Source Code for Module parsimony.utils.resampling

  1  # -*- coding: utf-8 -*- 
  2  """ 
  3  Created on Mon Jun 16 10:00:06 2014 
  4   
  5  Copyright (c) 2013-2014, CEA/DSV/I2BM/Neurospin. All rights reserved. 
  6   
  7  @author:  Tommy Löfstedt 
  8  @email:   lofstedt.tommy@gmail.com 
  9  @license: BSD 3-clause. 
 10  """ 
 11  import numpy as np 
 12   
 13  __all__ = ["k_fold", "stratified_k_fold", 
 14             "bootstrap", "stratified_bootstrap"] 
 15   
 16   
 17 -def k_fold(n, K=7): 
 18      """K-fold cross validation iterator. 
 19   
 20      Returns indices for training and test sets. 
 21   
 22      Parameters 
 23      ---------- 
 24      n : Positive integer greater than one. The number of samples. 
 25   
 26      K : Positive integer greater than or equal to two. The number of folds. 
 27      """ 
 28      all_ids = set(range(n)) 
 29      for k in xrange(K): 
 30          test = range(k, n, K) 
 31          train = all_ids.difference(test) 
 32   
 33          yield list(train), test 
 34   
 35   
 36 -def stratified_k_fold(y, K=7): 
 37      """Stratified k-fold cross validation iterator. 
 38   
 39      Returns indices for training and test sets. 
 40   
 41      Parameters 
 42      ---------- 
 43      y : Numpy array with n > 1 elements. The class labels. These labels are 
 44              used to stratify the folds. 
 45   
 46      K : Positive integer greater than or equal to two. The number of folds. 
 47      """ 
 48      y = np.array(y) 
 49      n = np.prod(y.shape) 
 50      y = np.reshape(y, (n, 1)) 
 51   
 52      # Assign the class labels to different folds. 
 53      labels, y_inverse = np.unique(y, return_inverse=True) 
 54      count = np.bincount(y_inverse) 
 55      classes = -np.ones(n) 
 56      for i in xrange(count.shape[0]): 
 57          c = count[i] 
 58          v = np.mod(np.arange(c), K) 
 59   
 60          classes[y_inverse == i] = v 
 61   
 62      # Assign any leftovers to existing classes. 
 63      m = np.max(classes) 
 64      if m > K - 1: 
 65          ind = np.where(classes == m)[0] 
 66          for i in range(len(ind)): 
 67              classes[ind[i]] = i 
 68   
 69      # Loop over the stratified classes and yield the given train and test set. 
 70      all_ids = set(range(n)) 
 71      for k in range(K): 
 72          test = np.where(classes == k)[0].tolist() 
 73          train = list(all_ids.difference(test)) 
 74   
 75          yield train, test 
 76   
 77   
 78 -def bootstrap(n, B=100, seed=None): 
 79      """Bootstrap sample iterator. 
 80   
 81      Returns indices for a bootstrap training set. 
 82   
 83      Parameters 
 84      ---------- 
 85      n : Positive integer greater than one. The number of samples. 
 86   
 87      B : Positive integer greater than or equal to two. The number of bootstrap 
 88              samples to draw. 
 89   
 90      seed : Integer. A random seed to initialise the random number generator 
 91              with. Use in order to obtain deterministic results. The seed is not 
 92              used if the seed is None. 
 93      """ 
 94      if seed is not None: 
 95          np.random.seed(seed) 
 96   
 97      for b in xrange(B): 
 98          sample = np.random.randint(0, n, size=n).tolist() 
 99   
100          yield sample 
101   
102   
103 -def stratified_bootstrap(y, B=100, seed=None): 
104      """Stratified bootstrap sample iterator. 
105   
106      Returns indices for a bootstrap training set. 
107   
108      Parameters 
109      ---------- 
110      y : Numpy array with n > 1 elements. The class labels. These labels are 
111              used to stratify the samples. 
112   
113      B : Positive integer greater than or equal to two. The number of bootstrap 
114              samples to draw. 
115   
116      seed : Integer. A random seed to initialise the random number generator 
117              with. Use in order to obtain deterministic results. The seed is not 
118              used if the seed is None. 
119      """ 
120      y = np.array(y) 
121      n = np.prod(y.shape) 
122      y = np.reshape(y, (n, 1)) 
123   
124      if seed is not None: 
125          np.random.seed(seed) 
126   
127      # Assign the class labels to different folds 
128      labels, y_inverse = np.unique(y, return_inverse=True) 
129      count = np.bincount(y_inverse).tolist() 
130      for b in xrange(B): 
131          sample = -np.ones(y.shape, dtype=np.int) 
132          for i in xrange(len(count)): 
133              c = count[i]  # Current class 
134              cls = y_inverse == i  # Find class among samples 
135              i = np.where(cls)[0]  # Class indices 
136              # Class sample 
137              s = np.random.randint(0, c, size=c) 
138   
139              # Save the samples 
140              sample[cls] = i[s].reshape((c, 1)) 
141   
142          yield sample.ravel().tolist() 
143