Source Code for Module parsimony.utils.classif_label

 1  # -*- coding: utf-8 -*- 
 2  """ 
 3  Created on Mon Mar 10 13:52:23 2014 
 4   
 5  Copyright (c) 2013-2014, CEA/DSV/I2BM/Neurospin. All rights reserved. 
 6   
 7  @author:  Edouard Duchesnay 
 8  @email:   edouard.duchesnay@cea.fr 
 9  @license: BSD 3-clause. 
10  """ 
11  import numpy as np 
12   
13   
14 -def class_weight_to_sample_weight(class_weight, y): 
15      """Estimate class weights for unbalanced datasets. 
16   
17      Parameters 
18      ---------- 
19      class_weight : dict, 'auto' or None 
20          If 'auto', class weights will be given inverse proportional 
21          to the frequency of the class in the data. sample_weight will sum 
22          to n_sample. 
23          If a dictionary is given, keys are classes and values 
24          are corresponding class weights. With two classes in {1, 0}, 
25          class_weight = {0:0.5, 1:0.5} is equivalent to class_weight == "auto" 
26          If None is given, the class weights will be uniform sample_weight==1. 
27   
28      y : array-like, shape (n_samples,) 
29          Array of original class labels per sample; 
30   
31      Returns 
32      ------- 
33      weight_vect : ndarray, shape (n_samples,) 
34          Array with weight_vect[i] the weight for i-th sample 
35   
36      Example 
37      ------- 
38      >>> y = [1, 1, 1, 0, 0, 2] 
39      >>> w = class_weight_to_sample_weight("auto", y) 
40      >>> print w.sum() == len(y) 
41      True 
42      >>> print ["%i:%.2f" % (l, np.sum(w[y==l])) for l in np.unique(y)] 
43      ['0:2.00', '1:2.00', '2:2.00'] 
44      >>> y = [1, 1, 1, 0, 0, 2] 
45      >>> w2 = class_weight_to_sample_weight({0:1./3, 1:1./3, 2:1./3}, y) 
46      >>> np.all(w2 == w) 
47      True 
48      """ 
49      if class_weight is None or len(class_weight) == 0: 
50          # uniform class weights 
51          return np.ones(y.shape, dtype=np.float64) 
52      # wik = n / nk * pk 
53      # pk: desire prior of class k (sum pk == 1) 
54      y = np.asarray(y) 
55      classes = np.unique(y) 
56      nk = np.bincount(y.astype(int).ravel()) 
57      n = float(y.shape[0]) 
58      if class_weight == 'auto': 
59          pk = 1. / classes.shape[0] 
60      else: 
61          if not isinstance(class_weight, dict): 
62              raise ValueError("class_weight must be dict, 'auto', or None," 
63                               " got: %r" % class_weight) 
64          pk = np.array([class_weight[k] for k in classes]) 
65      wk = n / nk * pk 
66      sample_weight = wk[np.searchsorted(classes, y)] 
67      return sample_weight 
68   
69   
70 -def check_labels(y): 
71      """ensure binary classification with 0, 1 labels""" 
72      nlevels = 2 
73      classes = np.unique(y) 
74      if len(classes) > nlevels: 
75          raise ValueError("Multinomial classification with more " \ 
76                          "than %i labels is not possible" % nlevels) 
77      classes_recoded = np.arange(len(classes)) 
78      if np.all(classes_recoded == classes): 
79          return y 
80      # Ensure labels are 0, 1 
81      y_recoded = np.zeros(y.shape, dtype=np.float64) 
82      for i in xrange(len(classes)): 
83          y_recoded[y == classes[i]] = classes_recoded[i] 
84      return y_recoded 
85