class TestSmartCacheWithRemoteOptimizer:
""" Tests SmartCache that's being tuned by the remote optimizer.
This test will:
1. Instantiate a SmartCache.
2. Create an MlosExperiment that connects to a remote or in-process optimizer.
3. Optimize the SmartCache with the help of the remote or in-process optimizer.
"""
def setup_method(self, method):
mlos_globals.init_mlos_global_context()
mlos_globals.mlos_global_context.start_clock()
self.logger = create_logger('TestSmartCacheWithRemoteOptimizer')
self.logger.level = logging.DEBUG
# Start up the gRPC service. Try a bunch of times before giving up.
#
max_num_tries = 100
num_tries = 0
for port in range(50051, 50051 + max_num_tries):
num_tries += 1
try:
self.server = OptimizerServicesServer(port=port,
num_threads=10)
self.server.start()
self.port = port
break
except:
self.logger.info(
f"Failed to create OptimizerMicroserviceServer on port {port}"
)
if num_tries == max_num_tries:
raise
self.optimizer_service_channel = grpc.insecure_channel(
f'localhost:{self.port}')
self.bayesian_optimizer_factory = BayesianOptimizerFactory(
grpc_channel=self.optimizer_service_channel, logger=self.logger)
self.mlos_agent = MlosAgent(
logger=self.logger,
communication_channel=mlos_globals.mlos_global_context.
communication_channel,
shared_config=mlos_globals.mlos_global_context.shared_config,
bayesian_optimizer_grpc_channel=self.optimizer_service_channel)
self.mlos_agent_thread = Thread(target=self.mlos_agent.run)
self.mlos_agent_thread.start()
global_values.declare_singletons(
) # TODO: having both globals and global_values is a problem
# Let's add the allowed component types
self.mlos_agent.add_allowed_component_type(SmartCache)
self.mlos_agent.add_allowed_component_type(SmartCacheWorkloadGenerator)
self.mlos_agent.set_configuration(
component_type=SmartCacheWorkloadGenerator,
new_config_values=Point(workload_type='cyclical_key_from_range',
cyclical_key_from_range_config=Point(
min=0, range_width=2048)))
# Let's create the workload
self.smart_cache_workload = SmartCacheWorkloadGenerator(
logger=self.logger)
self.optimizer = None
self.working_set_size_estimator = WorkingSetSizeEstimator()
self.hit_rate_monitor = HitRateMonitor()
self.smart_cache_experiment = MlosExperiment(
smart_component_types=[SmartCache],
telemetry_aggregators=[
self.working_set_size_estimator, self.hit_rate_monitor
])
self.optimization_problem = OptimizationProblem(
parameter_space=SmartCache.parameter_search_space,
objective_space=SimpleHypergrid(name="objectives",
dimensions=[
ContinuousDimension(
name="hit_rate",
min=0,
max=1)
]),
objectives=[Objective(name="hit_rate", minimize=False)])
def teardown_method(self, method):
mlos_globals.mlos_global_context.stop_clock()
self.mlos_agent.stop_all()
self.server.stop(grace=None).wait(timeout=1)
self.server.wait_for_termination(timeout=1)
self.optimizer_service_channel.close()
def test_smart_cache_with_remote_optimizer_on_a_timer(self):
""" Periodically invokes the optimizer to improve cache performance.
"""
optimizer_config = bayesian_optimizer_config_store.default
optimizer_config.homogeneous_random_forest_regression_model_config.decision_tree_regression_model_config.n_new_samples_before_refit = 5
self.optimizer = self.bayesian_optimizer_factory.create_remote_optimizer(
optimization_problem=self.optimization_problem,
optimizer_config=optimizer_config)
self.mlos_agent.start_experiment(self.smart_cache_experiment)
num_iterations = 101
for i in range(num_iterations):
smart_cache_workload_thread = Thread(
target=self.smart_cache_workload.run, args=(0.1, ))
smart_cache_workload_thread.start()
smart_cache_workload_thread.join()
current_cache_config = self.mlos_agent.get_configuration(
component_type=SmartCache)
features_df = current_cache_config.to_dataframe()
hit_rate = self.hit_rate_monitor.get_hit_rate()
num_requests = self.hit_rate_monitor.num_requests
working_set_size_estimate = self.working_set_size_estimator.estimate_working_set_size(
)
objectives_df = pd.DataFrame({'hit_rate': [hit_rate]})
self.optimizer.register(features_df, objectives_df)
new_config_values = self.optimizer.suggest()
self.mlos_agent.set_configuration(
component_type=SmartCache, new_config_values=new_config_values)
self.hit_rate_monitor.reset()
self.logger.info(
f"Previous config: {current_cache_config.to_json()}")
self.logger.info(
f"Estimated working set size: {working_set_size_estimate.chapman_estimator}. Hit rate: {hit_rate:.2f}. Num requests: {num_requests} "
)
self.mlos_agent.stop_experiment(self.smart_cache_experiment)
# Let's look at the goodness of fit.
#
multi_objective_gof_metrics = self.optimizer.compute_surrogate_model_goodness_of_fit(
)
for objective_name, random_forest_gof_metrics in multi_objective_gof_metrics:
# The model might not have used all of the samples, but should have used a majority of them (I expect about 90%), but 70% is a good sanity check
# and should make this test not very flaky.
assert random_forest_gof_metrics.last_refit_iteration_number > 0.5 * num_iterations
# Those relative errors should generally be between 0 and 1 unless the model's predictions are worse than predicting average...
# This unit tests occasionally doesn't have enough data to get us down to 1 so we'll pass the test if its less than 2.
# Note, the point of this test is to check sanity. We'll use a separate suite to evaluate models' performance from an ML standpoint.
self.logger.info(
f"Relative absolute error: {random_forest_gof_metrics.relative_absolute_error}"
)
self.logger.info(
f"Relative squared error: {random_forest_gof_metrics.relative_squared_error}"
)
assert random_forest_gof_metrics.relative_absolute_error is None or (
0 <= random_forest_gof_metrics.relative_absolute_error <= 2)
assert random_forest_gof_metrics.relative_squared_error is None or (
0 <= random_forest_gof_metrics.relative_squared_error <= 2)
# There is an invariant linking mean absolute error (MAE), root mean squared error (RMSE) and number of observations (n) let's assert it.
n = random_forest_gof_metrics.last_refit_iteration_number
self.logger.info(f"Last refit iteration number: {n}")
self.logger.info(
f"Mean absolute error: {random_forest_gof_metrics.mean_absolute_error}"
)
self.logger.info(
f"Root mean squared error: {random_forest_gof_metrics.root_mean_squared_error}"
)
assert random_forest_gof_metrics.mean_absolute_error <= random_forest_gof_metrics.root_mean_squared_error <= math.sqrt(
n) * random_forest_gof_metrics.mean_absolute_error
# We know that the sample confidence interval is wider (or equal to) prediction interval. So hit rates should be ordered accordingly.
assert random_forest_gof_metrics.sample_90_ci_hit_rate >= random_forest_gof_metrics.prediction_90_ci_hit_rate
class TestBayesianOptimizerGrpcClient(unittest.TestCase):
""" Tests the E2E Grpc Client-Service workflow.
"""
@classmethod
def setUpClass(cls):
warnings.simplefilter("error")
global_values.declare_singletons()
def setUp(self):
self.logger = create_logger(self.__class__.__name__)
# Start up the gRPC service.
#
self.server = OptimizerMicroserviceServer(port=50051, num_threads=10)
self.server.start()
self.optimizer_service_channel = grpc.insecure_channel('localhost:50051')
self.bayesian_optimizer_factory = BayesianOptimizerFactory(grpc_channel=self.optimizer_service_channel, logger=self.logger)
self.optimizer_monitor = OptimizerMonitor(grpc_channel=self.optimizer_service_channel, logger=self.logger)
objective_function_config = objective_function_config_store.get_config_by_name('2d_quadratic_concave_up')
self.objective_function = ObjectiveFunctionFactory.create_objective_function(objective_function_config)
self.optimization_problem = OptimizationProblem(
parameter_space=self.objective_function.parameter_space,
objective_space=self.objective_function.output_space,
objectives=[Objective(name='y', minimize=True)]
)
def tearDown(self):
""" We need to tear down the gRPC server here.
:return:
"""
self.server.stop(grace=None)
def test_echo(self):
optimizer_service_stub = OptimizerServiceStub(channel=self.optimizer_service_channel)
response = optimizer_service_stub.Echo(Empty())
self.assertTrue(isinstance(response, Empty))
def test_optimizer_with_default_config(self):
pre_existing_optimizers = {optimizer.id: optimizer for optimizer in self.optimizer_monitor.get_existing_optimizers()}
print(bayesian_optimizer_config_store.default)
bayesian_optimizer = self.bayesian_optimizer_factory.create_remote_optimizer(
optimization_problem=self.optimization_problem,
optimizer_config=bayesian_optimizer_config_store.default
)
post_existing_optimizers = {optimizer.id: optimizer for optimizer in self.optimizer_monitor.get_existing_optimizers()}
new_optimizers = {
optimizer_id: optimizer
for optimizer_id, optimizer in post_existing_optimizers.items()
if optimizer_id not in pre_existing_optimizers
}
self.assertTrue(len(new_optimizers) == 1)
new_optimizer_id = list(new_optimizers.keys())[0]
new_optimizer = new_optimizers[new_optimizer_id]
self.assertTrue(new_optimizer_id == bayesian_optimizer.id)
self.assertTrue(new_optimizer.optimizer_config == bayesian_optimizer.optimizer_config)
num_iterations = 100
registered_features_df, registered_objectives_df = self.optimize_quadratic(optimizer=bayesian_optimizer, num_iterations=num_iterations)
# Apparently the to_json/from_json loses precision so we explicitly lose it here so that we can do the comparison.
#
registered_features_json = registered_features_df.to_json(orient='index', double_precision=15)
registered_objectives_json = registered_objectives_df.to_json(orient='index', double_precision=15)
# Apparently the jitter is too good and we actually have to use the json strings or they will be optimized away.
#
assert len(registered_features_json) > 0
assert len(registered_objectives_json) > 0
registered_features_df = pd.read_json(registered_features_json, orient='index')
registered_objectives_df = pd.read_json(registered_objectives_json, orient='index')
observed_features_df, observed_objectives_df = bayesian_optimizer.get_all_observations()
self.assertTrue((np.abs(registered_features_df - observed_features_df) < 0.00000001).all().all())
self.assertTrue((np.abs(registered_objectives_df - observed_objectives_df) < 0.00000001).all().all())
# Let's look at the goodness of fit.
#
random_forest_gof_metrics = bayesian_optimizer.compute_surrogate_model_goodness_of_fit()
# The model might not have used all of the samples, but should have used a majority of them (I expect about 90%), but 70% is a good sanity check
# and should make this test not very flaky.
self.assertTrue(random_forest_gof_metrics.last_refit_iteration_number > 0.7 * num_iterations)
# The invariants below should be true for all surrogate models: the random forest, and all constituent decision trees. So let's iterate over them all.
models_gof_metrics = [random_forest_gof_metrics]
for model_gof_metrics in models_gof_metrics:
self.assertTrue(0 <= model_gof_metrics.relative_absolute_error <= 1) # This could fail if the models are really wrong. Not expected in this unit test though.
self.assertTrue(0 <= model_gof_metrics.relative_squared_error <= 1)
# There is an invariant linking mean absolute error (MAE), root mean squared error (RMSE) and number of observations (n) let's assert it.
n = model_gof_metrics.last_refit_iteration_number
self.assertTrue(model_gof_metrics.mean_absolute_error <= model_gof_metrics.root_mean_squared_error <= math.sqrt(n) * model_gof_metrics.mean_absolute_error)
# We know that the sample confidence interval is wider (or equal to) prediction interval. So hit rates should be ordered accordingly.
self.assertTrue(model_gof_metrics.sample_90_ci_hit_rate >= model_gof_metrics.prediction_90_ci_hit_rate)
self.assertTrue(0 <= model_gof_metrics.coefficient_of_determination <= 1)
def test_optimizer_with_random_config(self):
num_random_restarts = 10
for i in range(num_random_restarts):
optimizer_config = bayesian_optimizer_config_store.parameter_space.random()
optimizer_config.min_samples_required_for_guided_design_of_experiments = min(optimizer_config.min_samples_required_for_guided_design_of_experiments, 100)
if optimizer_config.surrogate_model_implementation == "HomogeneousRandomForestRegressionModel":
rf_config = optimizer_config.homogeneous_random_forest_regression_model_config
rf_config.n_estimators = min(rf_config.n_estimators, 20)
print(f"[{i+1}/{num_random_restarts}] Creating a bayesian optimizer with config: {optimizer_config}")
bayesian_optimizer = self.bayesian_optimizer_factory.create_remote_optimizer(
optimization_problem=self.optimization_problem,
optimizer_config=optimizer_config
)
registered_features_df, registered_objectives_df = self.optimize_quadratic(optimizer=bayesian_optimizer, num_iterations=12)
# Apparently the to_json/from_json loses precision so we explicitly lose it here so that we can do the comparison.
#
registered_features_json = registered_features_df.to_json(orient='index', double_precision=15)
registered_objectives_json = registered_objectives_df.to_json(orient='index', double_precision=15)
# Apparently the jitter is too good and we actually have to use the json strings or they will be optimized away.
#
assert len(registered_features_json) > 0
assert len(registered_objectives_json) > 0
registered_features_df = pd.read_json(registered_features_json, orient='index')
registered_objectives_df = pd.read_json(registered_objectives_json, orient='index')
observed_features_df, observed_objectives_df = bayesian_optimizer.get_all_observations()
self.assertTrue((np.abs(registered_features_df - observed_features_df) < 0.00000001).all().all())
self.assertTrue((np.abs(registered_objectives_df - observed_objectives_df) < 0.00000001).all().all())
@unittest.skip(reason="Not implemented yet.")
def test_optimizer_with_named_config(self):
...
def optimize_quadratic(self, optimizer, num_iterations):
registered_features_df = None
registered_objectives_df = None
old_optimum = np.inf
for i in range(num_iterations):
suggested_params = optimizer.suggest()
suggested_params_df = suggested_params.to_dataframe()
y = self.objective_function.evaluate_point(suggested_params)
optimizer.register(suggested_params_df, y.to_dataframe())
if registered_features_df is None:
registered_features_df = suggested_params_df
else:
registered_features_df = registered_features_df.append(suggested_params_df, ignore_index=True)
if registered_objectives_df is None:
registered_objectives_df = y.to_dataframe()
else:
registered_objectives_df = registered_objectives_df.append(y.to_dataframe(), ignore_index=True)
best_params, optimum = optimizer.optimum()
# ensure current optimum doesn't go up
assert optimum.y <= old_optimum
old_optimum = optimum.y
print(f"[{i+1}/{num_iterations}]Best Params: {best_params}, Best Value: {optimum.y}")
return registered_features_df, registered_objectives_df
class TestBayesianOptimizerGrpcClient:
""" Tests the E2E Grpc Client-Service workflow.
"""
@classmethod
def setup_class(cls):
warnings.simplefilter("error")
global_values.declare_singletons()
def setup_method(self, method):
self.logger = create_logger(self.__class__.__name__)
# Start up the gRPC service. Try a bunch of times before giving up.
#
max_num_tries = 100
num_tries = 0
for port in range(50051, 50051 + max_num_tries):
num_tries += 1
try:
self.server = OptimizerServicesServer(port=port,
num_threads=10)
self.server.start()
self.port = port
break
except:
self.logger.info(
f"Failed to create OptimizerMicroserviceServer on port {port}"
)
if num_tries == max_num_tries:
raise
self.optimizer_service_channel = grpc.insecure_channel(
f'localhost:{self.port}')
self.bayesian_optimizer_factory = BayesianOptimizerFactory(
grpc_channel=self.optimizer_service_channel, logger=self.logger)
self.optimizer_monitor = OptimizerMonitor(
grpc_channel=self.optimizer_service_channel, logger=self.logger)
objective_function_config = objective_function_config_store.get_config_by_name(
'2d_quadratic_concave_up')
self.objective_function = ObjectiveFunctionFactory.create_objective_function(
objective_function_config)
self.optimization_problem = OptimizationProblem(
parameter_space=self.objective_function.parameter_space,
objective_space=self.objective_function.output_space,
objectives=[Objective(name='y', minimize=True)])
def teardown_method(self, method):
""" We need to tear down the gRPC server here.
:return:
"""
self.server.stop(grace=None).wait(timeout=1)
self.server.wait_for_termination(timeout=1)
self.optimizer_service_channel.close()
def test_echo(self):
optimizer_service_stub = OptimizerServiceStub(
channel=self.optimizer_service_channel)
response = optimizer_service_stub.Echo(Empty())
assert isinstance(response, Empty)
def test_optimizer_with_default_config(self):
pre_existing_optimizers = {
optimizer.id: optimizer
for optimizer in self.optimizer_monitor.get_existing_optimizers()
}
print(bayesian_optimizer_config_store.default)
bayesian_optimizer = self.bayesian_optimizer_factory.create_remote_optimizer(
optimization_problem=self.optimization_problem,
optimizer_config=bayesian_optimizer_config_store.default)
post_existing_optimizers = {
optimizer.id: optimizer
for optimizer in self.optimizer_monitor.get_existing_optimizers()
}
new_optimizers = {
optimizer_id: optimizer
for optimizer_id, optimizer in post_existing_optimizers.items()
if optimizer_id not in pre_existing_optimizers
}
assert len(new_optimizers) == 1
new_optimizer_id = list(new_optimizers.keys())[0]
new_optimizer = new_optimizers[new_optimizer_id]
assert new_optimizer_id == bayesian_optimizer.id
assert new_optimizer.optimizer_config == bayesian_optimizer.optimizer_config
num_iterations = 100
registered_features_df, registered_objectives_df = self.optimize_objective_function(
optimizer=bayesian_optimizer,
objective_function=self.objective_function,
num_iterations=num_iterations)
# Apparently the to_json/from_json loses precision so we explicitly lose it here so that we can do the comparison.
#
registered_features_json = registered_features_df.to_json(
orient='index', double_precision=15)
registered_objectives_json = registered_objectives_df.to_json(
orient='index', double_precision=15)
# Apparently the jitter is too good and we actually have to use the json strings or they will be optimized away.
#
assert len(registered_features_json) > 0
assert len(registered_objectives_json) > 0
registered_features_df = pd.read_json(registered_features_json,
orient='index')
registered_objectives_df = pd.read_json(registered_objectives_json,
orient='index')
observed_features_df, observed_objectives_df, _ = bayesian_optimizer.get_all_observations(
)
assert (np.abs(registered_features_df - observed_features_df) <
0.00000001).all().all()
assert (np.abs(registered_objectives_df - observed_objectives_df) <
0.00000001).all().all()
# Assert that the observations and predictions are returned in the right order from the remote optimizer
#
parameters_df, objectives_df, _ = bayesian_optimizer.get_all_observations(
)
predictions_df = bayesian_optimizer.predict(
parameter_values_pandas_frame=parameters_df).get_dataframe()
assert parameters_df.index.intersection(predictions_df.index).equals(
predictions_df.index)
# Let's look at the goodness of fit.
#
multi_objective_gof_metrics = bayesian_optimizer.compute_surrogate_model_goodness_of_fit(
)
for objective_name, random_forest_gof_metrics in multi_objective_gof_metrics:
# The model might not have used all of the samples, but should have used a majority of them (I expect about 90%), but 50% is a good sanity check
# and should make this test not very flaky.
assert random_forest_gof_metrics.last_refit_iteration_number > 0.5 * num_iterations
# The invariants below should be true for all surrogate models: the random forest, and all constituent decision trees. So let's iterate over them all.
models_gof_metrics = [random_forest_gof_metrics]
for model_gof_metrics in models_gof_metrics:
assert 0 <= model_gof_metrics.relative_absolute_error <= 1 # This could fail if the models are really wrong. Not expected in this unit test though.
assert 0 <= model_gof_metrics.relative_squared_error <= 1
# There is an invariant linking mean absolute error (MAE), root mean squared error (RMSE) and number of observations (n) let's assert it.
n = model_gof_metrics.last_refit_iteration_number
assert model_gof_metrics.mean_absolute_error <= model_gof_metrics.root_mean_squared_error <= math.sqrt(
n) * model_gof_metrics.mean_absolute_error
# We know that the sample confidence interval is wider (or equal to) prediction interval. So hit rates should be ordered accordingly.
assert model_gof_metrics.sample_90_ci_hit_rate >= model_gof_metrics.prediction_90_ci_hit_rate
assert 0 <= model_gof_metrics.coefficient_of_determination <= 1
@pytest.mark.parametrize("i", [i for i in range(10)])
def test_optimizer_with_random_config(self, i):
optimizer_config = bayesian_optimizer_config_store.parameter_space.random(
)
optimizer_config.min_samples_required_for_guided_design_of_experiments = max(
min(
optimizer_config.
min_samples_required_for_guided_design_of_experiments, 100),
20)
if optimizer_config.surrogate_model_implementation == "HomogeneousRandomForestRegressionModel":
rf_config = optimizer_config.homogeneous_random_forest_regression_model_config
rf_config.n_estimators = min(rf_config.n_estimators, 20)
if optimizer_config.surrogate_model_implementation == MultiObjectiveRegressionEnhancedRandomForest.__name__:
optimizer_config.min_samples_required_for_guided_design_of_experiments = 25
rerf_model_config = optimizer_config.regression_enhanced_random_forest_regression_model_config
rerf_model_config.max_basis_function_degree = min(
rerf_model_config.max_basis_function_degree, 2)
# increased polynomial degree requires more data to estimate model parameters (poly term coefficients)
optimizer_config.min_samples_required_for_guided_design_of_experiments += 25 * (
rerf_model_config.max_basis_function_degree - 1)
rf_model_config = rerf_model_config.sklearn_random_forest_regression_model_config
rf_model_config.perform_initial_random_forest_hyper_parameter_search = False
rf_model_config.max_depth = min(rf_model_config.max_depth, 10)
rf_model_config.n_jobs = min(rf_model_config.n_jobs, 4)
print(
f"[{i+1}] Creating a bayesian optimizer with config: {optimizer_config}"
)
bayesian_optimizer = self.bayesian_optimizer_factory.create_remote_optimizer(
optimization_problem=self.optimization_problem,
optimizer_config=optimizer_config)
registered_features_df, registered_objectives_df = self.optimize_objective_function(
optimizer=bayesian_optimizer,
objective_function=self.objective_function,
num_iterations=12)
# Apparently the to_json/from_json loses precision so we explicitly lose it here so that we can do the comparison.
#
registered_features_json = registered_features_df.to_json(
orient='index', double_precision=15)
registered_objectives_json = registered_objectives_df.to_json(
orient='index', double_precision=15)
# Apparently the jitter is too good and we actually have to use the json strings or they will be optimized away.
#
assert len(registered_features_json) > 0
assert len(registered_objectives_json) > 0
registered_features_df = pd.read_json(registered_features_json,
orient='index')
registered_objectives_df = pd.read_json(registered_objectives_json,
orient='index')
observed_features_df, observed_objectives_df, _ = bayesian_optimizer.get_all_observations(
)
assert (np.abs(registered_features_df - observed_features_df) <
0.00000001).all().all()
assert (np.abs(registered_objectives_df - observed_objectives_df) <
0.00000001).all().all()
@pytest.mark.parametrize("i", [i for i in range(10)])
def test_optimizer_with_random_config_random_objective(self, i):
objective_function_config = objective_function_config_store.parameter_space.random(
)
objective_function = ObjectiveFunctionFactory.create_objective_function(
objective_function_config)
optimization_problem = objective_function.default_optimization_problem
optimizer_config = bayesian_optimizer_config_store.parameter_space.random(
)
optimizer_config.min_samples_required_for_guided_design_of_experiments = max(
min(
optimizer_config.
min_samples_required_for_guided_design_of_experiments, 100),
20)
if optimizer_config.surrogate_model_implementation == "HomogeneousRandomForestRegressionModel":
rf_config = optimizer_config.homogeneous_random_forest_regression_model_config
rf_config.n_estimators = min(rf_config.n_estimators, 20)
if optimizer_config.surrogate_model_implementation == MultiObjectiveRegressionEnhancedRandomForest.__name__:
optimizer_config.min_samples_required_for_guided_design_of_experiments = 25
rerf_model_config = optimizer_config.regression_enhanced_random_forest_regression_model_config
rerf_model_config.max_basis_function_degree = min(
rerf_model_config.max_basis_function_degree, 2)
# increased polynomial degree requires more data to estimate model parameters (poly term coefficients)
optimizer_config.min_samples_required_for_guided_design_of_experiments += 25 * (
rerf_model_config.max_basis_function_degree - 1)
rf_model_config = rerf_model_config.sklearn_random_forest_regression_model_config
rf_model_config.perform_initial_random_forest_hyper_parameter_search = False
rf_model_config.max_depth = min(rf_model_config.max_depth, 10)
rf_model_config.n_jobs = min(rf_model_config.n_jobs, 4)
print(
f"[{i+1}] Creating a bayesian optimizer with config: {optimizer_config} \n\n\nObjective function config: {objective_function_config}"
)
bayesian_optimizer = self.bayesian_optimizer_factory.create_remote_optimizer(
optimization_problem=optimization_problem,
optimizer_config=optimizer_config)
registered_params_df, registered_objectives_df = self.optimize_objective_function(
optimizer=bayesian_optimizer,
objective_function=objective_function,
num_iterations=20)
# Apparently the to_json/from_json loses precision so we explicitly lose it here so that we can do the comparison.
#
registered_features_json = registered_params_df.to_json(
orient='index', double_precision=15)
registered_objectives_json = registered_objectives_df.to_json(
orient='index', double_precision=15)
# Apparently the jitter is too good and we actually have to use the json strings or they will be optimized away.
#
assert len(registered_features_json) > 0
assert len(registered_objectives_json) > 0
registered_params_df = pd.read_json(registered_features_json,
orient='index')
registered_objectives_df = pd.read_json(registered_objectives_json,
orient='index')
observed_params_df, observed_objectives_df, _ = bayesian_optimizer.get_all_observations(
)
numeric_params_names = [
dimension.name
for dimension in optimization_problem.parameter_space.dimensions
if (isinstance(dimension, (ContinuousDimension, DiscreteDimension))
or (isinstance(dimension, CategoricalDimension)
and dimension.is_numeric)) and (
dimension.name in registered_params_df.columns) and (
dimension.name in observed_params_df.columns)
]
numeric_params_df = registered_params_df[numeric_params_names]
observed_numeric_params_df = observed_params_df[numeric_params_names]
assert (np.abs(
numeric_params_df.fillna(0) - observed_numeric_params_df.fillna(0))
< 0.00000001).all().all()
assert (np.abs(registered_objectives_df - observed_objectives_df) <
0.00000001).all().all()
@staticmethod
def optimize_objective_function(optimizer, objective_function,
num_iterations):
registered_features_df = None
registered_objectives_df = None
# Let's make sure that the optimum for the first objective doesn't get worse throughout the optimization loop.
#
first_objective = optimizer.optimization_problem.objectives[0]
if first_objective.minimize:
old_optimum = np.inf
else:
old_optimum = -np.inf
for i in range(num_iterations):
suggested_params = optimizer.suggest()
suggested_params_df = suggested_params.to_dataframe()
objective_values = objective_function.evaluate_point(
suggested_params)
optimizer.register(suggested_params_df,
objective_values.to_dataframe())
if registered_features_df is None:
registered_features_df = suggested_params_df
else:
registered_features_df = registered_features_df.append(
suggested_params_df, ignore_index=True)
if registered_objectives_df is None:
registered_objectives_df = objective_values.to_dataframe()
else:
registered_objectives_df = registered_objectives_df.append(
objective_values.to_dataframe(), ignore_index=True)
best_params, optimum = optimizer.optimum()
# ensure current optimum doesn't go up
#
if first_objective.minimize:
assert optimum[first_objective.name] <= old_optimum
else:
assert optimum[first_objective.name] >= old_optimum
old_optimum = optimum[first_objective.name]
print(
f"[{i+1}/{num_iterations}]Best Params: {best_params}, Best Value: {optimum[first_objective.name]}"
)
return registered_features_df, registered_objectives_df