def _prepare_classifiers(self):
if self.classifier_params is None:
self.classifier_params = dict()
self.classifier_name.set_params(**self.classifier_params)
if self.select_features is not None:
if self.selection_params is None:
self.selection_params = {
'nfeatures': 0.1,
'min_count': 2,
'minfeatures': 100
}
self.selection_params['method'] = self.select_features
self.first_selector = MulticlassFeatureSelector(
local=True,
method=self.select_features,
min_count=self.min_count,
nfeatures=-1)
feature_selector = MulticlassFeatureSelector(
**self.selection_params)
self.classifier_ = Pipeline([('selector', feature_selector),
('classifier', self.classifier_name)])
else:
self.first_selector = ZeroFeatureRemover()
self.classifier_ = self.classifier_name
return self
def _prepare_classifiers(self):
if self.classifier_params is None:
self.classifier_params = dict()
self.classifier_name.set_params(**self.classifier_params)
if self.select_features is not None:
if self.selection_params is None:
self.selection_params = {'nfeatures': 0.1, 'min_count': 2, 'minfeatures': 100}
self.selection_params['method'] = self.select_features
self.first_selector = MulticlassFeatureSelector(
local=True, method=self.select_features, min_count=self.min_count, nfeatures=-1)
feature_selector = MulticlassFeatureSelector(**self.selection_params)
self.classifier_ = Pipeline([('selector', feature_selector),
('classifier', self.classifier_name)])
else:
self.first_selector = ZeroFeatureRemover()
self.classifier_ = self.classifier_name
return self
class LocalTransformationClassifier(BaseEstimator, ClassifierMixin):
def __init__(self, left_context_length=2, right_context_length=3, has_central_context=True,
classifier_name=sklm.LogisticRegression(), classifier_params=None,
multiclass=False, select_features=False, selection_params = None,
min_count=3, smallest_prob=0.01, min_probs_ratio=0.75):
self.left_context_length = left_context_length
self.right_context_length = right_context_length
self.has_central_context = has_central_context
self.select_features = True
self.substrs = []
self.classifier_name = classifier_name
self.classifier_params = classifier_params
self.multiclass = multiclass
self.select_features = select_features
self.selection_params = selection_params
self.smallest_prob = smallest_prob
self.min_probs_ratio = min_probs_ratio
self.min_count = min_count
# инициализация параметров
self.classifiers = dict()
self.classifier_classes = dict()
self.predefined_answers = dict()
self.selector_name = None
self.selectors = dict()
self.feature_codes = dict()
self.features = []
self.features_number = 0
def _prepare_classifiers(self):
if self.classifier_params is None:
self.classifier_params = dict()
self.classifier_name.set_params(**self.classifier_params)
if self.select_features is not None:
if self.selection_params is None:
self.selection_params = {'nfeatures': 0.1, 'min_count': 2, 'minfeatures': 100}
self.selection_params['method'] = self.select_features
self.first_selector = MulticlassFeatureSelector(
local=True, method=self.select_features, min_count=self.min_count, nfeatures=-1)
feature_selector = MulticlassFeatureSelector(**self.selection_params)
self.classifier_ = Pipeline([('selector', feature_selector),
('classifier', self.classifier_name)])
else:
self.first_selector = ZeroFeatureRemover()
self.classifier_ = self.classifier_name
return self
def fit(self, X, y):
"""
Применяет классификатор к данным.
X --- набор троек вида (фрагмент, парадигма, левый контекст, правый контекст)
"""
self._prepare_classifiers()
substrs, X_train, y = self._preprocess_input(X, y, create_features=True,
retain_multiple=self.multiclass, return_y=True)
X_train = self.first_selector.fit_transform(X_train, y)
self._fit_classifiers(substrs, X_train, y)
return self
def _fit_classifiers(self, substrs, X_train, y):
"""
Обучает классификаторы для каждой из подстрок, к которой планируется его применять
"""
if issparse(X_train):
X_train = X_train.tocsr()
substr_indexes_in_train = _find_positions(substrs)
for substr, indexes in substr_indexes_in_train.items():
X_curr, y_curr = X_train[indexes,:], [y[i] for i in indexes]
X_curr.sort_indices()
max_index = max(y_curr)
# в случае, если y_curr содержит только один класс, не обучаем классификатор
if min(y_curr) == max_index:
self.predefined_answers[substr] = max_index
continue
self.classifiers[substr] = clone(self.classifier_)
self.classifiers[substr].fit(X_curr, y_curr)
self.classifier_classes[substr] = self.classifiers[substr].classes_
return
def predict(self, X):
probs = self._predict_proba(X)
if not self.multiclass:
class_indexes = [indices[0] for indices, _ in probs]
answer = [[x] for x in np.take(self.classes_, class_indexes)]
else:
answer = [np.take(self.classes_, indices) for indices
in extract_classes_from_sparse_probs(probs, self.min_probs_ratio)]
return answer
def predict_proba(self, X):
# probs = self._predict_proba(X)
# answer = np.zeros(dtype=np.float64, shape=(len(X), len(self.classes_)))
# for i, (indices, elem_probs) in enumerate(probs):
# answer[i][indices] = elem_probs
# return answer
return self._predict_proba(X, return_arrays=True)
def _predict_proba(self, X, return_arrays=False):
"""
Аргументы:
------------
X: list of tuples, список троек вида (преобразуемая подстрока, левый контекст, правый контекст)
Возвращает:
------------
answer: list of tuples, список пар вида (indices, probs),
indices: array, shape=(len(self.classes), dtype=int,
список кодов трансформаций в порядке убывания их вероятности,
probs: array, shape=(len(self.classes), dtype=float,
вероятности трансформаций в порядке убывания.
"""
substrs, X_test = self._preprocess_input(X, create_features=False, retain_multiple=False)
X_test = self.first_selector.transform(X_test)
substr_indexes = _find_positions(substrs)
if return_arrays:
answer = np.zeros(shape=(len(substrs), len(self.classes_)))
else:
answer = [None for _ in substrs]
for substr, indexes in substr_indexes.items():
# print(substr, end=" ")
cls = self.classifiers.get(substr, None)
if cls is not None:
# X_curr = self.selectors[substr].transform(X_test[indexes])
curr_smallest_prob = self.smallest_prob / len(cls.classes_)
X_curr = X_test[indexes,:]
probs = cls.predict_proba(X_curr)
# probs[np.where(probs < curr_smallest_prob)] = 0.0
probs[np.where(probs < self.smallest_prob)] = 0.0
if return_arrays:
answer[np.ix_(indexes, cls.classes_)] = probs
else:
for index, row in zip(indexes, probs):
nonzero_indices = row.nonzero()[0]
row = row[nonzero_indices]
indices_order = np.flipud(np.argsort(row))
current_indices = nonzero_indices[indices_order]
answer[index] = (np.take(cls.classes_, current_indices),
row[indices_order] / row.sum())
# for index, row in zip(indexes, probs):
# nonzero_indices = row.nonzero()[0]
# row = row[nonzero_indices]
# indices_order = np.flipud(np.argsort(row))
# current_indices = nonzero_indices[indices_order]
# elem_probs = row[indices_order]
# probs_sum, last_index = elem_probs[0], 1
# while probs_sum < 1.0 - self.smallest_prob:
# probs_sum += elem_probs[last_index]
# last_index += 1
# class_indexes = np.take(cls.classes_, current_indices[:(last_index)])
# answer[index] = (class_indexes, elem_probs[:(last_index)] / probs_sum)
else:
# такой подстроки не было в обучающей выборке,
# поэтому для неё возращаем отсутствие замены
code = self.predefined_answers.get(substr, 0)
if return_arrays:
answer[indexes, code] = 1.0
else:
for index in indexes:
answer[index] = ([code], 1.0)
return answer
def _preprocess_input(self, X, y=None, create_features=False,
retain_multiple=False, return_y=False, sparse_type='csr'):
"""
Аргументы:
----------
y: list of lists or None, список, i-ый элемент которого содержит классы
для X[i] из обучающей выборки
create_features: bool, optional(default=False), создаются ли новые признаки
или им сопоставляется значение ``неизвестный суффикс/префикс''
retain_multiple: bool, optional(default=False), может ли возвращаться
несколько меток классов для одного объекта
return_y: возвращается ли y (True в случае обработки обучающей выборки,
т.к. в этом случае y преобразуется из списка списков в один список)
"""
## СОКРАТИТЬ
if create_features:
# создаём новые признаки, вызывается при обработке обучающей выборки
self._create_basic_features()
self.features_number = len(self.features)
# создаём функцию для обработки неизвестных признаков
## ПЕРЕПИСАТЬ КАК В ОБЫЧНОМ КЛАССИФИКАТОРЕ
# _process_unknown_feature = self._create_unknown_feature_processor(create_features)
if y is None:
y = [[None]] * len(X)
retain_multiple = False
y_new = list(chain(*y))
else:
self.classes_, y_new = np.unique(list(chain(*y)), return_inverse=True)
X_with_temporary_codes = []
for _, left_context, right_context in X:
active_features_codes = []
# обрабатываем левые контексты
for length in range(1, self.left_context_length + 1):
feature = left_context[-length: ]
feature_code, to_break =\
self._get_feature_code(feature, (len(feature) != length),
'left', create_new=create_features)
active_features_codes.append(feature_code)
if to_break:
break
# правые контексты
for length in range(1, self.right_context_length + 1):
feature = right_context[:length]
feature_code, to_break =\
self._get_feature_code(feature, (len(feature) != length),
'right', create_new=create_features)
active_features_codes.append(feature_code)
if to_break:
break
# центральный контекст
if self.has_central_context:
feature = left_context[-1:] + '#' + right_context[:1]
if feature.startswith('#'):
feature = '^' + feature
if feature.endswith('#'):
feature += '$'
feature_code, to_break =\
self._get_feature_code(feature, False, 'center', create_new=create_features)
if feature_code > 0:
active_features_codes.append(feature_code)
X_with_temporary_codes.append(active_features_codes)
if create_features:
# при равной длине порядок x$, x#, $x, #x
feature_order = sorted(enumerate(self.features),
key=(lambda x:(len(x[1]), x[1].startswith('#'),
x[1].startswith('$'), x[1].endswith('#'))))
self.features = [x[1] for x in feature_order]
# перекодируем временные признаки
temporary_features_recoding = [None] * self.features_number
for new_code, (code, _) in enumerate(feature_order):
temporary_features_recoding[code] = new_code
for i, elem in enumerate(X_with_temporary_codes):
X_with_temporary_codes[i] = [temporary_features_recoding[code] for code in elem]
self.feature_codes = {feature: code for code, feature in enumerate(self.features)}
# сохраняем данные для классификации
rows, cols, curr_row = [], [], 0
for codes, labels in zip(X_with_temporary_codes, y):
# каждый объект обучающей выборки размножается k раз,
# где k --- число классов, к которым он принадлежит, например, пара
# X[i]=x, y[i]=[1, 2, 6] будет преобразована в <x, 1>, <x, 2>, <x, 6>
number_of_rows_to_add = 1 if not(retain_multiple) else len(labels)
for i in range(number_of_rows_to_add):
rows.extend([curr_row for _ in codes])
cols.extend(codes)
curr_row += 1
# сохраняем преобразованные данные в разреженную матрицу
data = np.ones(shape=(len(rows,)), dtype=float)
if sparse_type == 'csr':
sparse_matrix = csr_matrix
elif sparse_type == 'csc':
sparse_matrix = csc_matrix
X_train = sparse_matrix((data, (rows, cols)), shape=(curr_row, self.features_number))
# сохраняем подстроки, которые и нужно классифицировать
substrs = [x[0] for x in X]
if return_y:
return substrs, X_train, y_new
else:
return substrs, X_train
def _create_basic_features(self):
"""
Создаёт базовые признаки ещё до чтения входа
Возвращает:
-----------
features: list, список созданных признаков
"""
self.features, self.feature_codes = [], dict()
# признаки для контекстов, не встречавшихся в обучающей выборке
# правые контексты
self.features.extend(("-" * length + '#')
for length in range(1, self.right_context_length + 1))
# короткие правые контексты
self.features.extend(("-" * length + '$')
for length in range(self.right_context_length))
# левые контексты
self.features.extend(('#' + "-" * length)
for length in range(1, self.right_context_length + 1))
# короткие левые контексты
self.features.extend(('^' + "-" * length)
for length in range(self.right_context_length))
if self.has_central_context:
self.features.append('-#-')
for code, feature in enumerate(self.features):
self.feature_codes[feature] = code
return
# def _create_new_feature(self, feature, is_full, side):
# if is_full:
# delim = '^' if side == 'left' else '$'
# to_break = True
# else:
# delim, to_break = '#', False
# if side == 'left':
# feature = delim + feature
# elif side == 'right':
# feature += delim
# feature_code = self.feature_codes.get(feature, None)
# if feature_code is None:
# feature_code = self._process_unknown_feature(feature, side)
# return feature_code, to_break
def _get_feature_code(self, feature, is_full, side, create_new=False):
"""
Возвращает код признака feature
"""
if is_full:
delim = '^' if side == 'left' else '$'
to_break = True
else:
delim, to_break = '#', False
if side == 'left':
feature = delim + feature
elif side == 'right':
feature += delim
code = self.feature_codes.get(feature, -1)
if code < 0:
if create_new:
self.features.append(feature)
self.feature_codes[feature] = code = self.features_number
self.features_number += 1
else:
if side == 'right':
partial_features = ['-' * start + feature[start:]
for start in range(1, len(feature))]
elif side == 'left':
partial_features = [feature[:(len(feature)-start)] + '-' * start
for start in range(1, len(feature))][::-1]
else: # side == 'center'
partial_features = []
code, to_break = -1, None
for partial_feature in partial_features:
code = self.feature_codes.get(partial_feature, -1)
if code > 0:
break
if code > 0:
self.feature_codes[feature] = code
return code, to_break
class LocalTransformationClassifier(BaseEstimator, ClassifierMixin):
def __init__(self,
left_context_length=2,
right_context_length=3,
has_central_context=True,
classifier_name=sklm.LogisticRegression(),
classifier_params=None,
multiclass=False,
select_features=False,
selection_params=None,
min_count=3,
smallest_prob=0.01,
min_probs_ratio=0.75):
self.left_context_length = left_context_length
self.right_context_length = right_context_length
self.has_central_context = has_central_context
self.select_features = True
self.substrs = []
self.classifier_name = classifier_name
self.classifier_params = classifier_params
self.multiclass = multiclass
self.select_features = select_features
self.selection_params = selection_params
self.smallest_prob = smallest_prob
self.min_probs_ratio = min_probs_ratio
self.min_count = min_count
# инициализация параметров
self.classifiers = dict()
self.classifier_classes = dict()
self.predefined_answers = dict()
self.selector_name = None
self.selectors = dict()
self.feature_codes = dict()
self.features = []
self.features_number = 0
def _prepare_classifiers(self):
if self.classifier_params is None:
self.classifier_params = dict()
self.classifier_name.set_params(**self.classifier_params)
if self.select_features is not None:
if self.selection_params is None:
self.selection_params = {
'nfeatures': 0.1,
'min_count': 2,
'minfeatures': 100
}
self.selection_params['method'] = self.select_features
self.first_selector = MulticlassFeatureSelector(
local=True,
method=self.select_features,
min_count=self.min_count,
nfeatures=-1)
feature_selector = MulticlassFeatureSelector(
**self.selection_params)
self.classifier_ = Pipeline([('selector', feature_selector),
('classifier', self.classifier_name)])
else:
self.first_selector = ZeroFeatureRemover()
self.classifier_ = self.classifier_name
return self
def fit(self, X, y):
"""
Применяет классификатор к данным.
X --- набор троек вида (фрагмент, парадигма, левый контекст, правый контекст)
"""
self._prepare_classifiers()
substrs, X_train, y = self._preprocess_input(
X,
y,
create_features=True,
retain_multiple=self.multiclass,
return_y=True)
X_train = self.first_selector.fit_transform(X_train, y)
self._fit_classifiers(substrs, X_train, y)
return self
def _fit_classifiers(self, substrs, X_train, y):
"""
Обучает классификаторы для каждой из подстрок, к которой планируется его применять
"""
if issparse(X_train):
X_train = X_train.tocsr()
substr_indexes_in_train = _find_positions(substrs)
for substr, indexes in substr_indexes_in_train.items():
X_curr, y_curr = X_train[indexes, :], [y[i] for i in indexes]
X_curr.sort_indices()
max_index = max(y_curr)
# в случае, если y_curr содержит только один класс, не обучаем классификатор
if min(y_curr) == max_index:
self.predefined_answers[substr] = max_index
continue
self.classifiers[substr] = clone(self.classifier_)
self.classifiers[substr].fit(X_curr, y_curr)
self.classifier_classes[substr] = self.classifiers[substr].classes_
return
def predict(self, X):
probs = self._predict_proba(X)
if not self.multiclass:
class_indexes = [indices[0] for indices, _ in probs]
answer = [[x] for x in np.take(self.classes_, class_indexes)]
else:
answer = [
np.take(self.classes_, indices)
for indices in extract_classes_from_sparse_probs(
probs, self.min_probs_ratio)
]
return answer
def predict_proba(self, X):
# probs = self._predict_proba(X)
# answer = np.zeros(dtype=np.float64, shape=(len(X), len(self.classes_)))
# for i, (indices, elem_probs) in enumerate(probs):
# answer[i][indices] = elem_probs
# return answer
return self._predict_proba(X, return_arrays=True)
def _predict_proba(self, X, return_arrays=False):
"""
Аргументы:
------------
X: list of tuples, список троек вида (преобразуемая подстрока, левый контекст, правый контекст)
Возвращает:
------------
answer: list of tuples, список пар вида (indices, probs),
indices: array, shape=(len(self.classes), dtype=int,
список кодов трансформаций в порядке убывания их вероятности,
probs: array, shape=(len(self.classes), dtype=float,
вероятности трансформаций в порядке убывания.
"""
substrs, X_test = self._preprocess_input(X,
create_features=False,
retain_multiple=False)
X_test = self.first_selector.transform(X_test)
substr_indexes = _find_positions(substrs)
if return_arrays:
answer = np.zeros(shape=(len(substrs), len(self.classes_)))
else:
answer = [None for _ in substrs]
for substr, indexes in substr_indexes.items():
# print(substr, end=" ")
cls = self.classifiers.get(substr, None)
if cls is not None:
# X_curr = self.selectors[substr].transform(X_test[indexes])
curr_smallest_prob = self.smallest_prob / len(cls.classes_)
X_curr = X_test[indexes, :]
probs = cls.predict_proba(X_curr)
# probs[np.where(probs < curr_smallest_prob)] = 0.0
probs[np.where(probs < self.smallest_prob)] = 0.0
if return_arrays:
answer[np.ix_(indexes, cls.classes_)] = probs
else:
for index, row in zip(indexes, probs):
nonzero_indices = row.nonzero()[0]
row = row[nonzero_indices]
indices_order = np.flipud(np.argsort(row))
current_indices = nonzero_indices[indices_order]
answer[index] = (np.take(cls.classes_,
current_indices),
row[indices_order] / row.sum())
# for index, row in zip(indexes, probs):
# nonzero_indices = row.nonzero()[0]
# row = row[nonzero_indices]
# indices_order = np.flipud(np.argsort(row))
# current_indices = nonzero_indices[indices_order]
# elem_probs = row[indices_order]
# probs_sum, last_index = elem_probs[0], 1
# while probs_sum < 1.0 - self.smallest_prob:
# probs_sum += elem_probs[last_index]
# last_index += 1
# class_indexes = np.take(cls.classes_, current_indices[:(last_index)])
# answer[index] = (class_indexes, elem_probs[:(last_index)] / probs_sum)
else:
# такой подстроки не было в обучающей выборке,
# поэтому для неё возращаем отсутствие замены
code = self.predefined_answers.get(substr, 0)
if return_arrays:
answer[indexes, code] = 1.0
else:
for index in indexes:
answer[index] = ([code], 1.0)
return answer
def _preprocess_input(self,
X,
y=None,
create_features=False,
retain_multiple=False,
return_y=False,
sparse_type='csr'):
"""
Аргументы:
----------
y: list of lists or None, список, i-ый элемент которого содержит классы
для X[i] из обучающей выборки
create_features: bool, optional(default=False), создаются ли новые признаки
или им сопоставляется значение ``неизвестный суффикс/префикс''
retain_multiple: bool, optional(default=False), может ли возвращаться
несколько меток классов для одного объекта
return_y: возвращается ли y (True в случае обработки обучающей выборки,
т.к. в этом случае y преобразуется из списка списков в один список)
"""
## СОКРАТИТЬ
if create_features:
# создаём новые признаки, вызывается при обработке обучающей выборки
self._create_basic_features()
self.features_number = len(self.features)
# создаём функцию для обработки неизвестных признаков
## ПЕРЕПИСАТЬ КАК В ОБЫЧНОМ КЛАССИФИКАТОРЕ
# _process_unknown_feature = self._create_unknown_feature_processor(create_features)
if y is None:
y = [[None]] * len(X)
retain_multiple = False
y_new = list(chain(*y))
else:
self.classes_, y_new = np.unique(list(chain(*y)),
return_inverse=True)
X_with_temporary_codes = []
for _, left_context, right_context in X:
active_features_codes = []
# обрабатываем левые контексты
for length in range(1, self.left_context_length + 1):
feature = left_context[-length:]
feature_code, to_break =\
self._get_feature_code(feature, (len(feature) != length),
'left', create_new=create_features)
active_features_codes.append(feature_code)
if to_break:
break
# правые контексты
for length in range(1, self.right_context_length + 1):
feature = right_context[:length]
feature_code, to_break =\
self._get_feature_code(feature, (len(feature) != length),
'right', create_new=create_features)
active_features_codes.append(feature_code)
if to_break:
break
# центральный контекст
if self.has_central_context:
feature = left_context[-1:] + '#' + right_context[:1]
if feature.startswith('#'):
feature = '^' + feature
if feature.endswith('#'):
feature += '$'
feature_code, to_break =\
self._get_feature_code(feature, False, 'center', create_new=create_features)
if feature_code > 0:
active_features_codes.append(feature_code)
X_with_temporary_codes.append(active_features_codes)
if create_features:
# при равной длине порядок x$, x#, $x, #x
feature_order = sorted(enumerate(self.features),
key=(lambda x:
(len(x[1]), x[1].startswith('#'), x[1].
startswith('$'), x[1].endswith('#'))))
self.features = [x[1] for x in feature_order]
# перекодируем временные признаки
temporary_features_recoding = [None] * self.features_number
for new_code, (code, _) in enumerate(feature_order):
temporary_features_recoding[code] = new_code
for i, elem in enumerate(X_with_temporary_codes):
X_with_temporary_codes[i] = [
temporary_features_recoding[code] for code in elem
]
self.feature_codes = {
feature: code
for code, feature in enumerate(self.features)
}
# сохраняем данные для классификации
rows, cols, curr_row = [], [], 0
for codes, labels in zip(X_with_temporary_codes, y):
# каждый объект обучающей выборки размножается k раз,
# где k --- число классов, к которым он принадлежит, например, пара
# X[i]=x, y[i]=[1, 2, 6] будет преобразована в <x, 1>, <x, 2>, <x, 6>
number_of_rows_to_add = 1 if not (retain_multiple) else len(labels)
for i in range(number_of_rows_to_add):
rows.extend([curr_row for _ in codes])
cols.extend(codes)
curr_row += 1
# сохраняем преобразованные данные в разреженную матрицу
data = np.ones(shape=(len(rows, )), dtype=float)
if sparse_type == 'csr':
sparse_matrix = csr_matrix
elif sparse_type == 'csc':
sparse_matrix = csc_matrix
X_train = sparse_matrix((data, (rows, cols)),
shape=(curr_row, self.features_number))
# сохраняем подстроки, которые и нужно классифицировать
substrs = [x[0] for x in X]
if return_y:
return substrs, X_train, y_new
else:
return substrs, X_train
def _create_basic_features(self):
"""
Создаёт базовые признаки ещё до чтения входа
Возвращает:
-----------
features: list, список созданных признаков
"""
self.features, self.feature_codes = [], dict()
# признаки для контекстов, не встречавшихся в обучающей выборке
# правые контексты
self.features.extend(
("-" * length + '#')
for length in range(1, self.right_context_length + 1))
# короткие правые контексты
self.features.extend(("-" * length + '$')
for length in range(self.right_context_length))
# левые контексты
self.features.extend(
('#' + "-" * length)
for length in range(1, self.right_context_length + 1))
# короткие левые контексты
self.features.extend(('^' + "-" * length)
for length in range(self.right_context_length))
if self.has_central_context:
self.features.append('-#-')
for code, feature in enumerate(self.features):
self.feature_codes[feature] = code
return
# def _create_new_feature(self, feature, is_full, side):
# if is_full:
# delim = '^' if side == 'left' else '$'
# to_break = True
# else:
# delim, to_break = '#', False
# if side == 'left':
# feature = delim + feature
# elif side == 'right':
# feature += delim
# feature_code = self.feature_codes.get(feature, None)
# if feature_code is None:
# feature_code = self._process_unknown_feature(feature, side)
# return feature_code, to_break
def _get_feature_code(self, feature, is_full, side, create_new=False):
"""
Возвращает код признака feature
"""
if is_full:
delim = '^' if side == 'left' else '$'
to_break = True
else:
delim, to_break = '#', False
if side == 'left':
feature = delim + feature
elif side == 'right':
feature += delim
code = self.feature_codes.get(feature, -1)
if code < 0:
if create_new:
self.features.append(feature)
self.feature_codes[feature] = code = self.features_number
self.features_number += 1
else:
if side == 'right':
partial_features = [
'-' * start + feature[start:]
for start in range(1, len(feature))
]
elif side == 'left':
partial_features = [
feature[:(len(feature) - start)] + '-' * start
for start in range(1, len(feature))
][::-1]
else: # side == 'center'
partial_features = []
code, to_break = -1, None
for partial_feature in partial_features:
code = self.feature_codes.get(partial_feature, -1)
if code > 0:
break
if code > 0:
self.feature_codes[feature] = code
return code, to_break