77
88import sys
99import warnings
10+ import numbers
1011from abc import ABCMeta , abstractmethod
1112
1213import numpy as np
2021from ..utils .validation import check_random_state
2122from ..model_selection import check_cv
2223from ..utils .extmath import safe_sparse_dot
23- from ..utils .fixes import _joblib_parallel_args
24- from ..utils .validation import check_is_fitted
24+ from ..utils .fixes import _astype_copy_false , _joblib_parallel_args
25+ from ..utils .validation import check_is_fitted , _check_sample_weight
2526from ..utils .validation import column_or_1d
2627
2728from . import _cd_fast as cd_fast
2829
2930
31+ def _set_order (X , y , order = 'C' ):
32+ """Change the order of X and y if necessary.
33+
34+ Parameters
35+ ----------
36+ X : {array-like, sparse matrix} of shape (n_samples, n_features)
37+ Training data.
38+
39+ y : ndarray of shape (n_samples,)
40+ Target values.
41+
42+ order : {None, 'C', 'F'}
43+ If 'C', dense arrays are returned as C-ordered, sparse matrices in csr
44+ format. If 'F', dense arrays are return as F-ordered, sparse matrices
45+ in csc format.
46+
47+ Returns
48+ -------
49+ X : {array-like, sparse matrix} of shape (n_samples, n_features)
50+ Training data with guaranteed order.
51+
52+ y : ndarray of shape (n_samples,)
53+ Target values with guaranteed order.
54+ """
55+ if order not in None , 'C' , 'F' ]:
56+ raise ValueError ("Unknown value for order. Got {} instead of "
57+ "None, 'C' or 'F'." .format (order ))
58+ sparse_X = sparse .issparse (X )
59+ sparse_y = sparse .issparse (y )
60+ if order is not None :
61+ sparse_format = "csc" if order == "F" else "csr"
62+ if sparse_X :
63+ # As of scipy 1.1.0, new argument copy=False by default.
64+ # This is what we want.
65+ X = X .asformat (sparse_format , ** _astype_copy_false (X ))
66+ else :
67+ X = np .asarray (X , order = order )
68+ if sparse_y :
69+ y = y .asformat (sparse_format )
70+ else :
71+ y = np .asarray (y , order = order )
72+ return X , y
73+
74+
3075###############################################################################
3176# Paths functions
3277
@@ -661,7 +706,7 @@ def __init__(self, alpha=1.0, l1_ratio=0.5, fit_intercept=True,
661706 self .random_state = random_state
662707 self .selection = selection
663708
664- def fit (self , X , y , check_input = True ):
709+ def fit (self , X , y , sample_weight = None , check_input = True ):
665710 """Fit model with coordinate descent.
666711
667712 Parameters
@@ -673,6 +718,9 @@ def fit(self, X, y, check_input=True):
673718
674719 Target. Will be cast to X's dtype if necessary
675720
721+ sample_weight : float or array-like of shape (n_samples,), default=None
722+ Sample weight.
723+
676724 check_input : bool, default=True
677725 Allow to bypass several input checking.
678726 Don't use this parameter unless you know what you do.
@@ -709,18 +757,49 @@ def fit(self, X, y, check_input=True):
709757 y = check_array (y , order = 'F' , copy = False , dtype = X .dtype .type ,
710758 ensure_2d = False )
711759
712- # Ensure copying happens only once, don't do it again if done above
760+ n_samples , n_features = X .shape
761+ alpha = self .alpha
762+
763+ if isinstance (sample_weight , numbers .Number ):
764+ sample_weight = None
765+ if sample_weight is not None :
766+ if check_input :
767+ if sparse .issparse (X ):
768+ raise ValueError ("Sample weights do not (yet) support "
769+ "sparse matrices." )
770+ sample_weight = _check_sample_weight (sample_weight , X ,
771+ dtype = X .dtype )
772+ # simplify things by rescaling sw to sum up to n_samples
773+ # => np.average(x, weights=sw) = np.mean(sw * x)
774+ sample_weight *= (n_samples / np .sum (sample_weight ))
775+ # Objective function is:
776+ # 1/2 * np.average(squared error, weights=sw) + alpha * penalty
777+ # but coordinate descent minimizes:
778+ # 1/2 * sum(squared error) + alpha * penalty
779+ # enet_path therefore sets alpha = n_samples * alpha
780+ # With sw, enet_path should set alpha = sum(sw) * alpha
781+ # Therefore, we rescale alpha = sum(sw) / n_samples * alpha
782+ # Note: As we rescaled sample_weights to sum up to n_samples,
783+ # we don't need this
784+ # alpha *= np.sum(sample_weight) / n_samples
785+
786+ # Ensure copying happens only once, don't do it again if done above.
787+ # X and y will be rescaled if sample_weight is not None, order='F'
788+ # ensures that the returned X and y are still F-contiguous.
713789 should_copy = self .copy_X and not X_copied
714790 X , y , X_offset , y_offset , X_scale , precompute , Xy = \
715791 _pre_fit (X , y , None , self .precompute , self .normalize ,
716792 self .fit_intercept , copy = should_copy ,
717- check_input = check_input )
793+ check_input = check_input , sample_weight = sample_weight )
794+ # coordinate descent needs F-ordered arrays and _pre_fit might have
795+ # called _rescale_data
796+ if check_input or sample_weight is not None :
797+ X , y = _set_order (X , y , order = 'F' )
718798 if y .ndim == 1 :
719799 y = y [:, np .newaxis ]
720800 if Xy is not None and Xy .ndim == 1 :
721801 Xy = Xy [:, np .newaxis ]
722802
723- n_samples , n_features = X .shape
724803 n_targets = y .shape [1 ]
725804
726805 if self .selection not in 'cyclic' , 'random' ]:
@@ -745,7 +824,7 @@ def fit(self, X, y, check_input=True):
745824 _ , this_coef , this_dual_gap , this_iter = \
746825 self .path (X , y [:, k ],
747826 l1_ratio = self .l1_ratio , eps = None ,
748- n_alphas = None , alphas = [self . alpha ],
827+ n_alphas = None , alphas = [alpha ],
749828 precompute = precompute , Xy = this_Xy ,
750829 fit_intercept = False , normalize = False , copy_X = True ,
751830 verbose = False , tol = self .tol , positive = self .positive ,
@@ -1397,6 +1476,7 @@ def __init__(self, eps=1e-3, n_alphas=100, alphas=None, fit_intercept=True,
13971476 def _more_tags (self ):
13981477 return {'multioutput' : False }
13991478
1479+
14001480class ElasticNetCV (RegressorMixin , LinearModelCV ):
14011481 """Elastic Net model with iterative fitting along a regularization path.
14021482
0 commit comments