=config,
optimum_value=value
)
except ValueError as e:
print(e)
evaluation_report.success = True
except Exception as e:
evaluation_report.success = False
evaluation_report.exception = e
evaluation_report.exception_traceback = traceback.format_exc()
with traced(scope_name="evaluating_optimizer"):
"""Once the optimization is done, we performa final evaluation of the optimizer."""
if self.optimizer.trained:
gof_metrics = self.optimizer.compute_surrogate_model_goodness_of_fit()
regression_model_fit_state.set_gof_metrics(data_set_type=DataSetType.TRAIN, gof_metrics=gof_metrics)
for optimum_name, optimum_over_time in optima_over_time.items():
try:
config, value = self.optimizer.optimum(optimum_definition=optimum_over_time.optimum_definition, alpha=optimum_over_time.alpha)
optima_over_time[optimum_name].add_optimum_at_iteration(
iteration=self.optimizer_evaluator_config.num_iterations,
optimum_config=config,
optimum_value=value
)
except Exception as e:
print(e)
if self.optimizer_evaluator_config.include_execution_trace_in_report:
evaluation_report.execution_trace = mlos.global_values.tracer.trace_events
mlos.global_values.tracer.clear_events()
if self.optimizer_evaluator_config.include_pickled_optimizer_in_report:
evaluation_report.add_pickled_optimizer(iteration=i, pickled_optimizer=pickle.dumps(self.optimizer))
if self.optimizer_evaluator_config.include_pickled_objective_function_in_report:
evaluation_report.pickled_objective_function_final_state = pickle.dumps(self.objective_function)
if self.optimizer_evaluator_config.report_regression_model_goodness_of_fit:
evaluation_report.regression_model_goodness_of_fit_state = regression_model_fit_state
if self.optimizer_evaluator_config.report_optima_over_time:
evaluation_report.optima_over_time = optima_over_time
return evaluation_report