564 lines
20 KiB
Cython
564 lines
20 KiB
Cython
# Authors: Mathieu Blondel
|
|
# Olivier Grisel
|
|
# Peter Prettenhofer
|
|
# Lars Buitinck
|
|
# Giorgio Patrini
|
|
#
|
|
# License: BSD 3 clause
|
|
|
|
#!python
|
|
|
|
from libc.math cimport fabs, sqrt, pow
|
|
cimport numpy as np
|
|
import numpy as np
|
|
cimport cython
|
|
from cython cimport floating
|
|
from numpy.math cimport isnan
|
|
|
|
np.import_array()
|
|
|
|
ctypedef fused integral:
|
|
int
|
|
long long
|
|
|
|
ctypedef np.float64_t DOUBLE
|
|
|
|
|
|
def csr_row_norms(X):
|
|
"""L2 norm of each row in CSR matrix X."""
|
|
if X.dtype not in [np.float32, np.float64]:
|
|
X = X.astype(np.float64)
|
|
return _csr_row_norms(X.data, X.shape, X.indices, X.indptr)
|
|
|
|
|
|
def _csr_row_norms(np.ndarray[floating, ndim=1, mode="c"] X_data,
|
|
shape,
|
|
np.ndarray[integral, ndim=1, mode="c"] X_indices,
|
|
np.ndarray[integral, ndim=1, mode="c"] X_indptr):
|
|
cdef:
|
|
unsigned long long n_samples = shape[0]
|
|
unsigned long long i
|
|
integral j
|
|
double sum_
|
|
|
|
norms = np.empty(n_samples, dtype=X_data.dtype)
|
|
cdef floating[::1] norms_view = norms
|
|
|
|
for i in range(n_samples):
|
|
sum_ = 0.0
|
|
for j in range(X_indptr[i], X_indptr[i + 1]):
|
|
sum_ += X_data[j] * X_data[j]
|
|
norms_view[i] = sum_
|
|
|
|
return norms
|
|
|
|
|
|
def csr_mean_variance_axis0(X, weights=None, return_sum_weights=False):
|
|
"""Compute mean and variance along axis 0 on a CSR matrix
|
|
|
|
Uses a np.float64 accumulator.
|
|
|
|
Parameters
|
|
----------
|
|
X : CSR sparse matrix, shape (n_samples, n_features)
|
|
Input data.
|
|
|
|
weights : ndarray of shape (n_samples,), dtype=floating, default=None
|
|
If it is set to None samples will be equally weighted.
|
|
|
|
.. versionadded:: 0.24
|
|
|
|
return_sum_weights : bool, default=False
|
|
If True, returns the sum of weights seen for each feature.
|
|
|
|
.. versionadded:: 0.24
|
|
|
|
Returns
|
|
-------
|
|
means : float array with shape (n_features,)
|
|
Feature-wise means
|
|
|
|
variances : float array with shape (n_features,)
|
|
Feature-wise variances
|
|
|
|
sum_weights : ndarray of shape (n_features,), dtype=floating
|
|
Returned if return_sum_weights is True.
|
|
"""
|
|
if X.dtype not in [np.float32, np.float64]:
|
|
X = X.astype(np.float64)
|
|
|
|
if weights is None:
|
|
weights = np.ones(X.shape[0], dtype=X.dtype)
|
|
|
|
means, variances, sum_weights = _csr_mean_variance_axis0(
|
|
X.data, X.shape[0], X.shape[1], X.indices, X.indptr, weights)
|
|
|
|
if return_sum_weights:
|
|
return means, variances, sum_weights
|
|
return means, variances
|
|
|
|
|
|
def _csr_mean_variance_axis0(np.ndarray[floating, ndim=1, mode="c"] X_data,
|
|
unsigned long long n_samples,
|
|
unsigned long long n_features,
|
|
np.ndarray[integral, ndim=1] X_indices,
|
|
np.ndarray[integral, ndim=1] X_indptr,
|
|
np.ndarray[floating, ndim=1] weights):
|
|
# Implement the function here since variables using fused types
|
|
# cannot be declared directly and can only be passed as function arguments
|
|
cdef:
|
|
np.npy_intp i
|
|
unsigned long long row_ind
|
|
integral col_ind
|
|
np.float64_t diff
|
|
# means[j] contains the mean of feature j
|
|
np.ndarray[np.float64_t, ndim=1] means = np.zeros(n_features)
|
|
# variances[j] contains the variance of feature j
|
|
np.ndarray[np.float64_t, ndim=1] variances = np.zeros(n_features)
|
|
|
|
np.ndarray[np.float64_t, ndim=1] sum_weights = np.full(
|
|
fill_value=np.sum(weights, dtype=np.float64), shape=n_features)
|
|
np.ndarray[np.float64_t, ndim=1] sum_weights_nz = np.zeros(
|
|
shape=n_features)
|
|
np.ndarray[np.float64_t, ndim=1] correction = np.zeros(
|
|
shape=n_features)
|
|
|
|
np.ndarray[np.uint64_t, ndim=1] counts = np.full(
|
|
fill_value=weights.shape[0], shape=n_features, dtype=np.uint64)
|
|
np.ndarray[np.uint64_t, ndim=1] counts_nz = np.zeros(
|
|
shape=n_features, dtype=np.uint64)
|
|
|
|
for row_ind in range(len(X_indptr) - 1):
|
|
for i in range(X_indptr[row_ind], X_indptr[row_ind + 1]):
|
|
col_ind = X_indices[i]
|
|
if not isnan(X_data[i]):
|
|
means[col_ind] += <np.float64_t>(X_data[i]) * weights[row_ind]
|
|
# sum of weights where X[:, col_ind] is non-zero
|
|
sum_weights_nz[col_ind] += weights[row_ind]
|
|
# number of non-zero elements of X[:, col_ind]
|
|
counts_nz[col_ind] += 1
|
|
else:
|
|
# sum of weights where X[:, col_ind] is not nan
|
|
sum_weights[col_ind] -= weights[row_ind]
|
|
# number of non nan elements of X[:, col_ind]
|
|
counts[col_ind] -= 1
|
|
|
|
for i in range(n_features):
|
|
means[i] /= sum_weights[i]
|
|
|
|
for row_ind in range(len(X_indptr) - 1):
|
|
for i in range(X_indptr[row_ind], X_indptr[row_ind + 1]):
|
|
col_ind = X_indices[i]
|
|
if not isnan(X_data[i]):
|
|
diff = X_data[i] - means[col_ind]
|
|
# correction term of the corrected 2 pass algorithm.
|
|
# See "Algorithms for computing the sample variance: analysis
|
|
# and recommendations", by Chan, Golub, and LeVeque.
|
|
correction[col_ind] += diff * weights[row_ind]
|
|
variances[col_ind] += diff * diff * weights[row_ind]
|
|
|
|
for i in range(n_features):
|
|
if counts[i] != counts_nz[i]:
|
|
correction[i] -= (sum_weights[i] - sum_weights_nz[i]) * means[i]
|
|
correction[i] = correction[i]**2 / sum_weights[i]
|
|
if counts[i] != counts_nz[i]:
|
|
# only compute it when it's guaranteed to be non-zero to avoid
|
|
# catastrophic cancellation.
|
|
variances[i] += (sum_weights[i] - sum_weights_nz[i]) * means[i]**2
|
|
variances[i] = (variances[i] - correction[i]) / sum_weights[i]
|
|
|
|
if floating is float:
|
|
return (np.array(means, dtype=np.float32),
|
|
np.array(variances, dtype=np.float32),
|
|
np.array(sum_weights, dtype=np.float32))
|
|
else:
|
|
return means, variances, sum_weights
|
|
|
|
|
|
def csc_mean_variance_axis0(X, weights=None, return_sum_weights=False):
|
|
"""Compute mean and variance along axis 0 on a CSC matrix
|
|
|
|
Uses a np.float64 accumulator.
|
|
|
|
Parameters
|
|
----------
|
|
X : CSC sparse matrix, shape (n_samples, n_features)
|
|
Input data.
|
|
|
|
weights : ndarray of shape (n_samples,), dtype=floating, default=None
|
|
If it is set to None samples will be equally weighted.
|
|
|
|
.. versionadded:: 0.24
|
|
|
|
return_sum_weights : bool, default=False
|
|
If True, returns the sum of weights seen for each feature.
|
|
|
|
.. versionadded:: 0.24
|
|
|
|
Returns
|
|
-------
|
|
means : float array with shape (n_features,)
|
|
Feature-wise means
|
|
|
|
variances : float array with shape (n_features,)
|
|
Feature-wise variances
|
|
|
|
sum_weights : ndarray of shape (n_features,), dtype=floating
|
|
Returned if return_sum_weights is True.
|
|
"""
|
|
if X.dtype not in [np.float32, np.float64]:
|
|
X = X.astype(np.float64)
|
|
|
|
if weights is None:
|
|
weights = np.ones(X.shape[0], dtype=X.dtype)
|
|
|
|
means, variances, sum_weights = _csc_mean_variance_axis0(
|
|
X.data, X.shape[0], X.shape[1], X.indices, X.indptr, weights)
|
|
|
|
if return_sum_weights:
|
|
return means, variances, sum_weights
|
|
return means, variances
|
|
|
|
|
|
def _csc_mean_variance_axis0(np.ndarray[floating, ndim=1, mode="c"] X_data,
|
|
unsigned long long n_samples,
|
|
unsigned long long n_features,
|
|
np.ndarray[integral, ndim=1] X_indices,
|
|
np.ndarray[integral, ndim=1] X_indptr,
|
|
np.ndarray[floating, ndim=1] weights):
|
|
# Implement the function here since variables using fused types
|
|
# cannot be declared directly and can only be passed as function arguments
|
|
cdef:
|
|
np.npy_intp i
|
|
unsigned long long col_ind
|
|
integral row_ind
|
|
np.float64_t diff
|
|
# means[j] contains the mean of feature j
|
|
np.ndarray[np.float64_t, ndim=1] means = np.zeros(n_features)
|
|
# variances[j] contains the variance of feature j
|
|
np.ndarray[np.float64_t, ndim=1] variances = np.zeros(n_features)
|
|
|
|
np.ndarray[np.float64_t, ndim=1] sum_weights = np.full(
|
|
fill_value=np.sum(weights, dtype=np.float64), shape=n_features)
|
|
np.ndarray[np.float64_t, ndim=1] sum_weights_nz = np.zeros(
|
|
shape=n_features)
|
|
np.ndarray[np.float64_t, ndim=1] correction = np.zeros(
|
|
shape=n_features)
|
|
|
|
np.ndarray[np.uint64_t, ndim=1] counts = np.full(
|
|
fill_value=weights.shape[0], shape=n_features, dtype=np.uint64)
|
|
np.ndarray[np.uint64_t, ndim=1] counts_nz = np.zeros(
|
|
shape=n_features, dtype=np.uint64)
|
|
|
|
for col_ind in range(n_features):
|
|
for i in range(X_indptr[col_ind], X_indptr[col_ind + 1]):
|
|
row_ind = X_indices[i]
|
|
if not isnan(X_data[i]):
|
|
means[col_ind] += <np.float64_t>(X_data[i]) * weights[row_ind]
|
|
# sum of weights where X[:, col_ind] is non-zero
|
|
sum_weights_nz[col_ind] += weights[row_ind]
|
|
# number of non-zero elements of X[:, col_ind]
|
|
counts_nz[col_ind] += 1
|
|
else:
|
|
# sum of weights where X[:, col_ind] is not nan
|
|
sum_weights[col_ind] -= weights[row_ind]
|
|
# number of non nan elements of X[:, col_ind]
|
|
counts[col_ind] -= 1
|
|
|
|
for i in range(n_features):
|
|
means[i] /= sum_weights[i]
|
|
|
|
for col_ind in range(n_features):
|
|
for i in range(X_indptr[col_ind], X_indptr[col_ind + 1]):
|
|
row_ind = X_indices[i]
|
|
if not isnan(X_data[i]):
|
|
diff = X_data[i] - means[col_ind]
|
|
# correction term of the corrected 2 pass algorithm.
|
|
# See "Algorithms for computing the sample variance: analysis
|
|
# and recommendations", by Chan, Golub, and LeVeque.
|
|
correction[col_ind] += diff * weights[row_ind]
|
|
variances[col_ind] += diff * diff * weights[row_ind]
|
|
|
|
for i in range(n_features):
|
|
if counts[i] != counts_nz[i]:
|
|
correction[i] -= (sum_weights[i] - sum_weights_nz[i]) * means[i]
|
|
correction[i] = correction[i]**2 / sum_weights[i]
|
|
if counts[i] != counts_nz[i]:
|
|
# only compute it when it's guaranteed to be non-zero to avoid
|
|
# catastrophic cancellation.
|
|
variances[i] += (sum_weights[i] - sum_weights_nz[i]) * means[i]**2
|
|
variances[i] = (variances[i] - correction[i]) / sum_weights[i]
|
|
|
|
if floating is float:
|
|
return (np.array(means, dtype=np.float32),
|
|
np.array(variances, dtype=np.float32),
|
|
np.array(sum_weights, dtype=np.float32))
|
|
else:
|
|
return means, variances, sum_weights
|
|
|
|
|
|
def incr_mean_variance_axis0(X, last_mean, last_var, last_n, weights=None):
|
|
"""Compute mean and variance along axis 0 on a CSR or CSC matrix.
|
|
|
|
last_mean, last_var are the statistics computed at the last step by this
|
|
function. Both must be initialized to 0.0. last_n is the
|
|
number of samples encountered until now and is initialized at 0.
|
|
|
|
Parameters
|
|
----------
|
|
X : CSR or CSC sparse matrix, shape (n_samples, n_features)
|
|
Input data.
|
|
|
|
last_mean : float array with shape (n_features,)
|
|
Array of feature-wise means to update with the new data X.
|
|
|
|
last_var : float array with shape (n_features,)
|
|
Array of feature-wise var to update with the new data X.
|
|
|
|
last_n : float array with shape (n_features,)
|
|
Sum of the weights seen so far (if weights are all set to 1
|
|
this will be the same as number of samples seen so far, before X).
|
|
|
|
weights : float array with shape (n_samples,) or None. If it is set
|
|
to None samples will be equally weighted.
|
|
|
|
Returns
|
|
-------
|
|
updated_mean : float array with shape (n_features,)
|
|
Feature-wise means
|
|
|
|
updated_variance : float array with shape (n_features,)
|
|
Feature-wise variances
|
|
|
|
updated_n : int array with shape (n_features,)
|
|
Updated number of samples seen
|
|
|
|
Notes
|
|
-----
|
|
NaNs are ignored during the computation.
|
|
|
|
References
|
|
----------
|
|
T. Chan, G. Golub, R. LeVeque. Algorithms for computing the sample
|
|
variance: recommendations, The American Statistician, Vol. 37, No. 3,
|
|
pp. 242-247
|
|
|
|
Also, see the non-sparse implementation of this in
|
|
`utils.extmath._batch_mean_variance_update`.
|
|
|
|
"""
|
|
if X.dtype not in [np.float32, np.float64]:
|
|
X = X.astype(np.float64)
|
|
X_dtype = X.dtype
|
|
if weights is None:
|
|
weights = np.ones(X.shape[0], dtype=X_dtype)
|
|
elif weights.dtype not in [np.float32, np.float64]:
|
|
weights = weights.astype(np.float64, copy=False)
|
|
if last_n.dtype not in [np.float32, np.float64]:
|
|
last_n = last_n.astype(np.float64, copy=False)
|
|
|
|
return _incr_mean_variance_axis0(X.data,
|
|
np.sum(weights),
|
|
X.shape[1],
|
|
X.indices,
|
|
X.indptr,
|
|
X.format,
|
|
last_mean.astype(X_dtype, copy=False),
|
|
last_var.astype(X_dtype, copy=False),
|
|
last_n.astype(X_dtype, copy=False),
|
|
weights.astype(X_dtype, copy=False))
|
|
|
|
|
|
def _incr_mean_variance_axis0(np.ndarray[floating, ndim=1] X_data,
|
|
floating n_samples,
|
|
unsigned long long n_features,
|
|
np.ndarray[int, ndim=1] X_indices,
|
|
# X_indptr might be either in32 or int64
|
|
np.ndarray[integral, ndim=1] X_indptr,
|
|
str X_format,
|
|
np.ndarray[floating, ndim=1] last_mean,
|
|
np.ndarray[floating, ndim=1] last_var,
|
|
np.ndarray[floating, ndim=1] last_n,
|
|
# previous sum of the weights (ie float)
|
|
np.ndarray[floating, ndim=1] weights):
|
|
# Implement the function here since variables using fused types
|
|
# cannot be declared directly and can only be passed as function arguments
|
|
cdef:
|
|
np.npy_intp i
|
|
|
|
# last = stats until now
|
|
# new = the current increment
|
|
# updated = the aggregated stats
|
|
# when arrays, they are indexed by i per-feature
|
|
cdef:
|
|
np.ndarray[floating, ndim=1] new_mean
|
|
np.ndarray[floating, ndim=1] new_var
|
|
np.ndarray[floating, ndim=1] updated_mean
|
|
np.ndarray[floating, ndim=1] updated_var
|
|
|
|
if floating is float:
|
|
dtype = np.float32
|
|
else:
|
|
dtype = np.float64
|
|
|
|
new_mean = np.zeros(n_features, dtype=dtype)
|
|
new_var = np.zeros_like(new_mean, dtype=dtype)
|
|
updated_mean = np.zeros_like(new_mean, dtype=dtype)
|
|
updated_var = np.zeros_like(new_mean, dtype=dtype)
|
|
|
|
cdef:
|
|
np.ndarray[floating, ndim=1] new_n
|
|
np.ndarray[floating, ndim=1] updated_n
|
|
np.ndarray[floating, ndim=1] last_over_new_n
|
|
|
|
# Obtain new stats first
|
|
updated_n = np.zeros(shape=n_features, dtype=dtype)
|
|
last_over_new_n = np.zeros_like(updated_n, dtype=dtype)
|
|
|
|
# X can be a CSR or CSC matrix
|
|
if X_format == 'csr':
|
|
new_mean, new_var, new_n = _csr_mean_variance_axis0(
|
|
X_data, n_samples, n_features, X_indices, X_indptr, weights)
|
|
else: # X_format == 'csc'
|
|
new_mean, new_var, new_n = _csc_mean_variance_axis0(
|
|
X_data, n_samples, n_features, X_indices, X_indptr, weights)
|
|
|
|
# First pass
|
|
cdef bint is_first_pass = True
|
|
for i in range(n_features):
|
|
if last_n[i] > 0:
|
|
is_first_pass = False
|
|
break
|
|
|
|
if is_first_pass:
|
|
return new_mean, new_var, new_n
|
|
|
|
for i in range(n_features):
|
|
updated_n[i] = last_n[i] + new_n[i]
|
|
|
|
# Next passes
|
|
for i in range(n_features):
|
|
if new_n[i] > 0:
|
|
last_over_new_n[i] = dtype(last_n[i]) / dtype(new_n[i])
|
|
# Unnormalized stats
|
|
last_mean[i] *= last_n[i]
|
|
last_var[i] *= last_n[i]
|
|
new_mean[i] *= new_n[i]
|
|
new_var[i] *= new_n[i]
|
|
# Update stats
|
|
updated_var[i] = (
|
|
last_var[i] + new_var[i] +
|
|
last_over_new_n[i] / updated_n[i] *
|
|
(last_mean[i] / last_over_new_n[i] - new_mean[i])**2
|
|
)
|
|
updated_mean[i] = (last_mean[i] + new_mean[i]) / updated_n[i]
|
|
updated_var[i] /= updated_n[i]
|
|
else:
|
|
updated_var[i] = last_var[i]
|
|
updated_mean[i] = last_mean[i]
|
|
updated_n[i] = last_n[i]
|
|
|
|
return updated_mean, updated_var, updated_n
|
|
|
|
|
|
def inplace_csr_row_normalize_l1(X):
|
|
"""Inplace row normalize using the l1 norm"""
|
|
_inplace_csr_row_normalize_l1(X.data, X.shape, X.indices, X.indptr)
|
|
|
|
|
|
def _inplace_csr_row_normalize_l1(np.ndarray[floating, ndim=1] X_data,
|
|
shape,
|
|
np.ndarray[integral, ndim=1] X_indices,
|
|
np.ndarray[integral, ndim=1] X_indptr):
|
|
cdef unsigned long long n_samples = shape[0]
|
|
cdef unsigned long long n_features = shape[1]
|
|
|
|
# the column indices for row i are stored in:
|
|
# indices[indptr[i]:indices[i+1]]
|
|
# and their corresponding values are stored in:
|
|
# data[indptr[i]:indptr[i+1]]
|
|
cdef np.npy_intp i, j
|
|
cdef double sum_
|
|
|
|
for i in range(n_samples):
|
|
sum_ = 0.0
|
|
|
|
for j in range(X_indptr[i], X_indptr[i + 1]):
|
|
sum_ += fabs(X_data[j])
|
|
|
|
if sum_ == 0.0:
|
|
# do not normalize empty rows (can happen if CSR is not pruned
|
|
# correctly)
|
|
continue
|
|
|
|
for j in range(X_indptr[i], X_indptr[i + 1]):
|
|
X_data[j] /= sum_
|
|
|
|
|
|
def inplace_csr_row_normalize_l2(X):
|
|
"""Inplace row normalize using the l2 norm"""
|
|
_inplace_csr_row_normalize_l2(X.data, X.shape, X.indices, X.indptr)
|
|
|
|
|
|
def _inplace_csr_row_normalize_l2(np.ndarray[floating, ndim=1] X_data,
|
|
shape,
|
|
np.ndarray[integral, ndim=1] X_indices,
|
|
np.ndarray[integral, ndim=1] X_indptr):
|
|
cdef integral n_samples = shape[0]
|
|
cdef integral n_features = shape[1]
|
|
|
|
cdef np.npy_intp i, j
|
|
cdef double sum_
|
|
|
|
for i in range(n_samples):
|
|
sum_ = 0.0
|
|
|
|
for j in range(X_indptr[i], X_indptr[i + 1]):
|
|
sum_ += (X_data[j] * X_data[j])
|
|
|
|
if sum_ == 0.0:
|
|
# do not normalize empty rows (can happen if CSR is not pruned
|
|
# correctly)
|
|
continue
|
|
|
|
sum_ = sqrt(sum_)
|
|
|
|
for j in range(X_indptr[i], X_indptr[i + 1]):
|
|
X_data[j] /= sum_
|
|
|
|
|
|
def assign_rows_csr(X,
|
|
np.ndarray[np.npy_intp, ndim=1] X_rows,
|
|
np.ndarray[np.npy_intp, ndim=1] out_rows,
|
|
np.ndarray[floating, ndim=2, mode="c"] out):
|
|
"""Densify selected rows of a CSR matrix into a preallocated array.
|
|
|
|
Like out[out_rows] = X[X_rows].toarray() but without copying.
|
|
No-copy supported for both dtype=np.float32 and dtype=np.float64.
|
|
|
|
Parameters
|
|
----------
|
|
X : scipy.sparse.csr_matrix, shape=(n_samples, n_features)
|
|
X_rows : array, dtype=np.intp, shape=n_rows
|
|
out_rows : array, dtype=np.intp, shape=n_rows
|
|
out : array, shape=(arbitrary, n_features)
|
|
"""
|
|
cdef:
|
|
# npy_intp (np.intp in Python) is what np.where returns,
|
|
# but int is what scipy.sparse uses.
|
|
int i, ind, j
|
|
np.npy_intp rX
|
|
np.ndarray[floating, ndim=1] data = X.data
|
|
np.ndarray[int, ndim=1] indices = X.indices, indptr = X.indptr
|
|
|
|
if X_rows.shape[0] != out_rows.shape[0]:
|
|
raise ValueError("cannot assign %d rows to %d"
|
|
% (X_rows.shape[0], out_rows.shape[0]))
|
|
|
|
out[out_rows] = 0.
|
|
for i in range(X_rows.shape[0]):
|
|
rX = X_rows[i]
|
|
for ind in range(indptr[rX], indptr[rX + 1]):
|
|
j = indices[ind]
|
|
out[out_rows[i], j] = data[ind]
|