blackboxopt.optimizers.testing
Tests that can be imported and used to test optimizer implementations against this packages blackbox optimizer interface.
handles_conditional_space(optimizer_class, optimizer_kwargs)
Check if optimizer handles conditional i.e. hierarchical search spaces.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
optimizer_class |
Union[Type[blackboxopt.base.SingleObjectiveOptimizer], Type[blackboxopt.base.MultiObjectiveOptimizer]] |
Optimizer to test. |
required |
optimizer_kwargs |
dict |
Expected to contain additional arguments for initializing
the optimizer. ( |
required |
Returns:
| Type | Description |
|---|---|
|
Source code in blackboxopt/optimizers/testing.py
def handles_conditional_space(
optimizer_class: Union[
Type[SingleObjectiveOptimizer], Type[MultiObjectiveOptimizer]
],
optimizer_kwargs: dict,
):
"""Check if optimizer handles conditional i.e. hierarchical search spaces.
Args:
optimizer_class: Optimizer to test.
optimizer_kwargs: Expected to contain additional arguments for initializing
the optimizer. (`search_space` and `objective(s)` are set automatically
by the test.)
Returns:
`True` if the test is passed.
"""
space = ps.ParameterSpace()
space.add(ps.CategoricalParameter("optimizer", ("adam", "sgd")))
space.add(ps.ContinuousParameter("lr", (0.0001, 0.1), transformation="log"))
space.add(
ps.ContinuousParameter("momentum", (0.0, 1.0)),
lambda optimizer: optimizer == "sgd",
)
opt = _initialize_optimizer(
optimizer_class,
optimizer_kwargs,
objective=Objective("loss", False),
objectives=[Objective("loss", False)],
space=space,
)
for _ in range(10):
es = opt.generate_evaluation_specification()
dummy_loss = es.configuration.get("momentum", 1.0) * es.configuration["lr"] ** 2
opt.report(es.create_evaluation({"loss": dummy_loss}))
handles_reporting_evaluations_list(optimizer_class, optimizer_kwargs)
Check if optimizer's report method can process an iterable of evaluations.
All optimizers should be able to allow reporting batches of evaluations. It's up to the optimizer's implementation, if evaluations in a batch are processed one by one like if they were reported individually, or if a batch is handled differently.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
optimizer_class |
Union[Type[blackboxopt.base.SingleObjectiveOptimizer], Type[blackboxopt.base.MultiObjectiveOptimizer]] |
Optimizer to test. |
required |
optimizer_kwargs |
dict |
Expected to contain additional arguments for initializing
the optimizer. ( |
required |
Returns:
| Type | Description |
|---|---|
|
Source code in blackboxopt/optimizers/testing.py
def handles_reporting_evaluations_list(
optimizer_class: Union[
Type[SingleObjectiveOptimizer], Type[MultiObjectiveOptimizer]
],
optimizer_kwargs: dict,
):
"""Check if optimizer's report method can process an iterable of evaluations.
All optimizers should be able to allow reporting batches of evaluations. It's up to
the optimizer's implementation, if evaluations in a batch are processed
one by one like if they were reported individually, or if a batch is handled
differently.
Args:
optimizer_class: Optimizer to test.
optimizer_kwargs: Expected to contain additional arguments for initializing
the optimizer. (`search_space` and `objective(s)` are set automatically
by the test.)
Returns:
`True` if the test is passed.
"""
opt = _initialize_optimizer(
optimizer_class,
optimizer_kwargs,
objective=Objective("loss", False),
objectives=[Objective("loss", False)],
seed=42,
)
evaluations = []
for _ in range(3):
es = opt.generate_evaluation_specification()
evaluation = es.create_evaluation(objectives={"loss": 0.42})
evaluations.append(evaluation)
opt.report(evaluations)
is_deterministic_when_reporting_shuffled_evaluations(optimizer_class, optimizer_kwargs)
Check if determinism isn't affected by the order of initially reported data.
Repeatedly initialize the optimizer with the same parameter space and a fixed seed. Report a set of initial evaluations in randomized order as initial data. Start optimizing and check if the generated configurations for all optimizers are equal.
By doing multiple evaluations, this tests covers effects that become visible after a while, e.g. only after stages got completed in staged iteration samplers.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
optimizer_class |
Union[Type[blackboxopt.base.SingleObjectiveOptimizer], Type[blackboxopt.base.MultiObjectiveOptimizer]] |
Optimizer to test. |
required |
optimizer_kwargs |
dict |
Expected to contain additional arguments for initializing
the optimizer. ( |
required |
Returns:
| Type | Description |
|---|---|
|
Source code in blackboxopt/optimizers/testing.py
def is_deterministic_when_reporting_shuffled_evaluations(
optimizer_class: Union[
Type[SingleObjectiveOptimizer], Type[MultiObjectiveOptimizer]
],
optimizer_kwargs: dict,
):
"""Check if determinism isn't affected by the order of initially reported data.
Repeatedly initialize the optimizer with the same parameter space and a fixed seed.
Report a set of initial evaluations in randomized order as initial data. Start
optimizing and check if the generated configurations for all optimizers are equal.
By doing multiple evaluations, this tests covers effects that become visible after
a while, e.g. only after stages got completed in staged iteration samplers.
Args:
optimizer_class: Optimizer to test.
optimizer_kwargs: Expected to contain additional arguments for initializing
the optimizer. (`search_space` and `objective(s)` are set automatically
by the test.)
Returns:
`True` if the test is passed.
"""
space = ps.ParameterSpace()
space.add(ps.ContinuousParameter("p1", (0, 1)))
def _run_experiment_1d(es):
x = es.configuration["p1"]
_x = np.copy(np.atleast_2d(x))
params = np.array([0.75, 0.0, -10.0, 0.0, 0.0])
y = np.polyval(params, _x)
return float(np.squeeze(y))
runs: Dict[int, Dict] = {0: {}, 1: {}}
for run_idx, run in runs.items():
run["evaluations"] = []
opt = _initialize_optimizer(
optimizer_class,
optimizer_kwargs,
objective=Objective("loss", False),
objectives=[Objective("loss", False)],
space=space,
seed=0,
)
# Report initial data in different order
eval_specs = [opt.generate_evaluation_specification() for _ in range(5)]
run["inital_evaluations"] = [
es.create_evaluation(objectives={"loss": _run_experiment_1d(es)})
for es in eval_specs
]
random.seed(run_idx)
random.shuffle(run["inital_evaluations"])
opt.report(run["inital_evaluations"])
# Start optimizing
for _ in range(5):
es = opt.generate_evaluation_specification()
evaluation = es.create_evaluation(
objectives={"loss": _run_experiment_1d(es)}
)
opt.report(evaluation)
run["evaluations"].append(evaluation)
initial_configs_run_0 = [e.configuration for e in runs[0]["inital_evaluations"]]
initial_configs_run_1 = [e.configuration for e in runs[1]["inital_evaluations"]]
configs_run_0 = [e.configuration for e in runs[0]["evaluations"]]
configs_run_1 = [e.configuration for e in runs[1]["evaluations"]]
assert initial_configs_run_0 != initial_configs_run_1
assert configs_run_0 == configs_run_1
is_deterministic_with_fixed_seed_and_larger_space(optimizer_class, optimizer_kwargs)
Check if optimizer is deterministic.
Repeatedly initialize the optimizer with the same parameter space and a fixed seed, get an evaluation specification, report a placeholder result and get another evaluation specification. The configuration of all final evaluation specifications should be equal.
This tests covers multiple parameter types by using a mixed search space.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
optimizer_class |
Union[Type[blackboxopt.base.SingleObjectiveOptimizer], Type[blackboxopt.base.MultiObjectiveOptimizer]] |
Optimizer to test. |
required |
optimizer_kwargs |
dict |
Expected to contain additional arguments for initializing
the optimizer. ( |
required |
Returns:
| Type | Description |
|---|---|
|
Source code in blackboxopt/optimizers/testing.py
def is_deterministic_with_fixed_seed_and_larger_space(
optimizer_class: Union[
Type[SingleObjectiveOptimizer], Type[MultiObjectiveOptimizer]
],
optimizer_kwargs: dict,
):
"""Check if optimizer is deterministic.
Repeatedly initialize the optimizer with the same parameter space and a fixed seed,
get an evaluation specification, report a placeholder result and get another
evaluation specification. The configuration of all final evaluation specifications
should be equal.
This tests covers multiple parameter types by using a mixed search space.
Args:
optimizer_class: Optimizer to test.
optimizer_kwargs: Expected to contain additional arguments for initializing
the optimizer. (`search_space` and `objective(s)` are set automatically
by the test.)
Returns:
`True` if the test is passed.
"""
final_configurations = []
for _ in range(2):
opt = _initialize_optimizer(
optimizer_class,
optimizer_kwargs,
objective=Objective("loss", False),
objectives=[Objective("loss", False)],
)
es1 = opt.generate_evaluation_specification()
evaluation1 = es1.create_evaluation(objectives={"loss": 0.42})
opt.report(evaluation1)
es2 = opt.generate_evaluation_specification()
final_configurations.append(es2.configuration.copy())
assert final_configurations[0] == final_configurations[1]
optimize_single_parameter_sequentially_for_n_max_evaluations(optimizer_class, optimizer_kwargs, n_max_evaluations=20)
[summary]
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
optimizer_class |
Union[Type[blackboxopt.base.SingleObjectiveOptimizer], Type[blackboxopt.base.MultiObjectiveOptimizer]] |
[description] |
required |
optimizer_kwargs |
dict |
[description] |
required |
n_max_evaluations |
int |
[description] |
20 |
Returns:
| Type | Description |
|---|---|
[description] |
Source code in blackboxopt/optimizers/testing.py
def optimize_single_parameter_sequentially_for_n_max_evaluations(
optimizer_class: Union[
Type[SingleObjectiveOptimizer], Type[MultiObjectiveOptimizer]
],
optimizer_kwargs: dict,
n_max_evaluations: int = 20,
):
"""[summary]
Args:
optimizer_class: [description]
optimizer_kwargs: [description]
n_max_evaluations: [description]
Returns:
[description]
"""
def quadratic_function(p1):
return p1**2
assert issubclass(optimizer_class, Optimizer), (
"The default test suite is only applicable for implementations of "
"blackboxopt.base.Optimizer"
)
optimizer = _initialize_optimizer(
optimizer_class,
optimizer_kwargs,
objective=Objective("loss", False),
objectives=[Objective("loss", False), Objective("score", True)],
)
eval_spec = optimizer.generate_evaluation_specification()
if issubclass(optimizer_class, MultiObjectiveOptimizer):
evaluation = eval_spec.create_evaluation(
objectives={"loss": None, "score": None}
)
else:
evaluation = eval_spec.create_evaluation(objectives={"loss": None})
optimizer.report(evaluation)
for _ in range(n_max_evaluations):
try:
eval_spec = optimizer.generate_evaluation_specification()
except OptimizationComplete:
break
loss = quadratic_function(p1=eval_spec.configuration["p1"])
if issubclass(optimizer_class, MultiObjectiveOptimizer):
evaluation_result = {"loss": loss, "score": -loss}
else:
evaluation_result = {"loss": loss}
evaluation = eval_spec.create_evaluation(objectives=evaluation_result)
optimizer.report(evaluation)
raises_evaluation_error_when_reporting_unknown_objective(optimizer_class, optimizer_kwargs)
Check if optimizer's report method raises exception in case objective is unknown.
Also make sure that the faulty evaluations (and only those) are included in the exception.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
optimizer_class |
Union[Type[blackboxopt.base.SingleObjectiveOptimizer], Type[blackboxopt.base.MultiObjectiveOptimizer]] |
Optimizer to test. |
required |
optimizer_kwargs |
dict |
Expected to contain additional arguments for initializing
the optimizer. ( |
required |
Returns:
| Type | Description |
|---|---|
|
Source code in blackboxopt/optimizers/testing.py
def raises_evaluation_error_when_reporting_unknown_objective(
optimizer_class: Union[
Type[SingleObjectiveOptimizer], Type[MultiObjectiveOptimizer]
],
optimizer_kwargs: dict,
):
"""Check if optimizer's report method raises exception in case objective is unknown.
Also make sure that the faulty evaluations (and only those) are included in the
exception.
Args:
optimizer_class: Optimizer to test.
optimizer_kwargs: Expected to contain additional arguments for initializing
the optimizer. (`search_space` and `objective(s)` are set automatically
by the test.)
Returns:
`True` if the test is passed.
"""
opt = _initialize_optimizer(
optimizer_class,
optimizer_kwargs,
objective=Objective("loss", False),
objectives=[Objective("loss", False)],
)
es_1 = opt.generate_evaluation_specification()
es_2 = opt.generate_evaluation_specification()
es_3 = opt.generate_evaluation_specification()
# NOTE: The following is not using pytest.raises because this would add pytest as
# a regular dependency to blackboxopt.
try:
evaluation_1 = es_1.create_evaluation(objectives={"loss": 1})
evaluation_2 = es_2.create_evaluation(objectives={"unknown_objective": 2})
evaluation_3 = es_3.create_evaluation(objectives={"loss": 4})
opt.report([evaluation_1, evaluation_2, evaluation_3])
raise AssertionError(
f"Optimizer {optimizer_class} did not raise an ObjectivesError when a "
+ "result including an unknown objective name was reported."
)
except EvaluationsError as exception:
invalid_evaluations = [e for e, _ in exception.evaluations_with_errors]
assert len(invalid_evaluations) == 1
assert evaluation_2 in invalid_evaluations
respects_fixed_parameter(optimizer_class, optimizer_kwargs)
Check if optimizer's generated evaluation specifications contain the values a parameter in the search space was fixed to.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
optimizer_class |
Union[Type[blackboxopt.base.SingleObjectiveOptimizer], Type[blackboxopt.base.MultiObjectiveOptimizer]] |
Optimizer to test. |
required |
optimizer_kwargs |
dict |
Expected to contain additional arguments for initializing
the optimizer. ( |
required |
Returns:
| Type | Description |
|---|---|
|
Source code in blackboxopt/optimizers/testing.py
def respects_fixed_parameter(
optimizer_class: Union[
Type[SingleObjectiveOptimizer], Type[MultiObjectiveOptimizer]
],
optimizer_kwargs: dict,
):
"""Check if optimizer's generated evaluation specifications contain the values
a parameter in the search space was fixed to.
Args:
optimizer_class: Optimizer to test.
optimizer_kwargs: Expected to contain additional arguments for initializing
the optimizer. (`search_space` and `objective(s)` are set automatically
by the test.)
Returns:
`True` if the test is passed.
"""
space = ps.ParameterSpace()
space.add(ps.ContinuousParameter("my_fixed_param", (-10.0, 200.0)))
space.add(ps.ContinuousParameter("x", (-2.0, 2.0)))
fixed_value = 1.0
space.fix(my_fixed_param=fixed_value)
opt = _initialize_optimizer(
optimizer_class,
optimizer_kwargs,
objective=Objective("loss", False),
objectives=[Objective("loss", False)],
space=space,
)
for _ in range(5):
es = opt.generate_evaluation_specification()
assert es.configuration["my_fixed_param"] == fixed_value
opt.report(
es.create_evaluation(objectives={"loss": es.configuration["x"] ** 2})
)