def test_init_cma_opts():
# type: () -> None
sampler = optuna.integration.CmaEsSampler(
x0={"x": 0, "y": 0},
sigma0=0.1,
cma_stds={"x": 1, "y": 1},
seed=1,
cma_opts={"popsize": 5},
independent_sampler=DeterministicRelativeSampler({}, {}),
)
study = optuna.create_study(sampler=sampler)
with patch("optuna.integration.cma._Optimizer") as mock_obj:
mock_obj.ask.return_value = {"x": -1, "y": -1}
study.optimize(
lambda t: t.suggest_int("x", -1, 1) + t.suggest_int("y", -1, 1), n_trials=2
)
assert mock_obj.mock_calls[0] == call(
{
"x": IntUniformDistribution(low=-1, high=1),
"y": IntUniformDistribution(low=-1, high=1),
},
{"x": 0, "y": 0},
0.1,
{"x": 1, "y": 1},
{"popsize": 5, "seed": 1, "verbose": -2},
)
def suggest_int(self,
name: str,
low: int,
high: int,
step: int = 1,
log: bool = False) -> int:
if step != 1:
if log:
raise ValueError(
"The parameter `step != 1` is not supported when `log` is True."
"The specified `step` is {}.".format(step))
else:
distribution: Union[
IntUniformDistribution,
IntLogUniformDistribution] = IntUniformDistribution(
low=low, high=high, step=step)
else:
if log:
distribution = IntLogUniformDistribution(low=low, high=high)
else:
distribution = IntUniformDistribution(low=low,
high=high,
step=step)
return int(self._suggest(name, distribution))
def test_init_cma_opts():
# type: () -> None
sampler = optuna.integration.CmaEsSampler(
x0={'x': 0, 'y': 0},
sigma0=0.1,
cma_stds={'x': 1, 'y': 1},
seed=1,
cma_opts={'popsize': 5},
independent_sampler=DeterministicRelativeSampler({}, {}))
study = optuna.create_study(sampler=sampler)
with patch('optuna.integration.cma._Optimizer') as mock_obj:
mock_obj.ask.return_value = {'x': -1, 'y': -1}
study.optimize(
lambda t: t.suggest_int('x', -1, 1) + t.suggest_int('y', -1, 1), n_trials=2)
assert mock_obj.mock_calls[0] == call({
'x': IntUniformDistribution(low=-1, high=1),
'y': IntUniformDistribution(low=-1, high=1)
}, {
'x': 0,
'y': 0
}, 0.1, {
'x': 1,
'y': 1
}, {
'popsize': 5,
'seed': 1,
'verbose': -2
})
def search_space() -> Dict[str, BaseDistribution]:
return {
"c": CategoricalDistribution(("a", "b")),
"d": DiscreteUniformDistribution(-1, 9, 2),
"i": IntUniformDistribution(-1, 1),
"ii": IntUniformDistribution(-1, 3, 2),
"l": LogUniformDistribution(0.001, 0.1),
"u": UniformDistribution(-2, 2),
}
def test_intersection_search_space() -> None:
search_space = optuna.samplers.IntersectionSearchSpace()
study = optuna.create_study()
# No trial.
assert search_space.calculate(study) == {}
assert search_space.calculate(study) == optuna.samplers.intersection_search_space(study)
# First trial.
study.optimize(lambda t: t.suggest_uniform("y", -3, 3) + t.suggest_int("x", 0, 10), n_trials=1)
assert search_space.calculate(study) == {
"x": IntUniformDistribution(low=0, high=10),
"y": UniformDistribution(low=-3, high=3),
}
assert search_space.calculate(study) == optuna.samplers.intersection_search_space(study)
# Returning sorted `OrderedDict` instead of `dict`.
assert search_space.calculate(study, ordered_dict=True) == OrderedDict(
[
("x", IntUniformDistribution(low=0, high=10)),
("y", UniformDistribution(low=-3, high=3)),
]
)
assert search_space.calculate(
study, ordered_dict=True
) == optuna.samplers.intersection_search_space(study, ordered_dict=True)
# Second trial (only 'y' parameter is suggested in this trial).
study.optimize(lambda t: t.suggest_uniform("y", -3, 3), n_trials=1)
assert search_space.calculate(study) == {"y": UniformDistribution(low=-3, high=3)}
assert search_space.calculate(study) == optuna.samplers.intersection_search_space(study)
# Failed or pruned trials are not considered in the calculation of
# an intersection search space.
def objective(trial, exception):
# type: (optuna.trial.Trial, Exception) -> float
trial.suggest_uniform("z", 0, 1)
raise exception
study.optimize(lambda t: objective(t, RuntimeError()), n_trials=1, catch=(RuntimeError,))
study.optimize(lambda t: objective(t, optuna.exceptions.TrialPruned()), n_trials=1)
assert search_space.calculate(study) == {"y": UniformDistribution(low=-3, high=3)}
assert search_space.calculate(study) == optuna.samplers.intersection_search_space(study)
# If two parameters have the same name but different distributions,
# those are regarded as different parameters.
study.optimize(lambda t: t.suggest_uniform("y", -1, 1), n_trials=1)
assert search_space.calculate(study) == {}
assert search_space.calculate(study) == optuna.samplers.intersection_search_space(study)
# The search space remains empty once it is empty.
study.optimize(lambda t: t.suggest_uniform("y", -3, 3) + t.suggest_int("x", 0, 10), n_trials=1)
assert search_space.calculate(study) == {}
assert search_space.calculate(study) == optuna.samplers.intersection_search_space(study)
def test_not_contained_param() -> None:
trial = create_trial(
value=0.2,
params={"x": 1.0},
distributions={"x": UniformDistribution(1.0, 10.0)},
)
with pytest.warns(UserWarning):
assert trial.suggest_float("x", 10.0, 100.0) == 1.0
trial = create_trial(
value=0.2,
params={"x": 1.0},
distributions={"x": LogUniformDistribution(1.0, 10.0)},
)
with pytest.warns(UserWarning):
assert trial.suggest_float("x", 10.0, 100.0, log=True) == 1.0
trial = create_trial(
value=0.2,
params={"x": 1.0},
distributions={"x": DiscreteUniformDistribution(1.0, 10.0, 1.0)},
)
with pytest.warns(UserWarning):
assert trial.suggest_float("x", 10.0, 100.0, step=1.0) == 1.0
trial = create_trial(
value=0.2,
params={"x": 1.0},
distributions={"x": IntUniformDistribution(1, 10)},
)
with pytest.warns(UserWarning):
assert trial.suggest_int("x", 10, 100) == 1
trial = create_trial(
value=0.2,
params={"x": 1},
distributions={"x": IntUniformDistribution(1, 10, 1)},
)
with pytest.warns(UserWarning):
assert trial.suggest_int("x", 10, 100, 1) == 1
trial = create_trial(
value=0.2,
params={"x": 1},
distributions={"x": IntLogUniformDistribution(1, 10)},
)
with pytest.warns(UserWarning):
assert trial.suggest_int("x", 10, 100, log=True) == 1
def test_distributions(storage_init_func):
# type: (Callable[[], storages.BaseStorage]) -> None
def objective(trial):
# type: (Trial) -> float
trial.suggest_uniform("a", 0, 10)
trial.suggest_loguniform("b", 0.1, 10)
trial.suggest_discrete_uniform("c", 0, 10, 1)
trial.suggest_int("d", 0, 10)
trial.suggest_categorical("e", ["foo", "bar", "baz"])
trial.suggest_int("f", 1, 10, log=True)
return 1.0
study = create_study(storage_init_func())
study.optimize(objective, n_trials=1)
assert study.best_trial.distributions == {
"a": UniformDistribution(low=0, high=10),
"b": LogUniformDistribution(low=0.1, high=10),
"c": DiscreteUniformDistribution(low=0, high=10, q=1),
"d": IntUniformDistribution(low=0, high=10),
"e": CategoricalDistribution(choices=("foo", "bar", "baz")),
"f": IntLogUniformDistribution(low=1, high=10),
}
def test_sample_single_distribution(
sampler_class: Callable[[], BaseSampler]) -> None:
relative_search_space = {
"a": UniformDistribution(low=1.0, high=1.0),
"b": LogUniformDistribution(low=1.0, high=1.0),
"c": DiscreteUniformDistribution(low=1.0, high=1.0, q=1.0),
"d": IntUniformDistribution(low=1, high=1),
"e": IntLogUniformDistribution(low=1, high=1),
"f": CategoricalDistribution([1]),
"g": FloatDistribution(low=1.0, high=1.0),
"h": FloatDistribution(low=1.0, high=1.0, log=True),
"i": FloatDistribution(low=1.0, high=1.0, step=1.0),
"j": IntDistribution(low=1, high=1),
"k": IntDistribution(low=1, high=1, log=True),
}
with warnings.catch_warnings():
warnings.simplefilter("ignore", optuna.exceptions.ExperimentalWarning)
sampler = sampler_class()
study = optuna.study.create_study(sampler=sampler)
# We need to test the construction of the model, so we should set `n_trials >= 2`.
for _ in range(2):
trial = study.ask(fixed_distributions=relative_search_space)
study.tell(trial, 1.0)
for param_name in relative_search_space.keys():
assert trial.params[param_name] == 1
def test_sample_relative():
# type: () -> None
relative_search_space = {
'a': UniformDistribution(low=0, high=5),
'b': CategoricalDistribution(choices=('foo', 'bar', 'baz')),
'c': IntUniformDistribution(low=20, high=50), # Not exist in `relative_params`.
}
relative_params = {
'a': 3.2,
'b': 'baz',
}
unknown_param_value = 30
sampler = FixedSampler( # type: ignore
relative_search_space, relative_params, unknown_param_value)
study = optuna.study.create_study(sampler=sampler)
def objective(trial):
# type: (Trial) -> float
# Predefined parameters are sampled by `sample_relative()` method.
assert trial.suggest_uniform('a', 0, 5) == 3.2
assert trial.suggest_categorical('b', ['foo', 'bar', 'baz']) == 'baz'
# Other parameters are sampled by `sample_independent()` method.
assert trial.suggest_int('c', 20, 50) == unknown_param_value
assert trial.suggest_loguniform('d', 1, 100) == unknown_param_value
assert trial.suggest_uniform('e', 20, 40) == unknown_param_value
return 0.0
study.optimize(objective, n_trials=10, catch=())
for trial in study.trials:
assert trial.params == {'a': 3.2, 'b': 'baz', 'c': 30, 'd': 30, 'e': 30}
def create_optuna_distribution_from_override(override: Override) -> Any:
value = override.value()
if not override.is_sweep_override():
return value
if override.is_choice_sweep():
assert isinstance(value, ChoiceSweep)
choices = [
x for x in override.sweep_iterator(transformer=Transformer.encode)
]
return CategoricalDistribution(choices)
if override.is_range_sweep():
choices = [
x for x in override.sweep_iterator(transformer=Transformer.encode)
]
return CategoricalDistribution(choices)
if override.is_interval_sweep():
assert isinstance(value, IntervalSweep)
if "log" in value.tags:
if "int" in value.tags:
return IntLogUniformDistribution(value.start, value.end)
return LogUniformDistribution(value.start, value.end)
else:
if "int" in value.tags:
return IntUniformDistribution(value.start, value.end)
return UniformDistribution(value.start, value.end)
raise NotImplementedError(
"{} is not supported by Optuna sweeper.".format(override))
def test_sample_relative() -> None:
relative_search_space: Dict[str, BaseDistribution] = {
"a": UniformDistribution(low=0, high=5),
"b": CategoricalDistribution(choices=("foo", "bar", "baz")),
"c": IntUniformDistribution(low=20, high=50), # Not exist in `relative_params`.
}
relative_params = {
"a": 3.2,
"b": "baz",
}
unknown_param_value = 30
sampler = FixedSampler( # type: ignore
relative_search_space, relative_params, unknown_param_value
)
study = optuna.study.create_study(sampler=sampler)
def objective(trial: Trial) -> float:
# Predefined parameters are sampled by `sample_relative()` method.
assert trial.suggest_uniform("a", 0, 5) == 3.2
assert trial.suggest_categorical("b", ["foo", "bar", "baz"]) == "baz"
# Other parameters are sampled by `sample_independent()` method.
assert trial.suggest_int("c", 20, 50) == unknown_param_value
assert trial.suggest_loguniform("d", 1, 100) == unknown_param_value
assert trial.suggest_uniform("e", 20, 40) == unknown_param_value
return 0.0
study.optimize(objective, n_trials=10, catch=())
for trial in study.trials:
assert trial.params == {"a": 3.2, "b": "baz", "c": 30, "d": 30, "e": 30}
def test_search_space_transform_encoding() -> None:
trans = _SearchSpaceTransform({"x0": IntUniformDistribution(0, 3)})
assert len(trans.column_to_encoded_columns) == 1
numpy.testing.assert_equal(trans.column_to_encoded_columns[0], numpy.array([0]))
numpy.testing.assert_equal(trans.encoded_column_to_column, numpy.array([0]))
trans = _SearchSpaceTransform({"x0": CategoricalDistribution(["foo", "bar", "baz"])})
assert len(trans.column_to_encoded_columns) == 1
numpy.testing.assert_equal(trans.column_to_encoded_columns[0], numpy.array([0, 1, 2]))
numpy.testing.assert_equal(trans.encoded_column_to_column, numpy.array([0, 0, 0]))
trans = _SearchSpaceTransform(
{
"x0": UniformDistribution(0, 3),
"x1": CategoricalDistribution(["foo", "bar", "baz"]),
"x3": DiscreteUniformDistribution(0, 1, q=0.2),
}
)
assert len(trans.column_to_encoded_columns) == 3
numpy.testing.assert_equal(trans.column_to_encoded_columns[0], numpy.array([0]))
numpy.testing.assert_equal(trans.column_to_encoded_columns[1], numpy.array([1, 2, 3]))
numpy.testing.assert_equal(trans.column_to_encoded_columns[2], numpy.array([4]))
numpy.testing.assert_equal(trans.encoded_column_to_column, numpy.array([0, 1, 1, 1, 2]))
def test_search_space_transform_untransform_params() -> None:
search_space = {
"x0": DiscreteUniformDistribution(0, 1, q=0.2),
"x1": CategoricalDistribution(["foo", "bar", "baz", "qux"]),
"x2": IntLogUniformDistribution(1, 10),
"x3": CategoricalDistribution(["quux", "quuz"]),
"x4": UniformDistribution(2, 3),
"x5": LogUniformDistribution(1, 10),
"x6": IntUniformDistribution(2, 4),
"x7": CategoricalDistribution(["corge"]),
}
params = {
"x0": 0.2,
"x1": "qux",
"x2": 1,
"x3": "quux",
"x4": 2.0,
"x5": 1.0,
"x6": 2,
"x7": "corge",
}
trans = _SearchSpaceTransform(search_space)
trans_params = trans.transform(params)
untrans_params = trans.untransform(trans_params)
for name in params.keys():
assert untrans_params[name] == params[name]
def create_optuna_distribution_from_config(
config: MutableMapping[str, Any]) -> BaseDistribution:
kwargs = dict(config)
if isinstance(config["type"], str):
kwargs["type"] = DistributionType[config["type"]]
param = DistributionConfig(**kwargs)
if param.type == DistributionType.categorical:
assert param.choices is not None
return CategoricalDistribution(param.choices)
if param.type == DistributionType.int:
assert param.low is not None
assert param.high is not None
if param.log:
return IntLogUniformDistribution(int(param.low), int(param.high))
step = int(param.step) if param.step is not None else 1
return IntUniformDistribution(int(param.low),
int(param.high),
step=step)
if param.type == DistributionType.float:
assert param.low is not None
assert param.high is not None
if param.log:
return LogUniformDistribution(param.low, param.high)
if param.step is not None:
return DiscreteUniformDistribution(param.low, param.high,
param.step)
return UniformDistribution(param.low, param.high)
raise NotImplementedError(
f"{param.type} is not supported by Optuna sweeper.")
def create_optuna_distribution_from_override(override: Override) -> Any:
value = override.value()
if not override.is_sweep_override():
return value
choices: List[CategoricalChoiceType] = []
if override.is_choice_sweep():
assert isinstance(value, ChoiceSweep)
for x in override.sweep_iterator(transformer=Transformer.encode):
assert isinstance(
x, (str, int, float, bool)
), f"A choice sweep expects str, int, float, or bool type. Got {type(x)}."
choices.append(x)
return CategoricalDistribution(choices)
if override.is_range_sweep():
assert isinstance(value, RangeSweep)
assert value.start is not None
assert value.stop is not None
if value.shuffle:
for x in override.sweep_iterator(transformer=Transformer.encode):
assert isinstance(
x, (str, int, float, bool)
), f"A choice sweep expects str, int, float, or bool type. Got {type(x)}."
choices.append(x)
return CategoricalDistribution(choices)
return IntUniformDistribution(int(value.start),
int(value.stop),
step=int(value.step))
if override.is_interval_sweep():
assert isinstance(value, IntervalSweep)
assert value.start is not None
assert value.end is not None
if "log" in value.tags:
if isinstance(value.start, int) and isinstance(value.end, int):
return IntLogUniformDistribution(int(value.start),
int(value.end))
return LogUniformDistribution(value.start, value.end)
else:
if isinstance(value.start, int) and isinstance(value.end, int):
return IntUniformDistribution(value.start, value.end)
return UniformDistribution(value.start, value.end)
raise NotImplementedError(
f"{override} is not supported by Optuna sweeper.")
def search_space():
# type: () -> Dict[str, BaseDistribution]
return {
'c': CategoricalDistribution(('a', 'b')),
'd': DiscreteUniformDistribution(-1, 9, 2),
'i': IntUniformDistribution(-1, 1),
'l': LogUniformDistribution(0.001, 0.1),
'u': UniformDistribution(-2, 2),
}
def test_relative_parameters(storage_mode: str) -> None:
relative_search_space = {
"x": UniformDistribution(low=5, high=6),
"y": UniformDistribution(low=5, high=6),
}
relative_params = {"x": 5.5, "y": 5.5, "z": 5.5}
sampler = DeterministicRelativeSampler(relative_search_space,
relative_params) # type: ignore
with StorageSupplier(storage_mode) as storage:
study = create_study(storage=storage, sampler=sampler)
def create_trial() -> Trial:
return Trial(study,
study._storage.create_new_trial(study._study_id))
# Suggested from `relative_params`.
trial0 = create_trial()
distribution0 = UniformDistribution(low=0, high=100)
assert trial0._suggest("x", distribution0) == 5.5
# Not suggested from `relative_params` (due to unknown parameter name).
trial1 = create_trial()
distribution1 = distribution0
assert trial1._suggest("w", distribution1) != 5.5
# Not suggested from `relative_params` (due to incompatible value range).
trial2 = create_trial()
distribution2 = UniformDistribution(low=0, high=5)
assert trial2._suggest("x", distribution2) != 5.5
# Error (due to incompatible distribution class).
trial3 = create_trial()
distribution3 = IntUniformDistribution(low=1, high=100)
with pytest.raises(ValueError):
trial3._suggest("y", distribution3)
# Error ('z' is included in `relative_params` but not in `relative_search_space`).
trial4 = create_trial()
distribution4 = UniformDistribution(low=0, high=10)
with pytest.raises(ValueError):
trial4._suggest("z", distribution4)
# Error (due to incompatible distribution class).
trial5 = create_trial()
distribution5 = IntLogUniformDistribution(low=1, high=100)
with pytest.raises(ValueError):
trial5._suggest("y", distribution5)
def test_suggest_int(storage_init_func: Callable[[], storages.BaseStorage]) -> None:
mock = Mock()
mock.side_effect = [1, 2]
sampler = samplers.RandomSampler()
with patch.object(sampler, "sample_independent", mock) as mock_object:
study = create_study(storage_init_func(), sampler=sampler)
trial = Trial(study, study._storage.create_new_trial(study._study_id))
distribution = IntUniformDistribution(low=0, high=3)
assert trial._suggest("x", distribution) == 1 # Test suggesting a param.
assert trial._suggest("x", distribution) == 1 # Test suggesting the same param.
assert trial._suggest("y", distribution) == 2 # Test suggesting a different param.
assert trial.params == {"x": 1, "y": 2}
assert mock_object.call_count == 2
def restore_old_distribution(distribution_json: str) -> str:
distribution = json_to_distribution(distribution_json)
old_distribution: BaseDistribution
# Float distributions.
if isinstance(distribution, FloatDistribution):
if distribution.log:
old_distribution = LogUniformDistribution(
low=distribution.low,
high=distribution.high,
)
else:
if distribution.step is not None:
old_distribution = DiscreteUniformDistribution(
low=distribution.low,
high=distribution.high,
q=distribution.step,
)
else:
old_distribution = UniformDistribution(
low=distribution.low,
high=distribution.high,
)
# Integer distributions.
elif isinstance(distribution, IntDistribution):
if distribution.log:
old_distribution = IntLogUniformDistribution(
low=distribution.low,
high=distribution.high,
step=distribution.step,
)
else:
old_distribution = IntUniformDistribution(
low=distribution.low,
high=distribution.high,
step=distribution.step,
)
# Categorical distribution.
else:
old_distribution = distribution
return distribution_to_json(old_distribution)
def test_intersection_search_space():
# type: () -> None
study = optuna.create_study()
# No trial.
assert optuna.samplers.intersection_search_space(study) == {}
# First trial.
study.optimize(
lambda t: t.suggest_int("x", 0, 10) + t.suggest_uniform("y", -3, 3),
n_trials=1)
assert optuna.samplers.intersection_search_space(study) == {
"x": IntUniformDistribution(low=0, high=10),
"y": UniformDistribution(low=-3, high=3),
}
# Second trial (only 'y' parameter is suggested in this trial).
study.optimize(lambda t: t.suggest_uniform("y", -3, 3), n_trials=1)
assert optuna.samplers.intersection_search_space(study) == {
"y": UniformDistribution(low=-3, high=3)
}
# Failed or pruned trials are not considered in the calculation of
# an intersection search space.
def objective(trial, exception):
# type: (optuna.trial.Trial, Exception) -> float
trial.suggest_uniform("z", 0, 1)
raise exception
study.optimize(lambda t: objective(t, RuntimeError()),
n_trials=1,
catch=(RuntimeError, ))
study.optimize(lambda t: objective(t, optuna.exceptions.TrialPruned()),
n_trials=1)
assert optuna.samplers.intersection_search_space(study) == {
"y": UniformDistribution(low=-3, high=3)
}
# If two parameters have the same name but different distributions,
# those are regarded as different trials.
study.optimize(lambda t: t.suggest_uniform("y", -1, 1), n_trials=1)
assert optuna.samplers.intersection_search_space(study) == {}
def test_frozen_trial_suggest_int() -> None:
trial = FrozenTrial(
number=0,
trial_id=0,
state=TrialState.COMPLETE,
value=0.2,
datetime_start=datetime.datetime.now(),
datetime_complete=datetime.datetime.now(),
params={"x": 1},
distributions={"x": IntUniformDistribution(0, 10)},
user_attrs={},
system_attrs={},
intermediate_values={},
)
assert trial.suggest_int("x", 0, 10) == 1
with pytest.raises(ValueError):
trial.suggest_int("y", 0, 10)
def test_is_compatible(search_space, x0):
# type: (Dict[str, BaseDistribution], Dict[str, Any]) -> None
optimizer = optuna.integration.cma._Optimizer(search_space, x0, 0.1,
None, {})
# Compatible.
trial = _create_frozen_trial(x0, search_space)
assert optimizer._is_compatible(trial)
# Compatible.
trial = _create_frozen_trial(
x0, dict(search_space, u=UniformDistribution(-10, 10)))
assert optimizer._is_compatible(trial)
# Compatible.
trial = _create_frozen_trial(
dict(x0, unknown=7),
dict(search_space, unknown=UniformDistribution(0, 10)))
assert optimizer._is_compatible(trial)
# Incompatible ('u' doesn't exist).
param = dict(x0)
del param['u']
dist = dict(search_space)
del dist['u']
trial = _create_frozen_trial(param, dist)
assert not optimizer._is_compatible(trial)
# Incompatible (the value of 'u' is out of range).
trial = _create_frozen_trial(
dict(x0, u=20), dict(search_space,
u=UniformDistribution(-100, 100)))
assert not optimizer._is_compatible(trial)
# Error (different distribution class).
trial = _create_frozen_trial(
x0, dict(search_space, u=IntUniformDistribution(-2, 2)))
with pytest.raises(ValueError):
optimizer._is_compatible(trial)
def test_distributions(storage_mode: str) -> None:
def objective(trial: Trial) -> float:
trial.suggest_float("a", 0, 10)
trial.suggest_float("b", 0.1, 10, log=True)
trial.suggest_float("c", 0, 10, step=1)
trial.suggest_int("d", 0, 10)
trial.suggest_categorical("e", ["foo", "bar", "baz"])
trial.suggest_int("f", 1, 10, log=True)
return 1.0
with StorageSupplier(storage_mode) as storage:
study = create_study(storage=storage)
study.optimize(objective, n_trials=1)
assert study.best_trial.distributions == {
"a": UniformDistribution(low=0, high=10),
"b": LogUniformDistribution(low=0.1, high=10),
"c": DiscreteUniformDistribution(low=0, high=10, q=1),
"d": IntUniformDistribution(low=0, high=10),
"e": CategoricalDistribution(choices=("foo", "bar", "baz")),
"f": IntLogUniformDistribution(low=1, high=10),
}
def _parse_optuna_search_space(self, search_space_path):
with open(search_space_path) as fin:
search_space = json.load(fin)
optuna_search_space: Dict[str, BaseDistribution] = {}
low_values: Dict[str, Any] = {}
for name, value in search_space.items():
if not isinstance(value, dict):
d = CategoricalDistribution((value, ))
low_values[name] = str(value)
optuna_search_space[name] = d
continue
sampling_strategy = value["sampling strategy"]
if sampling_strategy == "choice":
d = CategoricalDistribution(tuple(value["choices"]))
optuna_search_space[name] = d
low_values[name] = str(value["choices"][0])
continue
if sampling_strategy == "integer":
d = IntUniformDistribution(value["bounds"][0],
value["bounds"][1])
optuna_search_space[name] = d
elif sampling_strategy == "uniform":
d = UniformDistribution(value["bounds"][0], value["bounds"][1])
optuna_search_space[name] = d
elif sampling_strategy == "loguniform":
d = LogUniformDistribution(value["bounds"][0],
value["bounds"][1])
optuna_search_space[name] = d
else:
raise ValueError(
f"Unknown sampling strategy: {sampling_strategy}.")
low_values[name] = str(value["bounds"][0])
return optuna_search_space, low_values
def test_group_decomposed_search_space() -> None:
search_space = _GroupDecomposedSearchSpace()
study = create_study()
# No trial.
assert search_space.calculate(study).search_spaces == []
# A single parameter.
study.optimize(lambda t: t.suggest_int("x", 0, 10), n_trials=1)
assert search_space.calculate(study).search_spaces == [{
"x":
IntUniformDistribution(low=0, high=10)
}]
# Disjoint parameters.
study.optimize(
lambda t: t.suggest_int("y", 0, 10) + t.suggest_float("z", -3, 3),
n_trials=1)
assert search_space.calculate(study).search_spaces == [
{
"x": IntUniformDistribution(low=0, high=10)
},
{
"y": IntUniformDistribution(low=0, high=10),
"z": UniformDistribution(low=-3, high=3),
},
]
# Parameters which include one of search spaces in the group.
study.optimize(
lambda t: t.suggest_int("y", 0, 10) + t.suggest_float("z", -3, 3) + t.
suggest_float("u", 1e-2, 1e2, log=True) + bool(
t.suggest_categorical("v", ["A", "B", "C"])),
n_trials=1,
)
assert search_space.calculate(study).search_spaces == [
{
"x": IntUniformDistribution(low=0, high=10)
},
{
"y": IntUniformDistribution(low=0, high=10),
"z": UniformDistribution(low=-3, high=3),
},
{
"u": LogUniformDistribution(low=1e-2, high=1e2),
"v": CategoricalDistribution(choices=["A", "B", "C"]),
},
]
# A parameter which is included by one of search spaces in thew group.
study.optimize(lambda t: t.suggest_float("u", 1e-2, 1e2, log=True),
n_trials=1)
assert search_space.calculate(study).search_spaces == [
{
"x": IntUniformDistribution(low=0, high=10)
},
{
"y": IntUniformDistribution(low=0, high=10),
"z": UniformDistribution(low=-3, high=3),
},
{
"u": LogUniformDistribution(low=1e-2, high=1e2)
},
{
"v": CategoricalDistribution(choices=["A", "B", "C"])
},
]
# Parameters whose intersection with one of search spaces in the group is not empty.
study.optimize(lambda t: t.suggest_int("y", 0, 10) + t.suggest_int(
"w", 2, 8, log=True),
n_trials=1)
assert search_space.calculate(study).search_spaces == [
{
"v": CategoricalDistribution(choices=["A", "B", "C"])
},
{
"x": IntUniformDistribution(low=0, high=10)
},
{
"u": LogUniformDistribution(low=1e-2, high=1e2)
},
{
"y": IntUniformDistribution(low=0, high=10)
},
{
"z": UniformDistribution(low=-3, high=3)
},
{
"w": IntLogUniformDistribution(low=2, high=8)
},
]
search_space = _GroupDecomposedSearchSpace()
study = create_study()
# Failed or pruned trials are not considered in the calculation of
# an intersection search space.
def objective(trial: Trial, exception: Exception) -> float:
trial.suggest_float("a", 0, 1)
raise exception
study.optimize(lambda t: objective(t, RuntimeError()),
n_trials=1,
catch=(RuntimeError, ))
study.optimize(lambda t: objective(t, TrialPruned()), n_trials=1)
assert search_space.calculate(study).search_spaces == []
# If two parameters have the same name but different distributions,
# the first one takes priority.
study.optimize(lambda t: t.suggest_float("a", -1, 1), n_trials=1)
study.optimize(lambda t: t.suggest_float("a", 0, 1), n_trials=1)
assert search_space.calculate(study).search_spaces == [{
"a":
UniformDistribution(low=-1, high=1)
}]
import pytest
from optuna._transform import _SearchSpaceTransform
from optuna.distributions import BaseDistribution
from optuna.distributions import CategoricalDistribution
from optuna.distributions import DiscreteUniformDistribution
from optuna.distributions import IntLogUniformDistribution
from optuna.distributions import IntUniformDistribution
from optuna.distributions import LogUniformDistribution
from optuna.distributions import UniformDistribution
@pytest.mark.parametrize(
"param,distribution",
[
(0, IntUniformDistribution(0, 3)),
(1, IntLogUniformDistribution(1, 10)),
(2, IntUniformDistribution(0, 10, step=2)),
(0.0, UniformDistribution(0, 3)),
(1.0, LogUniformDistribution(1, 10)),
(0.2, DiscreteUniformDistribution(0, 1, q=0.2)),
("foo", CategoricalDistribution(["foo"])),
("bar", CategoricalDistribution(["foo", "bar", "baz"])),
],
)
def test_search_space_transform_shapes_dtypes(param: Any, distribution: BaseDistribution) -> None:
trans = _SearchSpaceTransform({"x0": distribution})
trans_params = trans.transform({"x0": param})
if isinstance(distribution, CategoricalDistribution):
expected_bounds_shape = (len(distribution.choices), 2)
def suggest_int(self, name: str, low: int, high: int, step: int = 1, log: bool = False) -> int:
"""Suggest a value for the integer parameter.
The value is sampled from the integers in :math:`[\\mathsf{low}, \\mathsf{high}]`.
Example:
Suggest the number of trees in `RandomForestClassifier <https://scikit-learn.org/
stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html>`_.
.. testcode::
import numpy as np
from sklearn.datasets import load_iris
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
import optuna
X, y = load_iris(return_X_y=True)
X_train, X_valid, y_train, y_valid = train_test_split(X, y)
def objective(trial):
n_estimators = trial.suggest_int("n_estimators", 50, 400)
clf = RandomForestClassifier(n_estimators=n_estimators, random_state=0)
clf.fit(X_train, y_train)
return clf.score(X_valid, y_valid)
study = optuna.create_study(direction="maximize")
study.optimize(objective, n_trials=3)
Args:
name:
A parameter name.
low:
Lower endpoint of the range of suggested values. ``low`` is included in the range.
high:
Upper endpoint of the range of suggested values. ``high`` is included in the range.
step:
A step of discretization.
.. note::
Note that :math:`\\mathsf{high}` is modified if the range is not divisible by
:math:`\\mathsf{step}`. Please check the warning messages to find the changed
values.
.. note::
The method returns one of the values in the sequence
:math:`\\mathsf{low}, \\mathsf{low} + \\mathsf{step}, \\mathsf{low} + 2 *
\\mathsf{step}, \\dots, \\mathsf{low} + k * \\mathsf{step} \\le
\\mathsf{high}`, where :math:`k` denotes an integer.
.. note::
The ``step != 1`` and ``log`` arguments cannot be used at the same time.
To set the ``step`` argument :math:`\\mathsf{step} \\ge 2`, set the
``log`` argument to :obj:`False`.
log:
A flag to sample the value from the log domain or not.
.. note::
If ``log`` is true, at first, the range of suggested values is divided into
grid points of width 1. The range of suggested values is then converted to
a log domain, from which a value is sampled. The uniformly sampled
value is re-converted to the original domain and rounded to the nearest grid
point that we just split, and the suggested value is determined.
For example, if `low = 2` and `high = 8`, then the range of suggested values is
`[2, 3, 4, 5, 6, 7, 8]` and lower values tend to be more sampled than higher
values.
.. note::
The ``step != 1`` and ``log`` arguments cannot be used at the same time.
To set the ``log`` argument to :obj:`True`, set the ``step`` argument to 1.
Raises:
:exc:`ValueError`:
If ``step != 1`` and ``log = True`` are specified.
.. seealso::
:ref:`configurations` tutorial describes more details and flexible usages.
"""
if step != 1:
if log:
raise ValueError(
"The parameter `step != 1` is not supported when `log` is True."
"The specified `step` is {}.".format(step)
)
else:
distribution: Union[
IntUniformDistribution, IntLogUniformDistribution
] = IntUniformDistribution(low=low, high=high, step=step)
else:
if log:
distribution = IntLogUniformDistribution(low=low, high=high)
else:
distribution = IntUniformDistribution(low=low, high=high, step=step)
self._check_distribution(name, distribution)
return int(self._suggest(name, distribution))
np.floating,
)
# Check all points are multiples of distribution.q.
points = points
points -= distribution.low
points /= distribution.q
round_points = np.round(points)
np.testing.assert_almost_equal(round_points, points)
@parametrize_sampler
@pytest.mark.parametrize(
"distribution",
[
IntUniformDistribution(-10, 10),
IntUniformDistribution(0, 10),
IntUniformDistribution(-10, 0),
IntUniformDistribution(-10, 10, 2),
IntUniformDistribution(0, 10, 2),
IntUniformDistribution(-10, 0, 2),
],
)
def test_int(sampler_class: Callable[[], BaseSampler],
distribution: IntUniformDistribution) -> None:
study = optuna.study.create_study(sampler=sampler_class())
points = np.array([
study.sampler.sample_independent(study, _create_new_trial(study), "x",
distribution) for _ in range(100)
])
def suggest_int(self, name, low, high, step=1, log=False):
# type: (str, int, int, int, bool) -> int
"""Suggest a value for the integer parameter.
The value is sampled from the integers in :math:`[\\mathsf{low}, \\mathsf{high}]`, and the
step of discretization is :math:`\\mathsf{step}`. More specifically, this method returns
one of the values in the sequence :math:`\\mathsf{low}, \\mathsf{low} + \\mathsf{step},
\\mathsf{low} + 2 * \\mathsf{step}, \\dots, \\mathsf{low} + k * \\mathsf{step} \\le
\\mathsf{high}`, where :math:`k` denotes an integer. Note that :math:`\\mathsf{high}` is
modified if the range is not divisible by :math:`\\mathsf{step}`. Please check the warning
messages to find the changed values.
Example:
Suggest the number of trees in `RandomForestClassifier <https://scikit-learn.org/
stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html>`_.
.. testcode::
import numpy as np
from sklearn.datasets import load_iris
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
import optuna
X, y = load_iris(return_X_y=True)
X_train, X_valid, y_train, y_valid = train_test_split(X, y)
def objective(trial):
n_estimators = trial.suggest_int('n_estimators', 50, 400)
clf = RandomForestClassifier(n_estimators=n_estimators, random_state=0)
clf.fit(X_train, y_train)
return clf.score(X_valid, y_valid)
study = optuna.create_study(direction='maximize')
study.optimize(objective, n_trials=3)
Args:
name:
A parameter name.
low:
Lower endpoint of the range of suggested values. ``low`` is included in the range.
high:
Upper endpoint of the range of suggested values. ``high`` is included in the range.
step:
A step of discretization.
log:
A flag to sample the value from the log domain or not.
If ``log`` is true, at first, the range of suggested values is divided into grid
points of width ``step``. The range of suggested values is then converted to a log
domain, from which a value is uniformly sampled. The uniformly sampled value is
re-converted to the original domain and rounded to the nearest grid point that we
just split, and the suggested value is determined.
For example,
if `low = 2`, `high = 8` and `step = 2`,
then the range of suggested values is divided by ``step`` as `[2, 4, 6, 8]`
and lower values tend to be more sampled than higher values.
"""
if log:
distribution = IntLogUniformDistribution(
low=low, high=high, step=step
) # type: Union[IntUniformDistribution, IntLogUniformDistribution]
else:
distribution = IntUniformDistribution(low=low, high=high, step=step)
self._check_distribution(name, distribution)
if distribution.low == distribution.high:
return self._set_new_param_or_get_existing(name, distribution.low, distribution)
return int(self._suggest(name, distribution))
@mark.parametrize(
"input, expected",
[
(
{
"type": "categorical",
"choices": [1, 2, 3]
},
CategoricalDistribution([1, 2, 3]),
),
({
"type": "int",
"low": 0,
"high": 10
}, IntUniformDistribution(0, 10)),
(
{
"type": "int",
"low": 0,
"high": 10,
"step": 2
},
IntUniformDistribution(0, 10, step=2),
),
({
"type": "int",
"low": 0,
"high": 5
}, IntUniformDistribution(0, 5)),
(