def evaluate(self, study: Study,
params: Optional[List[str]]) -> Dict[str, float]:
distributions = _get_distributions(study, params)
params_data, values_data = _get_study_data(study, distributions)
evaluator = fANOVA(
X=params_data,
Y=values_data,
config_space=_get_configuration_space(distributions),
max_features=max(1, int(params_data.shape[1] * 0.7)),
)
individual_importances = {}
for i, name in enumerate(evaluator.cs.get_hyperparameter_names()):
imp = evaluator.quantify_importance((i, ))
imp = imp[(i, )]["individual importance"]
individual_importances[name] = imp
tot_importance = sum(v for v in individual_importances.values())
for name in individual_importances.keys():
individual_importances[name] /= tot_importance
param_importances = OrderedDict(
reversed(
sorted(
individual_importances.items(),
key=lambda name_and_importance: name_and_importance[1],
)))
return param_importances
def evaluate(self,
study: Study,
params: Optional[List[str]] = None) -> Dict[str, float]:
distributions = _get_distributions(study, params)
params_data, values_data = _get_study_data(study, distributions)
if params_data.size == 0: # `params` were given but as an empty list.
return OrderedDict()
params_data, cols_to_raw_cols = _encode_categorical(
params_data, distributions.values())
forest = self._forest
forest.fit(params_data, values_data)
feature_importances = forest.feature_importances_
feature_importances_reduced = numpy.zeros(len(distributions))
numpy.add.at(feature_importances_reduced, cols_to_raw_cols,
feature_importances)
param_importances = OrderedDict()
param_names = list(distributions.keys())
for i in feature_importances_reduced.argsort()[::-1]:
param_importances[
param_names[i]] = feature_importances_reduced[i].item()
return param_importances
def evaluate(self, study: Study, params: Optional[List[str]] = None) -> Dict[str, float]:
distributions = _get_distributions(study, params)
params_data, values_data = _get_study_data(study, distributions)
if params_data.size == 0: # `params` were given but as an empty list.
return OrderedDict()
# Many (deep) copies of the search spaces are required during the tree traversal and using
# Optuna distributions will create a bottleneck.
# Therefore, search spaces (parameter distributions) are represented by a single
# `numpy.ndarray`, coupled with a list of flags that indicate whether they are categorical
# or not.
search_spaces = numpy.empty((len(distributions), 2), dtype=numpy.float64)
search_spaces_is_categorical = []
for i, distribution in enumerate(distributions.values()):
if isinstance(distribution, CategoricalDistribution):
search_spaces[i, 0] = 0
search_spaces[i, 1] = len(distribution.choices)
search_spaces_is_categorical.append(True)
elif isinstance(
distribution,
(
DiscreteUniformDistribution,
IntLogUniformDistribution,
IntUniformDistribution,
LogUniformDistribution,
UniformDistribution,
),
):
search_spaces[i, 0] = distribution.low
search_spaces[i, 1] = distribution.high
search_spaces_is_categorical.append(False)
else:
assert False
evaluator = self._evaluator
evaluator.fit(
X=params_data,
y=values_data,
search_spaces=search_spaces,
search_spaces_is_categorical=search_spaces_is_categorical,
)
importances = {}
for i, name in enumerate(distributions.keys()):
importance, _ = evaluator.get_importance((i,))
importances[name] = importance
total_importance = sum(importances.values())
for name in importances.keys():
importances[name] /= total_importance
sorted_importances = OrderedDict(
reversed(
sorted(importances.items(), key=lambda name_and_importance: name_and_importance[1])
)
)
return sorted_importances
