def test_check_X_enforce_min_columns():
X, y = make_classification_problem(n_columns=2)
msg = r"columns"
with pytest.raises(ValueError, match=msg):
check_X(X, enforce_min_columns=3)
with pytest.raises(ValueError, match=msg):
check_X_y(X, y, enforce_min_columns=3)
def test_check_X_enforce_univariate():
X, y = make_classification_problem(n_columns=2)
msg = r"univariate"
with pytest.raises(ValueError, match=msg):
check_X(X, enforce_univariate=True)
with pytest.raises(ValueError, match=msg):
check_X_y(X, y, enforce_univariate=True)
def test_check_enforce_min_instances():
X, y = make_classification_problem(n_instances=3)
msg = r"instance"
with pytest.raises(ValueError, match=msg):
check_X(X, enforce_min_instances=4)
with pytest.raises(ValueError, match=msg):
check_X_y(X, y, enforce_min_instances=4)
with pytest.raises(ValueError, match=msg):
check_y(y, enforce_min_instances=4)
def _set_oob_score(self, X, y):
"""Compute out-of-bag score"""
check_X_y(X, y)
check_X(X, enforce_univariate=True)
n_classes_ = self.n_classes_
n_samples = y.shape[0]
oob_decision_function = []
oob_score = 0.0
predictions = [
np.zeros((n_samples, n_classes_[k]))
for k in range(self.n_outputs_)
]
n_samples_bootstrap = _get_n_samples_bootstrap(n_samples,
self.max_samples)
for estimator in self.estimators_:
final_estimator = estimator.steps[-1][1]
unsampled_indices = _generate_unsampled_indices(
final_estimator.random_state, n_samples, n_samples_bootstrap)
p_estimator = estimator.predict_proba(X.iloc[unsampled_indices, :])
if self.n_outputs_ == 1:
p_estimator = [p_estimator]
for k in range(self.n_outputs_):
predictions[k][unsampled_indices, :] += p_estimator[k]
for k in range(self.n_outputs_):
if (predictions[k].sum(axis=1) == 0).any():
warn("Some inputs do not have OOB scores. "
"This probably means too few trees were used "
"to compute any reliable oob estimates.")
decision = (predictions[k] /
predictions[k].sum(axis=1)[:, np.newaxis])
oob_decision_function.append(decision)
oob_score += np.mean(y[:, k] == np.argmax(predictions[k], axis=1),
axis=0)
if self.n_outputs_ == 1:
self.oob_decision_function_ = oob_decision_function[0]
else:
self.oob_decision_function_ = oob_decision_function
self.oob_score_ = oob_score / self.n_outputs_
def fit(self, X, y):
"""Perform a shapelet transform then builds a random forest.
Contract default for ST is 5 hours
----------
X : array-like or sparse matrix of shape = [n_instances,
series_length] or shape = [n_instances,n_columns]
The training input samples. If a Pandas data frame is passed it
must have a single column (i.e. univariate
classification. RISE has no bespoke method for multivariate
classification as yet.
y : array-like, shape = [n_instances] The class labels.
Returns
-------
self : object
"""
X, y = check_X_y(X, y, enforce_univariate=True)
self.n_classes = np.unique(y).shape[0]
self.classes_ = class_distribution(np.asarray(y).reshape(-1, 1))[0][0]
self.classifier.fit(X, y)
# self.shapelet_transform.fit(X,y)
# print("Shapelet Search complete")
# self.st_X =self.shapelet_transform.transform(X)
# print("Transform complete")
# X = np.asarray([a.values for a in X.iloc[:, 0]])
# self.classifier.fit(X,y)
# print("Build classifier complete")
self._is_fitted = True
return self
def fit(self, X, y):
"""
Build the classifier on the training set (X, y)
----------
X : array-like or sparse matrix of shape = [n_instances, n_columns]
The training input samples. If a Pandas data frame is passed,
column 0 is extracted.
y : array-like, shape = [n_instances]
The class labels.
Returns
-------
self : object
"""
X, y = check_X_y(X, y, enforce_univariate=True)
self.X = dataset_properties.positive_dataframe_indices(X)
self.random_state = check_random_state(self.random_state)
# setup label encoding
if self.label_encoder is None:
self.label_encoder = LabelEncoder()
y = self.label_encoder.fit_transform(y)
self.y = y
self.classes_ = self.label_encoder.classes_
if self.distance_measure is None:
if self.get_distance_measure is None:
self.get_distance_measure = self.setup_distance_measure(self)
self.distance_measure = self.get_distance_measure(self)
self.X_exemplar, self.y_exemplar = self.pick_exemplars(self)
self._is_fitted = True
return self
def fit(self, X, y):
"""
Build the classifier on the training set (X, y)
----------
X : array-like or sparse matrix of shape = [n_instances, n_columns]
The training input samples. If a Pandas data frame is passed,
column 0 is extracted.
y : array-like, shape = [n_instances]
The class labels.
Returns
-------
self : object
"""
X, y = check_X_y(X, y, enforce_univariate=True)
self.X = dataset_properties.positive_dataframe_indices(X)
self.random_state = check_random_state(self.random_state)
if self.find_stump is None:
self.find_stump = best_of_n_stumps(self.n_stump_evaluations)
# setup label encoding
if self.label_encoder is None:
self.label_encoder = LabelEncoder()
y = self.label_encoder.fit_transform(y)
self.y = y
self.classes_ = self.label_encoder.classes_
if self.distance_measure is None:
if self.get_distance_measure is None:
self.get_distance_measure = self.setup_distance_measure(self)
self.distance_measure = self.get_distance_measure(self)
self.stump = self.find_stump(self)
n_branches = len(self.stump.y_exemplar)
self.branches = [None] * n_branches
if self.depth < self.max_depth:
for index in range(n_branches):
sub_y = self.stump.y_branches[index]
if not self.is_leaf(sub_y):
sub_tree = ProximityTree(
random_state=self.random_state,
get_exemplars=self.get_exemplars,
distance_measure=self.distance_measure,
setup_distance_measure=self.setup_distance_measure,
get_distance_measure=self.get_distance_measure,
get_gain=self.get_gain,
is_leaf=self.is_leaf,
verbosity=self.verbosity,
max_depth=self.max_depth,
n_jobs=self.n_jobs,
)
sub_tree.label_encoder = self.label_encoder
sub_tree.depth = self.depth + 1
self.branches[index] = sub_tree
sub_X = self.stump.X_branches[index]
sub_tree.fit(sub_X, sub_y)
self._is_fitted = True
return self
def check_and_clean_data(X, y=None, input_checks=True):
'''
Performs basic sktime data checks and prepares the train data for input to
Keras models.
:param X: the train data
:param y: teh train labels
:param input_checks: whether to perform the basic sktime checks
:return: X
'''
if input_checks:
if y is None:
check_X(X)
else:
check_X_y(X, y)
# want data in form: [instances = n][timepoints = m][dimensions = d]
if isinstance(X, pd.DataFrame):
if _is_nested_dataframe(X):
if X.shape[1] > 1:
# we have multiple columns, AND each cell contains a series,
# so this is a multidimensional problem
X = _multivariate_nested_df_to_array(X)
else:
# we have a single column containing a series, treat this as
# a univariate problem
X = _univariate_nested_df_to_array(X)
else:
# we have multiple columns each containing a primitive, treat as
# univariate series
X = _univariate_df_to_array(X)
if len(X.shape) == 2:
# add a dimension to make it multivariate with one dimension
X = X.reshape(
X.shape[0], X.shape[1], 1
) # go from [n][m] to [n][m][d=1]
return X
def fit(self, X, y):
X, y = check_X_y(X, y, enforce_univariate=True)
sfa = self.transformer.fit_transform(X, y)
self.transformed_data = [series.to_dict() for series in sfa.iloc[:, 0]]
self.class_vals = y
self.num_classes = np.unique(y).shape[0]
self.classes_ = class_distribution(np.asarray(y).reshape(-1, 1))[0][0]
for index, classVal in enumerate(self.classes_):
self.class_dictionary[classVal] = index
self._is_fitted = True
return self
def fit(self, X, y):
"""Build a forest of trees from the training set (X, y) using random
intervals and summary features
Parameters
----------
X : array-like or sparse matrix of shape = [n_instances,
series_length] or shape = [n_instances,n_columns]
The training input samples. If a Pandas data frame is passed it
must have a single column (i.e. univariate
classification. RISE has no bespoke method for multivariate
classification as yet.
y : array-like, shape = [n_instances] The class labels.
Returns
-------
self : object
"""
X, y = check_X_y(X, y, enforce_univariate=True, coerce_to_numpy=True)
X = X.squeeze(1)
n_instances, self.series_length = X.shape
rng = check_random_state(self.random_state)
self.n_classes = np.unique(y).shape[0]
self.classes_ = class_distribution(np.asarray(y).reshape(-1, 1))[0][0]
self.n_intervals = int(math.sqrt(self.series_length))
if self.n_intervals == 0:
self.n_intervals = 1
if self.series_length < self.min_interval:
self.min_interval = self.series_length
self.intervals_ = [
_get_intervals(self.n_intervals, self.min_interval,
self.series_length, rng)
for _ in range(self.n_estimators)
]
self.estimators_ = Parallel(n_jobs=self.n_jobs)(
delayed(_fit_estimator)(
X,
y,
self.base_estimator,
self.intervals_[i],
self.random_state,
) for i in range(self.n_estimators))
self._is_fitted = True
return self
def fit(self, X, y=None):
"""Fit.
Parameters
----------
X : pd.DataFrame
nested pandas DataFrame of shape [n_samples, n_columns]
y : pd.Series or np.array
Target variable
Returns
-------
self : an instance of self
"""
# lazy imports to avoid hard dependency
from tsfresh.transformers.feature_selector import FeatureSelector
# input checks
if y is None:
raise ValueError(
f"{self.__class__.__name__} requires `y` in `fit`.")
X, y = check_X_y(X, y, coerce_to_pandas=True)
self.extractor_ = TSFreshFeatureExtractor(
default_fc_parameters=self.default_fc_parameters,
kind_to_fc_parameters=self.kind_to_fc_parameters,
chunksize=self.chunksize,
n_jobs=self.n_jobs,
show_warnings=self.show_warnings,
disable_progressbar=self.disable_progressbar,
profiling=self.profiling,
profiling_filename=self.profiling_filename,
profiling_sorting=self.profiling_sorting,
)
selection_params = self._get_selection_params()
extraction_param = self._get_extraction_params()
self.selector_ = FeatureSelector(
n_jobs=extraction_param["n_jobs"],
chunksize=extraction_param["chunksize"],
ml_task=self.ml_task,
**selection_params,
)
Xt = self.extractor_.fit_transform(X)
self.selector_.fit(Xt, y)
self._is_fitted = True
return self
def _set_oob_score(self, X, y):
"""
Compute out-of-bag scores."""
X, y = check_X_y(X, y, enforce_univariate=True)
n_samples = y.shape[0]
predictions = np.zeros((n_samples, self.n_outputs_))
n_predictions = np.zeros((n_samples, self.n_outputs_))
n_samples_bootstrap = _get_n_samples_bootstrap(
n_samples, self.max_samples
)
for estimator in self.estimators_:
final_estimator = estimator.steps[-1][1]
unsampled_indices = _generate_unsampled_indices(
final_estimator.random_state, n_samples, n_samples_bootstrap)
p_estimator = estimator.predict(
X[unsampled_indices, :], check_input=False)
if self.n_outputs_ == 1:
p_estimator = p_estimator[:, np.newaxis]
predictions[unsampled_indices, :] += p_estimator
n_predictions[unsampled_indices, :] += 1
if (n_predictions == 0).any():
warn("Some inputs do not have OOB scores. "
"This probably means too few trees were used "
"to compute any reliable oob estimates.")
n_predictions[n_predictions == 0] = 1
predictions /= n_predictions
self.oob_prediction_ = predictions
if self.n_outputs_ == 1:
self.oob_prediction_ = \
self.oob_prediction_.reshape((n_samples, ))
self.oob_score_ = 0.0
for k in range(self.n_outputs_):
self.oob_score_ += r2_score(y[:, k],
predictions[:, k])
self.oob_score_ /= self.n_outputs_
def fit(self, X, y):
if isinstance(X, pd.Series) or isinstance(X, pd.DataFrame):
X, y = check_X_y(X, y, enforce_univariate=True)
X = tabularize(X, return_array=True)
sfa = self.transformer.fit_transform(X, y)
self.transformed_data = sfa[0] # .iloc[:, 0]
self.class_vals = y
self.num_classes = np.unique(y).shape[0]
self.classes_ = class_distribution(np.asarray(y).reshape(-1, 1))[0][0]
for index, classVal in enumerate(self.classes_):
self.class_dictionary[classVal] = index
self._is_fitted = True
return self
def fit(self, X, y):
"""
Method to perform training on the classifier.
Parameters
----------
X - pandas dataframe of training data of shape [n_instances,1].
y - list of class labels of shape [n_instances].
Returns
-------
self : the shapeDTW object
"""
# Perform preprocessing on params.
if not (isinstance(self.shape_descriptor_function, str)):
raise TypeError("shape_descriptor_function must be an 'str'. \
Found '" +
type(self.shape_descriptor_function).__name__ +
"' instead.")
X, y = check_X_y(X, y, enforce_univariate=False)
if self.metric_params is None:
self.metric_params = {}
# If the shape descriptor is 'compound',
# calculate the appropriate weighting_factor
if self.shape_descriptor_function == "compound":
self._calculate_weighting_factor_value(X, y)
# Fit the SlidingWindowSegmenter
sw = SlidingWindowSegmenter(self.subsequence_length)
sw.fit(X)
self.sw = sw
# Transform the training data.
X = self._preprocess(X)
# Fit the kNN classifier
self.knn = KNeighborsTimeSeriesClassifier(n_neighbors=self.n_neighbors)
self.knn.fit(X, y)
self.classes_ = self.knn.classes_
return self
def fit(self, X, y):
"""
Build the classifier on the training set (X, y)
----------
X : array-like or sparse matrix of shape = [n_instances, n_columns]
The training input samples. If a Pandas data frame is passed,
column 0 is extracted.
y : array-like, shape = [n_instances]
The class labels.
Returns
-------
self : object
"""
X, y = check_X_y(X, y, enforce_univariate=True)
self.X = dataset_properties.positive_dataframe_indices(X)
self.random_state = check_random_state(self.random_state)
# setup label encoding
if self.label_encoder is None:
self.label_encoder = LabelEncoder()
y = self.label_encoder.fit_transform(y)
self.y = y
self.classes_ = self.label_encoder.classes_
if self.distance_measure is None:
if self.get_distance_measure is None:
self.get_distance_measure = self.setup_distance_measure_getter(
self)
self.distance_measure = self.get_distance_measure(self)
if self.n_jobs > 1 or self.n_jobs < 0:
parallel = Parallel(self.n_jobs)
self.trees = parallel(
delayed(self._fit_tree)
(X, y, index, self.random_state.randint(0, self.n_estimators))
for index in range(self.n_estimators))
else:
self.trees = [
self._fit_tree(X, y, index,
self.random_state.randint(0, self.n_estimators))
for index in range(self.n_estimators)
]
self._is_fitted = True
return self
def fit(self, X, y):
"""Perform a shapelet transform then builds a random forest.
Contract default for ST is 5 hours
----------
X : array-like or sparse matrix of shape = [n_instances,
series_length] or shape = [n_instances,n_columns]
The training input samples. If a Pandas data frame is passed it
must have a single column (i.e. univariate
classification. RISE has no bespoke method for multivariate
classification as yet.
y : array-like, shape = [n_instances] The class labels.
Returns
-------
self : object
"""
X, y = check_X_y(X, y, enforce_univariate=True)
# if y is a pd.series then convert to array.
if isinstance(y, pd.Series):
y = y.to_numpy()
# generate pipeline in fit so that random state can be propogated properly.
self.classifier_ = Pipeline([
('st',
ContractedShapeletTransform(
time_contract_in_mins=self.time_contract_in_mins,
verbose=False,
random_state=self.random_state)),
('rf',
RandomForestClassifier(n_estimators=self.n_estimators,
random_state=self.random_state))
])
self.n_classes_ = np.unique(y).shape[0]
self.classes_ = class_distribution(np.asarray(y).reshape(-1, 1))[0][0]
self.classifier_.fit(X, y)
self._is_fitted = True
return self
def fit(self, X, y):
"""Build an ensemble of 1-NN classifiers from th training set (X, y),
Parameters
----------
X : array-like or sparse matrix of shape = [n_instances, n_columns]
The training input samples. If a Pandas data frame is passed,
it must have a single column. BOSS not configured
to handle multivariate
y : array-like, shape = [n_instances] The class labels.
Returns
-------
self : object
"""
X, y = check_X_y(X, y, enforce_univariate=True)
# Derivative DTW (DDTW) uses the regular DTW algorithm on data that
# are transformed into derivatives.
# To increase the efficiency of DDTW we can pre-transform the data
# into derivatives, and then call the
# standard DTW algorithm on it, rather than transforming each series
# every time a distance calculation
# is made. Please note that using DDTW elsewhere will not benefit
# from this speed enhancement
if self.distance_measures.__contains__(
ddtw_c) or self.distance_measures.__contains__(wddtw_c):
der_X = DerivativeSlopeTransformer().fit_transform(X)
# reshape X for use with the efficient cython distance measures
der_X = np.array(
[np.asarray([x]).reshape(len(x), 1) for x in der_X.iloc[:, 0]])
else:
der_X = None
# reshape X for use with the efficient cython distance measures
X = np.array(
[np.asarray([x]).reshape(len(x), 1) for x in X.iloc[:, 0]])
self.train_accs_by_classifier = np.zeros(len(self.distance_measures))
self.train_preds_by_classifier = [None] * len(self.distance_measures)
self.estimators_ = [None] * len(self.distance_measures)
self.classes_ = class_distribution(np.asarray(y).reshape(-1, 1))[0][0]
rand = np.random.RandomState(self.random_state)
# The default EE uses all training instances for setting parameters,
# and 100 parameter options per
# elastic measure. The prop_train_in_param_finding and
# prop_of_param_options attributes of this class
# can be used to control this however, using less cases to optimise
# parameters on the training data
# and/or using less parameter options.
#
# For using less training instances the appropriate number of cases
# must be sampled from the data.
# This is achieved through the use of a deterministic
# StratifiedShuffleSplit
#
# For using less parameter options a RandomizedSearchCV is used in
# place of a GridSearchCV
param_train_x = None
der_param_train_x = None
param_train_y = None
# If using less cases for parameter optimisation, use the
# StratifiedShuffleSplit:
if self.proportion_train_in_param_finding < 1:
if self.verbose > 0:
print(
"Restricting training cases for parameter optimisation: ",
end="")
sss = StratifiedShuffleSplit(
n_splits=1,
test_size=1 - self.proportion_train_in_param_finding,
random_state=rand)
for train_index, test_index in sss.split(X, y):
param_train_x = X[train_index, :]
param_train_y = y[train_index]
if der_X is not None:
der_param_train_x = der_X[train_index, :]
if self.verbose > 0:
print("using " + str(len(param_train_x)) +
" training cases instead of " + str(len(X)) +
" for parameter optimisation")
# else, use the full training data for optimising parameters
else:
if self.verbose > 0:
print("Using all training cases for parameter optimisation")
param_train_x = X
param_train_y = y
if der_X is not None:
der_param_train_x = der_X
self.constituent_build_times = []
if self.verbose > 0:
print("Using " + str(100 * self.proportion_of_param_options) +
" parameter "
"options per "
"measure")
for dm in range(0, len(self.distance_measures)):
this_measure = self.distance_measures[dm]
# uses the appropriate training data as required (either full or
# smaller sample as per the StratifiedShuffleSplit)
param_train_to_use = param_train_x
full_train_to_use = X
if this_measure is ddtw_c or dm is wddtw_c:
param_train_to_use = der_param_train_x
full_train_to_use = der_X
if this_measure is ddtw_c:
this_measure = dtw_c
elif this_measure is wddtw_c:
this_measure = wdtw_c
start_build_time = time.time()
if self.verbose > 0:
if self.distance_measures[dm] is ddtw_c or \
self.distance_measures[dm] is wddtw_c:
print("Currently evaluating " +
str(self.distance_measures[dm].__name__) +
" (implemented as " + str(this_measure.__name__) +
" with pre-transformed derivative data)")
else:
print("Currently evaluating " +
str(self.distance_measures[dm].__name__))
# If 100 parameter options are being considered per measure,
# use a GridSearchCV
if self.proportion_of_param_options == 1:
grid = GridSearchCV(
estimator=KNeighborsTimeSeriesClassifier(
metric=this_measure, n_neighbors=1, algorithm="brute"),
param_grid=ElasticEnsemble._get_100_param_options(
self.distance_measures[dm], X),
cv=LeaveOneOut(),
scoring='accuracy',
verbose=self.verbose)
grid.fit(param_train_to_use, param_train_y)
# Else, used RandomizedSearchCV to randomly sample parameter
# options for each measure
else:
grid = RandomizedSearchCV(
estimator=KNeighborsTimeSeriesClassifier(
metric=this_measure, n_neighbors=1, algorithm="brute"),
param_distributions=ElasticEnsemble._get_100_param_options(
self.distance_measures[dm], X),
cv=LeaveOneOut(),
scoring='accuracy',
n_iter=100 * self.proportion_of_param_options,
random_state=rand,
verbose=self.verbose)
grid.fit(param_train_to_use, param_train_y)
# once the best parameter option has been estimated on the
# training data, perform a final pass with this parameter option
# to get the individual predictions with cross_cal_predict (
# Note: optimisation potentially possible here if a GridSearchCV
# was used previously. TO-DO: determine how to extract
# predictions for the best param option from GridSearchCV)
best_model = KNeighborsTimeSeriesClassifier(
algorithm="brute",
n_neighbors=1,
metric=this_measure,
metric_params=grid.best_params_['metric_params'])
preds = cross_val_predict(best_model,
full_train_to_use,
y,
cv=LeaveOneOut())
acc = accuracy_score(y, preds)
if self.verbose > 0:
print("Training accuracy for " +
str(self.distance_measures[dm].__name__) + ": " +
str(acc) + " (with parameter setting: " +
str(grid.best_params_['metric_params']) + ")")
# Finally, reset the classifier for this measure and parameter
# option, ready to be called for test classification
best_model = KNeighborsTimeSeriesClassifier(
algorithm="brute",
n_neighbors=1,
metric=this_measure,
metric_params=grid.best_params_['metric_params'])
best_model.fit(full_train_to_use, y)
end_build_time = time.time()
self.constituent_build_times.append(
str(end_build_time - start_build_time))
self.estimators_[dm] = best_model
self.train_accs_by_classifier[dm] = acc
self.train_preds_by_classifier[dm] = preds
self._is_fitted = True
return self
def fit(self, X, y):
"""Fit the model using X as training data and y as target values
Parameters
----------
X : sktime-format pandas dataframe with shape([n_cases,n_dimensions]),
or numpy ndarray with shape([n_cases,n_readings,n_dimensions])
y : {array-like, sparse matrix}
Target values of shape = [n_samples]
"""
X, y = check_X_y(X, y, enforce_univariate=False)
y = np.asarray(y)
X = nested_to_3d_numpy(X)
check_classification_targets(y)
# print(X)
# if internal cv is desired, the relevant flag forces a grid search
# to evaluate the possible values,
# find the best, and then set this classifier's params to match
if self._cv_for_params:
grid = GridSearchCV(
estimator=KNeighborsTimeSeriesClassifier(metric=self.metric,
n_neighbors=1,
algorithm="brute"),
param_grid=self._param_matrix,
cv=LeaveOneOut(),
scoring='accuracy'
)
grid.fit(X, y)
self.metric_params = grid.best_params_['metric_params']
if y.ndim == 1 or y.ndim == 2 and y.shape[1] == 1:
if y.ndim != 1:
warnings.warn("A column-vector y was passed when a 1d array "
"was expected. Please change the shape of y to "
"(n_samples, ), for example using ravel().",
DataConversionWarning, stacklevel=2)
self.outputs_2d_ = False
y = y.reshape((-1, 1))
else:
self.outputs_2d_ = True
self.classes_ = []
self._y = np.empty(y.shape, dtype=np.int)
for k in range(self._y.shape[1]):
classes, self._y[:, k] = np.unique(y[:, k], return_inverse=True)
self.classes_.append(classes)
if not self.outputs_2d_:
self.classes_ = self.classes_[0]
self._y = self._y.ravel()
if hasattr(check_array, '__wrapped__'):
temp = check_array.__wrapped__.__code__
check_array.__wrapped__.__code__ = _check_array_ts.__code__
else:
temp = check_array.__code__
check_array.__code__ = _check_array_ts.__code__
fx = self._fit(X)
if hasattr(check_array, '__wrapped__'):
check_array.__wrapped__.__code__ = temp
else:
check_array.__code__ = temp
self._is_fitted = True
return fx
def test_check_X_bad_input_args(X):
with pytest.raises(ValueError):
check_X(X)
with pytest.raises(ValueError):
check_X_y(X, y)
def fit(self, X, y):
"""Build a WEASEL+MUSE classifiers from the training set (X, y),
Parameters
----------
X : nested pandas DataFrame of shape [n_instances, 1]
Nested dataframe with univariate time-series in cells.
y : array-like, shape = [n_instances] The class labels.
Returns
-------
self : object
"""
X, y = check_X_y(X, y, coerce_to_pandas=True)
y = np.asarray(y)
# add first order differences in each dimension to TS
if self.use_first_order_differences:
X = self.add_first_order_differences(X)
# Window length parameter space dependent on series length
self.col_names = X.columns
rng = check_random_state(self.random_state)
self.n_dims = len(self.col_names)
self.highest_dim_bit = (math.ceil(math.log2(self.n_dims))) + 1
self.highest_bits = np.zeros(self.n_dims)
self.SFA_transformers = [[] for _ in range(self.n_dims)]
# the words of all dimensions and all time series
all_words = [dict() for _ in range(X.shape[0])]
# On each dimension, perform SFA
for ind, column in enumerate(self.col_names):
X_dim = X[[column]]
X_dim = from_nested_to_3d_numpy(X_dim)
series_length = X_dim.shape[
-1] # TODO compute minimum over all ts ?
# increment window size in steps of 'win_inc'
win_inc = self.compute_window_inc(series_length)
self.max_window = int(min(series_length, self.max_window))
self.window_sizes.append(
list(range(self.min_window, self.max_window, win_inc)))
self.highest_bits[ind] = math.ceil(math.log2(self.max_window)) + 1
for window_size in self.window_sizes[ind]:
transformer = SFA(
word_length=rng.choice(self.word_lengths),
alphabet_size=self.alphabet_size,
window_size=window_size,
norm=rng.choice(self.norm_options),
anova=self.anova,
binning_method=rng.choice(self.binning_strategies),
bigrams=self.bigrams,
remove_repeat_words=False,
lower_bounding=False,
save_words=False,
)
sfa_words = transformer.fit_transform(X_dim, y)
self.SFA_transformers[ind].append(transformer)
bag = sfa_words[0] # .iloc[:, 0]
# chi-squared test to keep only relevant features
relevant_features = {}
apply_chi_squared = self.chi2_threshold > 0
if apply_chi_squared:
bag_vec = DictVectorizer(sparse=False).fit_transform(bag)
chi2_statistics, p = chi2(bag_vec, y)
relevant_features = np.where(
chi2_statistics >= self.chi2_threshold)[0]
# merging bag-of-patterns of different window_sizes
# to single bag-of-patterns with prefix indicating
# the used window-length
highest = np.int32(self.highest_bits[ind])
for j in range(len(bag)):
for (key, value) in bag[j].items():
# chi-squared test
if (not apply_chi_squared) or (key
in relevant_features):
# append the prefices to the words to
# distinguish between window-sizes
word = MUSE.shift_left(key, highest, ind,
self.highest_dim_bit,
window_size)
all_words[j][word] = value
self.clf = make_pipeline(
DictVectorizer(sparse=False),
StandardScaler(with_mean=True, copy=False),
LogisticRegression(
max_iter=5000,
solver="liblinear",
dual=True,
# class_weight="balanced",
penalty="l2",
random_state=self.random_state,
),
)
self.clf.fit(all_words, y)
self._is_fitted = True
return self
def fit(self, X, y=None):
"""A method to fit the shapelet transform to a specified X and y
Parameters
----------
X: pandas DataFrame
The training input samples.
y: array-like or list
The class values for X
Returns
-------
self : FullShapeletTransform
This estimator
"""
X, y = check_X_y(X, y, enforce_univariate=True)
# if y is a pd.series then convert to array.
if isinstance(y, pd.Series):
y = y.to_numpy()
if type(
self) is ContractedShapeletTransform and \
self.time_contract_in_mins <= 0:
raise ValueError(
"Error: time limit cannot be equal to or less than 0")
X_lens = np.array([
len(X.iloc[r, 0]) for r in range(len(X))
]) # note, assumes all dimensions of a case are the same
# length. A shapelet would not be well defined if indices do not match!
X = np.array([[X.iloc[r, c].values for c in range(len(X.columns))]
for r in range(len(X))
]) # may need to pad with nans here for uneq length,
# look at later
num_ins = len(y)
distinct_class_vals = \
class_distribution(np.asarray(y).reshape(-1, 1))[0][0]
candidates_evaluated = 0
if type(self) is _RandomEnumerationShapeletTransform:
num_series_to_visit = min(self.num_cases_to_sample, len(y))
else:
num_series_to_visit = num_ins
shapelet_heaps_by_class = {
i: ShapeletPQ()
for i in distinct_class_vals
}
self._random_state = check_random_state(self.random_state)
# Here we establish the order of cases to sample. We need to sample
# x cases and y shapelets from each (where x = num_cases_to_sample
# and y = num_shapelets_to_sample_per_case). We could simply sample
# x cases without replacement and y shapelets from each case, but
# the idea is that if we are using a time contract we may extract
# all y shapelets from each x candidate and still have time remaining.
# Therefore, if we get a list of the indices of the series and
# shuffle them appropriately, we can go through the list again and
# extract
# another y shapelets from each series (if we have time).
# We also want to ensure that we visit all classes so we will visit
# in round-robin order. Therefore, the code below extracts the indices
# of all series by class, shuffles the indices for each class
# independently, and then combines them in alternating order. This
# results in
# a shuffled list of indices that are in alternating class order (
# e.g. 1,2,3,1,2,3,1,2,3,1...)
def _round_robin(*iterables):
sentinel = object()
return (a for x in zip_longest(*iterables, fillvalue=sentinel)
for a in x if a != sentinel)
case_ids_by_class = {
i: np.where(y == i)[0]
for i in distinct_class_vals
}
# if transform is random/contract then shuffle the data initially
# when determining which cases to visit
if type(self) is _RandomEnumerationShapeletTransform or type(
self) is ContractedShapeletTransform:
for i in range(len(distinct_class_vals)):
self._random_state.shuffle(
case_ids_by_class[distinct_class_vals[i]])
num_train_per_class = {
i: len(case_ids_by_class[i])
for i in case_ids_by_class
}
round_robin_case_order = _round_robin(
*[list(v) for k, v in case_ids_by_class.items()])
cases_to_visit = [(i, y[i]) for i in round_robin_case_order]
# this dictionary will be used to store all possible starting
# positions and shapelet lengths for a give series length. This
# is because we enumerate all possible candidates and sample without
# replacement when assessing a series. If we have two series
# of the same length then they will obviously have the same valid
# shapelet starting positions and lengths (especially in standard
# datasets where all series are equal length) so it makes sense to
# store the possible candidates and reuse, rather than
# recalculating each time
# Initially the dictionary will be empty, and each time a new series
# length is seen the dict will be updated. Next time that length
# is used the dict will have an entry so can simply reuse
possible_candidates_per_series_length = {}
# a flag to indicate if extraction should stop (contract has ended)
time_finished = False
# max time calculating a shapelet
# for timing the extraction when contracting
start_time = time.time()
def time_taken():
return time.time() - start_time
max_time_calc_shapelet = -1
time_last_shapelet = time_taken()
# for every series
case_idx = 0
while case_idx < len(cases_to_visit):
series_id = cases_to_visit[case_idx][0]
this_class_val = cases_to_visit[case_idx][1]
# minus 1 to remove this candidate from sums
binary_ig_this_class_count = num_train_per_class[this_class_val] - 1
binary_ig_other_class_count = (num_ins -
binary_ig_this_class_count - 1)
if self.verbose:
if type(self) == _RandomEnumerationShapeletTransform:
print("visiting series: " + str(series_id) + " (#" +
str(case_idx + 1) + "/" + str(num_series_to_visit) +
")")
else:
print("visiting series: " + str(series_id) + " (#" +
str(case_idx + 1) + ")")
this_series_len = len(X[series_id][0])
# The bound on possible shapelet lengths will differ
# series-to-series if using unequal length data.
# However, shapelets cannot be longer than the series, so set to
# the minimum of the series length
# and max shapelet length (which is inf by default)
if self.max_shapelet_length == -1:
this_shapelet_length_upper_bound = this_series_len
else:
this_shapelet_length_upper_bound = min(
this_series_len, self.max_shapelet_length)
# all possible start and lengths for shapelets within this
# series (calculates if series length is new, a simple look-up
# if not)
# enumerate all possible candidate starting positions and lengths.
# First, try to reuse if they have been calculated for a series
# of the same length before.
candidate_starts_and_lens = \
possible_candidates_per_series_length.get(
this_series_len)
# else calculate them for this series length and store for
# possible use again
if candidate_starts_and_lens is None:
candidate_starts_and_lens = [
[start, length] for start in range(
0, this_series_len - self.min_shapelet_length + 1)
for length in range(self.min_shapelet_length,
this_shapelet_length_upper_bound + 1)
if start + length <= this_series_len
]
possible_candidates_per_series_length[
this_series_len] = candidate_starts_and_lens
# default for full transform
candidates_to_visit = candidate_starts_and_lens
num_candidates_per_case = len(candidate_starts_and_lens)
# limit search otherwise:
if hasattr(self, "num_candidates_to_sample_per_case"):
num_candidates_per_case = min(
self.num_candidates_to_sample_per_case,
num_candidates_per_case)
cand_idx = list(
self._random_state.choice(list(
range(0, len(candidate_starts_and_lens))),
num_candidates_per_case,
replace=False))
candidates_to_visit = [
candidate_starts_and_lens[x] for x in cand_idx
]
for candidate_idx in range(num_candidates_per_case):
# if shapelet heap for this class is not full yet, set entry
# criteria to be the predetermined IG threshold
ig_cutoff = self.predefined_ig_rejection_level
# otherwise if we have max shapelets already, set the
# threshold as the IG of the current 'worst' shapelet we have
if shapelet_heaps_by_class[
this_class_val].get_size() >= \
self.max_shapelets_to_store_per_class:
ig_cutoff = max(
shapelet_heaps_by_class[this_class_val].peek()[0],
ig_cutoff)
cand_start_pos = candidates_to_visit[candidate_idx][0]
cand_len = candidates_to_visit[candidate_idx][1]
candidate = ShapeletTransform.zscore(
X[series_id][:, cand_start_pos:cand_start_pos + cand_len])
# now go through all other series and get a distance from
# the candidate to each
orderline = []
# initialise here as copy, decrease the new val each time we
# evaluate a comparison series
num_visited_this_class = 0
num_visited_other_class = 0
candidate_rejected = False
for comparison_series_idx in range(len(cases_to_visit)):
i = cases_to_visit[comparison_series_idx][0]
if y[i] != cases_to_visit[comparison_series_idx][1]:
raise ValueError("class match sanity test broken")
if i == series_id:
# don't evaluate candidate against own series
continue
if y[i] == this_class_val:
num_visited_this_class += 1
binary_class_identifier = 1 # positive for this class
else:
num_visited_other_class += 1
binary_class_identifier = -1 # negative for any
# other class
bsf_dist = np.inf
start_left = cand_start_pos
start_right = cand_start_pos + 1
if X_lens[i] == cand_len:
start_left = 0
start_right = 0
for num_cals in range(
max(1, int(np.ceil(
(X_lens[i] - cand_len) / 2)))): # max
# used to force iteration where series len ==
# candidate len
if start_left < 0:
start_left = X_lens[i] - 1 - cand_len
comparison = ShapeletTransform.zscore(
X[i][:, start_left:start_left + cand_len])
dist_left = np.linalg.norm(candidate - comparison)
bsf_dist = min(dist_left * dist_left, bsf_dist)
# for odd lengths
if start_left == start_right:
continue
# right
if start_right == X_lens[i] - cand_len + 1:
start_right = 0
comparison = ShapeletTransform.zscore(
X[i][:, start_right:start_right + cand_len])
dist_right = np.linalg.norm(candidate - comparison)
bsf_dist = min(dist_right * dist_right, bsf_dist)
start_left -= 1
start_right += 1
orderline.append((bsf_dist, binary_class_identifier))
# sorting required after each add for early IG abandon.
# timsort should be efficient as array is almost in
# order - insertion-sort like behaviour in this case.
# Can't use heap as need to traverse in order multiple
# times, not just access root
orderline.sort()
if len(orderline) > 2:
ig_upper_bound = \
ShapeletTransform.calc_early_binary_ig(
orderline, num_visited_this_class,
num_visited_other_class,
binary_ig_this_class_count -
num_visited_this_class,
binary_ig_other_class_count -
num_visited_other_class)
# print("upper: "+str(ig_upper_bound))
if ig_upper_bound <= ig_cutoff:
candidate_rejected = True
break
candidates_evaluated += 1
if self.verbose > 3 and candidates_evaluated % 100 == 0:
print("candidates evaluated: " + str(candidates_evaluated))
# only do if candidate was not rejected
if candidate_rejected is False:
final_ig = ShapeletTransform.calc_binary_ig(
orderline, binary_ig_this_class_count,
binary_ig_other_class_count)
accepted_candidate = Shapelet(series_id, cand_start_pos,
cand_len, final_ig,
candidate)
# add to min heap to store shapelets for this class
shapelet_heaps_by_class[this_class_val].push(
accepted_candidate)
# informal, but extra 10% allowance for self similar later
if shapelet_heaps_by_class[
this_class_val].get_size() > \
self.max_shapelets_to_store_per_class * 3:
shapelet_heaps_by_class[this_class_val].pop()
# Takes into account the use of the MAX shapelet calculation
# time to not exceed the time_limit (not exact, but likely a
# good guess).
if hasattr(self,
'time_contract_in_mins') and \
self.time_contract_in_mins \
> 0:
time_now = time_taken()
time_this_shapelet = (time_now - time_last_shapelet)
if time_this_shapelet > max_time_calc_shapelet:
max_time_calc_shapelet = time_this_shapelet
print(max_time_calc_shapelet)
time_last_shapelet = time_now
# add a little 1% leeway to the timing incase one run was slightly faster than
# another based on the CPU.
time_in_seconds = self.time_contract_in_mins * 60
max_shapelet_time_percentage = (max_time_calc_shapelet /
100.0) * 0.75
if (time_now + max_shapelet_time_percentage) > \
time_in_seconds:
if self.verbose > 0:
print("No more time available! It's been {0:02d}:{"
"1:02}".format(
int(round(time_now / 60, 3)),
int((round(time_now / 60, 3) -
int(round(time_now / 60, 3))) *
60)))
time_finished = True
break
else:
if self.verbose > 0:
if candidate_rejected is False:
print(
"Candidate finished. {0:02d}:{1:02} "
"remaining".format(
int(
round(
self.time_contract_in_mins -
time_now / 60, 3)),
int((round(
self.time_contract_in_mins -
time_now / 60, 3) - int(
round(
(self.time_contract_in_mins
- time_now) / 60, 3))) *
60)))
else:
print(
"Candidate rejected. {0:02d}:{1:02} "
"remaining".format(
int(
round((self.time_contract_in_mins -
time_now) / 60, 3)),
int((round(
(self.time_contract_in_mins -
time_now) / 60,
3) - int(
round(
(self.time_contract_in_mins
- time_now) / 60, 3))) *
60)))
# stopping condition: in case of iterative transform (i.e.
# num_cases_to_sample have been visited)
# in case of contracted transform (i.e. time
# limit has been reached)
case_idx += 1
if case_idx >= num_series_to_visit:
if hasattr(self,
'time_contract_in_mins') and time_finished is not \
True:
case_idx = 0
elif case_idx >= num_series_to_visit or time_finished:
if self.verbose > 0:
print("Stopping search")
break
# remove self similar here
# for each class value
# get list of shapelets
# sort by quality
# remove self similar
self.shapelets = []
for class_val in distinct_class_vals:
by_class_descending_ig = sorted(
shapelet_heaps_by_class[class_val].get_array(),
key=itemgetter(0),
reverse=True)
if self.remove_self_similar and len(by_class_descending_ig) > 0:
by_class_descending_ig = \
ShapeletTransform.remove_self_similar_shapelets(
by_class_descending_ig)
else:
# need to extract shapelets from tuples
by_class_descending_ig = [x[2] for x in by_class_descending_ig]
# if we have more than max_shapelet_per_class, trim to that
# amount here
if len(by_class_descending_ig) > \
self.max_shapelets_to_store_per_class:
max_n = self.max_shapelets_to_store_per_class
by_class_descending_ig = by_class_descending_ig[:max_n]
self.shapelets.extend(by_class_descending_ig)
# final sort so that all shapelets from all classes are in
# descending order of information gain
self.shapelets.sort(key=lambda x: x.info_gain, reverse=True)
self.is_fitted_ = True
# warn the user if fit did not produce any valid shapelets
if len(self.shapelets) == 0:
warnings.warn(
"No valid shapelets were extracted from this dataset and "
"calling the transform method "
"will raise an Exception. Please re-fit the transform with "
"other data and/or "
"parameter options.")
self._is_fitted = True
return self
def fit(self, X, y):
"""Build a forest of trees from the training set (X, y) using random
intervals and summary features
Parameters
----------
X : array-like or sparse matrix of shape = [n_instances,
series_length] or shape = [n_instances,n_columns]
The training input samples. If a Pandas data frame is passed it
must have a single column (i.e. univariate
classification. RISE has no bespoke method for multivariate
classification as yet.
y : array-like, shape = [n_instances] The class labels.
Returns
-------
self : object
"""
X, y = check_X_y(X, y, enforce_univariate=True)
X = tabularize(X, return_array=True)
n_instances, self.series_length = X.shape
rng = check_random_state(self.random_state)
self.n_classes = np.unique(y).shape[0]
self.classes_ = class_distribution(np.asarray(y).reshape(-1, 1))[0][0]
self.n_intervals = int(math.sqrt(self.series_length))
if self.n_intervals == 0:
self.n_intervals = 1
if self.series_length < self.min_interval:
self.min_interval = self.series_length
self.intervals = np.zeros((self.n_estimators, self.n_intervals, 2),
dtype=int)
for i in range(self.n_estimators):
transformed_x = np.empty(shape=(3 * self.n_intervals, n_instances))
# Find the random intervals for classifier i and concatentate
# features
for j in range(self.n_intervals):
self.intervals[i][j][0] = rng.randint(self.series_length -
self.min_interval)
length = rng.randint(self.series_length -
self.intervals[i][j][0] - 1)
if length < self.min_interval:
length = self.min_interval
self.intervals[i][j][1] = self.intervals[i][j][0] + length
# Transforms here, just hard coding it, so not configurable
means = np.mean(
X[:, self.intervals[i][j][0]:self.intervals[i][j][1]],
axis=1)
std_dev = np.std(
X[:, self.intervals[i][j][0]:self.intervals[i][j][1]],
axis=1)
slope = self._lsq_fit(
X[:, self.intervals[i][j][0]:self.intervals[i][j][1]])
transformed_x[3 * j] = means
transformed_x[3 * j + 1] = std_dev
transformed_x[3 * j + 2] = slope
tree = clone(self.base_estimator)
tree.set_params(**{"random_state": self.random_state})
transformed_x = transformed_x.T
tree.fit(transformed_x, y)
self.classifiers.append(tree)
self._is_fitted = True
return self
def fit(self, X, y, sample_weight=None):
"""
Build a forest of trees from the training set (X, y).
Parameters
----------
X : array-like or sparse matrix of shape (n_samples, n_features)
The training input samples. Internally, its dtype will be converted
to ``dtype=np.float32``. If a sparse matrix is provided, it will be
converted into a sparse ``csc_matrix``.
y : array-like of shape (n_samples,) or (n_samples, n_outputs)
The target values (class labels in classification, real numbers in
regression).
sample_weight : array-like of shape (n_samples,), default=None
Sample weights. If None, then samples are equally weighted. Splits
that would create child nodes with net zero or negative weight are
ignored while searching for a split in each node. In the case of
classification, splits are also ignored if they would result in any
single class carrying a negative weight in either child node.
Returns
-------
self : object
"""
X, y = check_X_y(X, y, enforce_univariate=True)
# Validate or convert input data
if sample_weight is not None:
sample_weight = check_array(sample_weight, ensure_2d=False)
if issparse(X):
# Pre-sort indices to avoid that each individual tree of the
# ensemble sorts the indices.
X.sort_indices()
# Remap output
self.n_columns = X.shape[1]
self.n_features_ = X.shape[1] if X.ndim == 2 else 1
y = np.atleast_1d(y)
if y.ndim == 2 and y.shape[1] == 1:
warn(
"A column-vector y was passed when a 1d array was"
" expected. Please change the shape of y to "
"(n_samples,), for example using ravel().",
DataConversionWarning,
stacklevel=2)
if y.ndim == 1:
# reshape is necessary to preserve the data contiguity against vs
# [:, np.newaxis] that does not.
y = np.reshape(y, (-1, 1))
self.n_outputs_ = y.shape[1]
y, expanded_class_weight = self._validate_y_class_weight(y)
if getattr(y, "dtype", None) != DOUBLE or not y.flags.contiguous:
y = np.ascontiguousarray(y, dtype=DOUBLE)
if expanded_class_weight is not None:
if sample_weight is not None:
sample_weight = sample_weight * expanded_class_weight
else:
sample_weight = expanded_class_weight
# Get bootstrap sample size
n_samples_bootstrap = _get_n_samples_bootstrap(
n_samples=X.shape[0], max_samples=self.max_samples)
# Check parameters
self._validate_estimator()
if not self.bootstrap and self.oob_score:
raise ValueError("Out of bag estimation only available"
" if bootstrap=True")
random_state = check_random_state(self.random_state)
if not self.warm_start or not hasattr(self, "estimators_"):
# Free allocated memory, if any
self.estimators_ = []
n_more_estimators = self.n_estimators - len(self.estimators_)
if n_more_estimators < 0:
raise ValueError('n_estimators=%d must be larger or equal to '
'len(estimators_)=%d when warm_start==True' %
(self.n_estimators, len(self.estimators_)))
elif n_more_estimators == 0:
warn("Warm-start fitting without increasing n_estimators does not "
"fit new trees.")
else:
if self.warm_start and len(self.estimators_) > 0:
# We draw from the random state to get the random state we
# would have got if we hadn't used a warm_start.
random_state.randint(MAX_INT, size=len(self.estimators_))
trees = [
self._make_estimator(append=False, random_state=random_state)
for i in range(n_more_estimators)
]
# Parallel loop: for standard random forests, the threading
# backend is preferred as the Cython code for fitting the trees
# is internally releasing the Python GIL making threading more
# efficient than multiprocessing in that case.
# However, in this case,for fitting pipelines in parallel,
# multiprocessing is more efficient.
trees = Parallel(n_jobs=self.n_jobs, verbose=self.verbose)(
delayed(_parallel_build_trees)(
t,
self,
X,
y,
sample_weight,
i,
len(trees),
verbose=self.verbose,
class_weight=self.class_weight,
n_samples_bootstrap=n_samples_bootstrap)
for i, t in enumerate(trees))
# Collect newly grown trees
self.estimators_.extend(trees)
if self.oob_score:
self._set_oob_score(X, y)
# Decapsulate classes_ attributes
if hasattr(self, "classes_") and self.n_outputs_ == 1:
self.n_classes_ = self.n_classes_[0]
self.classes_ = self.classes_[0]
self._is_fitted = True
return self
def fit(self, X, y):
"""Build a WEASEL classifiers from the training set (X, y),
Parameters
----------
X : nested pandas DataFrame of shape [n_instances, 1]
Nested dataframe with univariate time-series in cells.
y : array-like, shape = [n_instances] The class labels.
Returns
-------
self : object
"""
X, y = check_X_y(X, y, enforce_univariate=True, coerce_to_numpy=True)
# Window length parameter space dependent on series length
self.n_instances, self.series_length = X.shape[0], X.shape[-1]
win_inc = self.compute_window_inc()
self.max_window = int(min(self.series_length, self.max_window))
self.window_sizes = list(
range(self.min_window, self.max_window, win_inc))
self.highest_bit = (math.ceil(math.log2(self.max_window))) + 1
rng = check_random_state(self.random_state)
all_words = [dict() for x in range(len(X))]
for window_size in self.window_sizes:
transformer = SFA(
word_length=rng.choice(self.word_lengths),
alphabet_size=self.alphabet_size,
window_size=window_size,
norm=rng.choice(self.norm_options),
anova=self.anova,
# levels=rng.choice([1, 2, 3]),
binning_method=self.binning_strategy,
bigrams=self.bigrams,
remove_repeat_words=False,
lower_bounding=False,
save_words=False,
)
sfa_words = transformer.fit_transform(X, y)
self.SFA_transformers.append(transformer)
bag = sfa_words[0] # .iloc[:, 0]
# chi-squared test to keep only relevant features
relevant_features = {}
apply_chi_squared = self.chi2_threshold > 0
if apply_chi_squared:
bag_vec = DictVectorizer(sparse=False).fit_transform(bag)
chi2_statistics, p = chi2(bag_vec, y)
relevant_features = np.where(
chi2_statistics >= self.chi2_threshold)[0]
# merging bag-of-patterns of different window_sizes
# to single bag-of-patterns with prefix indicating
# the used window-length
for j in range(len(bag)):
for (key, value) in bag[j].items():
# chi-squared test
if (not apply_chi_squared) or (key in relevant_features):
# append the prefices to the words to
# distinguish between window-sizes
if isinstance(key, tuple):
word = (((key[0] << self.highest_bit) | key[1]) <<
3) | window_size
else:
# word = ((key << self.highest_bit) << 3) \
# | window_size
word = WEASEL.shift_left(key, self.highest_bit,
window_size)
all_words[j][word] = value
self.clf = make_pipeline(
DictVectorizer(sparse=False),
StandardScaler(with_mean=True, copy=False),
LogisticRegression(
max_iter=5000,
solver="liblinear",
dual=True,
# class_weight="balanced",
penalty="l2",
random_state=self.random_state,
),
)
self.clf.fit(all_words, y)
self._is_fitted = True
return self
def fit(self, X, y):
"""Build a forest of trees from the training set (X, y) using random
intervals and spectral features
Parameters
----------
X : array-like or sparse matrix of shape = [n_instances,
series_length] or shape = [n_instances,n_columns]
The training input samples. If a Pandas data frame is passed it
must have a single column (i.e. univariate
classification. RISE has no bespoke method for multivariate
classification as yet.
y : array-like, shape = [n_instances] The class labels.
Returns
-------
self : object
"""
X, y = check_X_y(X, y, enforce_univariate=True, coerce_to_numpy=True)
X = X.squeeze(1)
n_instances, self.series_length = X.shape
rng = check_random_state(self.random_state)
self.estimators_ = []
self.n_classes = np.unique(y).shape[0]
self.classes_ = class_distribution(np.asarray(y).reshape(-1, 1))[0][0]
self.intervals = np.zeros((self.n_estimators, 2), dtype=int)
self.intervals[0][0] = 0
self.intervals[0][1] = self.series_length
for i in range(1, self.n_estimators):
self.intervals[i][0] = rng.randint(self.series_length -
self.min_interval)
self.intervals[i][1] = rng.randint(
self.intervals[i][0] + self.min_interval, self.series_length)
# Check lag against global properties
self.acf_lag_ = self.acf_lag
if self.acf_lag > self.series_length - self.acf_min_values:
self.acf_lag_ = self.series_length - self.acf_min_values
if self.acf_lag < 0:
self.acf_lag_ = 1
self.lags = np.zeros(self.n_estimators, dtype=int)
for i in range(0, self.n_estimators):
temp_lag = self.acf_lag_
if (temp_lag > self.intervals[i][1] - self.intervals[i][0] -
self.acf_min_values):
temp_lag = (self.intervals[i][1] - self.intervals[i][0] -
self.acf_min_values)
if temp_lag < 0:
temp_lag = 1
self.lags[i] = int(temp_lag)
acf_x = np.empty(shape=(n_instances, self.lags[i]))
ps_len = (self.intervals[i][1] - self.intervals[i][0]) / 2
ps_x = np.empty(shape=(n_instances, int(ps_len)))
for j in range(0, n_instances):
acf_x[j] = acf(X[j, self.intervals[i][0]:self.intervals[i][1]],
temp_lag)
ps_x[j] = ps(X[j, self.intervals[i][0]:self.intervals[i][1]])
transformed_x = np.concatenate((acf_x, ps_x), axis=1)
# transformed_x=acf_x
tree = clone(self.base_estimator)
# set random state, but not the same, so that estimators vary
tree.set_params(
**{"random_state": rng.randint(np.iinfo(np.int32).max)})
tree.fit(transformed_x, y)
self.estimators_.append(tree)
self._is_fitted = True
return self
def fit(self, X, y):
"""Build an ensemble of individual TDE classifiers from the training
set (X,y), through randomising over the parameter space to a set
number of times then selecting new parameters using Gaussian
processes
Parameters
----------
X : nested pandas DataFrame of shape [n_instances, 1]
Nested dataframe with univariate time-series in cells.
y : array-like, shape = [n_instances] The class labels.
Returns
-------
self : object
"""
X, y = check_X_y(X, y, enforce_univariate=True)
y = y.values if isinstance(y, pd.Series) else y
self.time_limit = self.time_limit * 60
self.n_instances, self.series_length = X.shape[0], len(X.iloc[0, 0])
self.n_classes = np.unique(y).shape[0]
self.classes_ = class_distribution(np.asarray(y).reshape(-1, 1))[0][0]
for index, classVal in enumerate(self.classes_):
self.class_dictionary[classVal] = index
self.classifiers = []
self.weights = []
self.prev_parameters_x = []
self.prev_parameters_y = []
# Window length parameter space dependent on series length
max_window_searches = self.series_length / 4
max_window = int(self.series_length * self.max_win_len_prop)
win_inc = int((max_window - self.min_window) / max_window_searches)
if win_inc < 1:
win_inc = 1
possible_parameters = self._unique_parameters(max_window, win_inc)
num_classifiers = 0
start_time = time.time()
train_time = 0
subsample_size = int(self.n_instances * 0.7)
lowest_acc = 0
lowest_acc_idx = 0
if self.time_limit > 0:
self.n_parameter_samples = 0
rng = check_random_state(self.random_state)
while (train_time < self.time_limit or num_classifiers <
self.n_parameter_samples) and len(possible_parameters) > 0:
if num_classifiers < self.randomly_selected_params:
parameters = possible_parameters.pop(
rng.randint(0, len(possible_parameters)))
else:
gp = GaussianProcessRegressor(random_state=self.random_state)
gp.fit(self.prev_parameters_x, self.prev_parameters_y)
preds = gp.predict(possible_parameters)
parameters = possible_parameters.pop(
rng.choice(np.flatnonzero(preds == preds.max())))
subsample = rng.choice(self.n_instances,
size=subsample_size,
replace=False)
X_subsample = X.iloc[subsample, :]
y_subsample = y[subsample]
tde = IndividualTDE(*parameters,
alphabet_size=self.alphabet_size,
random_state=self.random_state)
tde.fit(X_subsample, y_subsample)
tde.accuracy = self._individual_train_acc(tde, y_subsample,
subsample_size,
lowest_acc)
weight = math.pow(tde.accuracy, 4)
if num_classifiers < self.max_ensemble_size:
if tde.accuracy < lowest_acc:
lowest_acc = tde.accuracy
lowest_acc_idx = num_classifiers
self.weights.append(weight)
self.classifiers.append(tde)
elif tde.accuracy > lowest_acc:
self.weights[lowest_acc_idx] = weight
self.classifiers[lowest_acc_idx] = tde
lowest_acc, lowest_acc_idx = self._worst_ensemble_acc()
self.prev_parameters_x.append(parameters)
self.prev_parameters_y.append(tde.accuracy)
num_classifiers += 1
train_time = time.time() - start_time
self.n_estimators = len(self.classifiers)
self.weight_sum = np.sum(self.weights)
self._is_fitted = True
return self
def fit(self, X, y):
"""Build an ensemble of BOSS classifiers from the training set (X,
y), either through randomising over the para
space to make a fixed size ensemble quickly or by creating a
variable size ensemble of those within a threshold
of the best
Parameters
----------
X : nested pandas DataFrame of shape [n_instances, 1]
Nested dataframe with univariate time-series in cells.
y : array-like, shape = [n_instances] The class labels.
Returns
-------
self : object
"""
X, y = check_X_y(X, y, enforce_univariate=True)
y = y.values if isinstance(y, pd.Series) else y
self.time_limit = self.time_limit * 60
self.n_instances, self.series_length = X.shape[0], len(X.iloc[0, 0])
self.n_classes = np.unique(y).shape[0]
self.classes_ = class_distribution(np.asarray(y).reshape(-1, 1))[0][0]
for index, classVal in enumerate(self.classes_):
self.class_dictionary[classVal] = index
self.classifiers = []
self.weights = []
# Window length parameter space dependent on series length
max_window_searches = self.series_length / 4
max_window = int(self.series_length * self.max_win_len_prop)
win_inc = int((max_window - self.min_window) / max_window_searches)
if win_inc < 1:
win_inc = 1
# cBOSS
if self.randomised_ensemble:
possible_parameters = self._unique_parameters(max_window, win_inc)
num_classifiers = 0
start_time = time.time()
train_time = 0
subsample_size = int(self.n_instances * 0.7)
lowest_acc = 0
lowest_acc_idx = 0
rng = check_random_state(self.random_state)
if self.time_limit > 0:
self.n_parameter_samples = 0
while (train_time < self.time_limit or num_classifiers <
self.n_parameter_samples) and len(possible_parameters) > 0:
parameters = possible_parameters.pop(
rng.randint(0, len(possible_parameters)))
subsample = rng.choice(self.n_instances, size=subsample_size,
replace=False)
X_subsample = X.iloc[subsample, :]
y_subsample = y[subsample]
boss = BOSSIndividual(*parameters,
alphabet_size=self.alphabet_size,
save_words=False,
random_state=self.random_state)
boss.fit(X_subsample, y_subsample)
boss._clean()
boss.accuracy = self._individual_train_acc(boss, y_subsample,
subsample_size,
lowest_acc)
weight = math.pow(boss.accuracy, 4)
if num_classifiers < self.max_ensemble_size:
if boss.accuracy < lowest_acc:
lowest_acc = boss.accuracy
lowest_acc_idx = num_classifiers
self.weights.append(weight)
self.classifiers.append(boss)
elif boss.accuracy > lowest_acc:
self.weights[lowest_acc_idx] = weight
self.classifiers[lowest_acc_idx] = boss
lowest_acc, lowest_acc_idx = self._worst_ensemble_acc()
num_classifiers += 1
train_time = time.time() - start_time
# BOSS
else:
max_acc = -1
min_max_acc = -1
for i, normalise in enumerate(self.norm_options):
for win_size in range(self.min_window, max_window + 1,
win_inc):
boss = BOSSIndividual(win_size, self.word_lengths[0],
normalise, self.alphabet_size,
save_words=True,
random_state=self.random_state)
boss.fit(X, y)
best_classifier_for_win_size = boss
best_acc_for_win_size = -1
# the used work length may be shorter
best_word_len = boss.transformer.word_length
for n, word_len in enumerate(self.word_lengths):
if n > 0:
boss = boss._shorten_bags(word_len)
boss.accuracy = self._individual_train_acc(
boss, y, self.n_instances, best_acc_for_win_size)
# print(win_size, boss.accuracy)
if boss.accuracy >= best_acc_for_win_size:
best_acc_for_win_size = boss.accuracy
best_classifier_for_win_size = boss
best_word_len = word_len
if self._include_in_ensemble(best_acc_for_win_size,
max_acc,
min_max_acc,
len(self.classifiers)):
best_classifier_for_win_size._clean()
best_classifier_for_win_size._set_word_len(
best_word_len)
self.classifiers.append(best_classifier_for_win_size)
# print("appending", best_acc_for_win_size, win_size)
if best_acc_for_win_size > max_acc:
max_acc = best_acc_for_win_size
self.classifiers = list(compress(
self.classifiers, [
classifier.accuracy >= max_acc *
self.threshold for c, classifier in
enumerate(self.classifiers)]))
min_max_acc, min_acc_ind = \
self._worst_ensemble_acc()
if len(self.classifiers) > self.max_ensemble_size:
if min_acc_ind > -1:
del self.classifiers[min_acc_ind]
min_max_acc, min_acc_ind = \
self._worst_ensemble_acc()
self.weights = [1 for n in range(len(self.classifiers))]
self.n_estimators = len(self.classifiers)
self.weight_sum = np.sum(self.weights)
self._is_fitted = True
return self
def fit(self, X, y):
"""Build a WEASEL classifiers from the training set (X, y),
Parameters
----------
X : nested pandas DataFrame of shape [n_instances, 1]
Nested dataframe with univariate time-series in cells.
y : array-like, shape = [n_instances] The class labels.
Returns
-------
self : object
"""
X, y = check_X_y(X, y, enforce_univariate=True)
y = y.values if isinstance(y, pd.Series) else y
# Window length parameter space dependent on series length
self.n_instances, self.series_length = X.shape[0], len(X.iloc[0, 0])
self.max_window = min(self.series_length, self.max_window)
self.window_sizes = list(
range(self.min_window, self.max_window, self.win_inc))
max_acc = -1
self.highest_bit = (math.ceil(math.log2(self.max_window)) + 1)
final_bag_vec = None
for norm in self.norm_options:
# transformers = []
for w, word_length in enumerate(self.word_lengths):
all_words = [dict() for x in range(len(X))]
transformers = []
for i, window_size in enumerate(self.window_sizes):
# if w == 0: # only compute once, otherwise shorten
transformer = SFA(word_length=np.max(word_length),
alphabet_size=self.alphabet_size,
window_size=window_size,
norm=norm,
anova=self.anova,
binning_method=self.binning_strategy,
bigrams=self.bigrams,
remove_repeat_words=False,
lower_bounding=False,
save_words=False)
sfa_words = transformer.fit_transform(X, y)
transformers.append(transformer)
# use the shortening of words trick
# sfa_words = transformers[i]._shorten_bags(word_length)
# TODO refactor? dicts not really needed here ...
bag = sfa_words.iloc[:, 0]
# chi-squared test to keep only relevent features
# bag_vec = DictVectorizer(sparse=False).fit_transform(bag)
# chi2_statistics, p = chi2(bag_vec, y)
# relevant_features = np.where(
# chi2_statistics >= self.chi2_threshold)[0]
# merging bag-of-patterns of different window_sizes
# to single bag-of-patterns with prefix indicating
# the used window-length
for j in range(len(bag)):
for (key, value) in bag[j].items():
# if key in relevant_features: # chi-squared test
# append the prefices to the words to
# distinguish between window-sizes
word = (key << self.highest_bit) | window_size
# X_all_words[j].append((word, value))
all_words[j][word] = value
# TODO use CountVectorizer instead on actual words ... ???
vectorizer = DictVectorizer(sparse=True)
bag_vec = vectorizer.fit_transform(all_words)
clf = LogisticRegression(max_iter=5000,
solver="liblinear",
dual=True,
penalty="l2",
random_state=self.random_state)
current_acc = cross_val_score(clf, bag_vec, y, cv=5).mean()
# clf = RandomForestClassifier(oob_score=True,
# n_estimators=1000,
# n_jobs=-1).fit(bag_vec, y)
# current_acc = clf.oob_score_
# print("Train acc:", norm, word_length, current_acc)
if current_acc > max_acc:
max_acc = current_acc
self.vectorizer = vectorizer
self.clf = clf
self.SFA_transformers = transformers
self.best_word_length = word_length
final_bag_vec = bag_vec
if max_acc == 1.0:
break # there can be no better model than 1.0
# # fit final model using all words
# for i, window_size in enumerate(self.window_sizes):
# self.SFA_transformers[i] = \
# SFA(word_length=np.max(self.word_lengths),
# alphabet_size=self.alphabet_size,
# window_size=window_size,
# norm=norm,
# anova=self.anova,
# binning_method=self.binning_strategy,
# bigrams=self.bigrams,
# remove_repeat_words=False,
# lower_bounding=False,
# save_words=False)
# self.SFA_transformers[i].fit_transform(X, y)
self.clf.fit(final_bag_vec, y)
self._is_fitted = True
return self