def compute_distances(self, x1, x2=None):
"""
The method
- extracts normalized continuous attributes and then uses `row_norms`
and `safe_sparse_do`t to compute the distance as x^2 - 2xy - y^2
(the trick from sklearn);
- calls a function in Cython that recomputes the distances between pairs
of rows that yielded nan
- calls a function in Cython that adds the contributions of discrete
columns
"""
callbacks = StepwiseCallbacks(self.callback, [20, 10, 50, 5, 15])
if self.continuous.any():
data1, data2 = self.continuous_columns(x1, x2, self.means,
np.sqrt(2 * self.vars))
# adapted from sklearn.metric.euclidean_distances
xx = row_norms(data1, squared=True)[:, np.newaxis]
if x2 is not None:
yy = row_norms(data2, squared=True)[np.newaxis, :]
else:
yy = xx.T
distances = _safe_sparse_dot(data1,
data2.T,
dense_output=True,
callback=callbacks.next())
distances *= -2
distances += xx
distances += yy
with np.errstate(invalid="ignore"): # Nans are fixed below
np.maximum(distances, 0, out=distances)
if x2 is None:
distances.flat[::distances.shape[0] + 1] = 0.0
fixer = _distance.fix_euclidean_rows_normalized if self.normalize \
else _distance.fix_euclidean_rows
fixer(distances, data1, data2, self.means, self.vars,
self.dist_missing2_cont, x2 is not None, callbacks.next())
else:
distances = np.zeros(
(x1.shape[0], (x2 if x2 is not None else x1).shape[0]))
if np.any(self.discrete):
data1, data2 = self.discrete_columns(x1, x2)
_distance.euclidean_rows_discrete(distances, data1, data2,
self.dist_missing_disc,
self.dist_missing2_disc, x2
is not None, callbacks.next())
if x2 is None:
_distance.lower_to_symmetric(distances, callbacks.next())
return _interruptible_sqrt(distances, callback=callbacks.next())
def compute_distances(self, x1, x2=None):
"""
The method
- extracts normalized continuous attributes and then uses `row_norms`
and `safe_sparse_do`t to compute the distance as x^2 - 2xy - y^2
(the trick from sklearn);
- calls a function in Cython that recomputes the distances between pairs
of rows that yielded nan
- calls a function in Cython that adds the contributions of discrete
columns
"""
if self.continuous.any():
data1, data2 = self.continuous_columns(
x1, x2, self.means, np.sqrt(2 * self.vars))
# adapted from sklearn.metric.euclidean_distances
xx = row_norms(data1, squared=True)[:, np.newaxis]
if x2 is not None:
yy = row_norms(data2, squared=True)[np.newaxis, :]
else:
yy = xx.T
distances = safe_sparse_dot(data1, data2.T, dense_output=True)
distances *= -2
distances += xx
distances += yy
with np.errstate(invalid="ignore"): # Nans are fixed below
np.maximum(distances, 0, out=distances)
if x2 is None:
distances.flat[::distances.shape[0] + 1] = 0.0
fixer = _distance.fix_euclidean_rows_normalized if self.normalize \
else _distance.fix_euclidean_rows
fixer(distances, data1, data2,
self.means, self.vars, self.dist_missing2_cont,
x2 is not None)
else:
distances = np.zeros((x1.shape[0],
(x2 if x2 is not None else x1).shape[0]))
if np.any(self.discrete):
data1, data2 = self.discrete_columns(x1, x2)
_distance.euclidean_rows_discrete(
distances, data1, data2, self.dist_missing_disc,
self.dist_missing2_disc, x2 is not None)
if x2 is None:
_distance.lower_to_symmetric(distances)
return np.sqrt(distances)
def compute_distances(self, x1, x2):
"""
The method
- extracts normalized continuous attributes and computes distances
ignoring the possibility of nans
- recomputes the distances between pairs of rows that yielded nans
- adds the contributions of discrete columns using the same function as
the Euclidean distance
"""
callbacks = StepwiseCallbacks(self.callback, [5, 5, 60, 30])
if self.continuous.any():
data1, data2 = self.continuous_columns(x1, x2, self.medians,
2 * self.mads)
distances = _distance.manhattan_rows_cont(data1, data2, x2
is not None,
callbacks.next())
if self.normalize:
_distance.fix_manhattan_rows_normalized(
distances, data1, data2, x2 is not None, callbacks.next())
else:
_distance.fix_manhattan_rows(distances, data1, data2,
self.medians, self.mads,
self.dist_missing2_cont, x2
is not None, callbacks.next())
else:
distances = np.zeros(
(x1.shape[0], (x2 if x2 is not None else x1).shape[0]))
if np.any(self.discrete):
data1, data2 = self.discrete_columns(x1, x2)
# For discrete attributes, Euclidean is same as Manhattan
_distance.euclidean_rows_discrete(distances, data1, data2,
self.dist_missing_disc,
self.dist_missing2_disc, x2
is not None, callbacks.next())
if x2 is None:
_distance.lower_to_symmetric(distances, callbacks.next())
return distances
def compute_distances(self, x1, x2):
"""
The method
- extracts normalized continuous attributes and computes distances
ignoring the possibility of nans
- recomputes the distances between pairs of rows that yielded nans
- adds the contributions of discrete columns using the same function as
the Euclidean distance
"""
if self.continuous.any():
data1, data2 = self.continuous_columns(
x1, x2, self.medians, 2 * self.mads)
distances = _distance.manhattan_rows_cont(
data1, data2, x2 is not None)
if self.normalize:
_distance.fix_manhattan_rows_normalized(
distances, data1, data2, x2 is not None)
else:
_distance.fix_manhattan_rows(
distances, data1, data2,
self.medians, self.mads, self.dist_missing2_cont,
x2 is not None)
else:
distances = np.zeros((x1.shape[0],
(x2 if x2 is not None else x1).shape[0]))
if np.any(self.discrete):
data1, data2 = self.discrete_columns(x1, x2)
# For discrete attributes, Euclidean is same as Manhattan
_distance.euclidean_rows_discrete(
distances, data1, data2, self.dist_missing_disc,
self.dist_missing2_disc, x2 is not None)
if x2 is None:
_distance.lower_to_symmetric(distances)
return distances
