Source Code for Module mvpa.measures.corrstability

 1  #emacs: -*- mode: python; py-indent-offset: 4; indent-tabs-mode: nil -*- 
 2  #ex: set sts=4 ts=4 sw=4 et: 
 3  ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ## 
 4  # 
 5  #   See COPYING file distributed along with the PyMVPA package for the 
 6  #   copyright and license terms. 
 7  # 
 8  ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ## 
 9  """FeaturewiseDatasetMeasure of stability of labels across chunks based 
10  on correlation.""" 
11   
12  __docformat__ = 'restructuredtext' 
13   
14  import numpy as N 
15   
16  from mvpa.measures.base import FeaturewiseDatasetMeasure 
17   
18 -class CorrStability(FeaturewiseDatasetMeasure): 
19      """`FeaturewiseDatasetMeasure` that assesses feature stability 
20      across runs for each unique label by correlating label activity 
21      for pairwise combinations of the chunks. 
22   
23      If there are multiple samples with the same label in a single 
24      chunk (as is typically the case) this algorithm will take the 
25      featurewise average of the sample activations to get a single 
26      value per label, per chunk. 
27   
28      """ 
29   
30 -    def __init__(self, attr='labels', **kwargs): 
31          """Initialize 
32   
33          :Parameters: 
34            attr : basestring 
35              Attribute to correlate across chunks. 
36          """ 
37          # init base classes first 
38          FeaturewiseDatasetMeasure.__init__(self, **kwargs) 
39   
40          self.__attr = attr 
41   
42   
43 -    def _call(self, dataset): 
44          """Computes featurewise scores.""" 
45   
46          # get the attributes (usally the labels) and the samples 
47          attrdata = eval('dataset.' + self.__attr) 
48          samples = dataset.samples 
49   
50          # take mean within chunks 
51          dat = [] 
52          labels = [] 
53          chunks = [] 
54          for c in dataset.uniquechunks: 
55              for l in N.unique(attrdata): 
56                  ind = (dataset.chunks==c)&(attrdata==l) 
57                  if ind.sum() == 0: 
58                      # no instances, so skip 
59                      continue 
60                  # append the mean, and the label/chunk info 
61                  dat.append(samples[ind,:].mean(0)) 
62                  labels.append(l) 
63                  chunks.append(c) 
64   
65          # convert to arrays 
66          dat = N.asarray(dat) 
67          labels = N.asarray(labels) 
68          chunks = N.asarray(chunks) 
69   
70          # get indices for correlation (all pairwise values across 
71          # chunks) 
72          ind1 = [] 
73          ind2 = [] 
74          for i,c1 in enumerate(N.unique(chunks)[:-1]): 
75              for c2 in N.unique(chunks)[i+1:]: 
76                  for l in N.unique(labels): 
77                      v1 = N.where((chunks==c1)&(labels==l))[0] 
78                      v2 = N.where((chunks==c2)&(labels==l))[0] 
79                      if labels[v1] == labels[v2]: 
80                          # the labels match, so add them 
81                          ind1.extend(v1) 
82                          ind2.extend(v2) 
83           
84          # convert the indices to arrays 
85          ind1 = N.asarray(ind1) 
86          ind2 = N.asarray(ind2) 
87   
88          # remove the mean from the datasets 
89          dat1 = dat[ind1,:] - dat[ind1,:].mean(0)[N.newaxis,:].repeat(dat[ind1,:].shape[0],0) 
90          dat2 = dat[ind2,:] - dat[ind2,:].mean(0)[N.newaxis,:].repeat(dat[ind2,:].shape[0],0) 
91   
92          # calculate the correlation from the covariance and std 
93          covar = (dat1*dat2).mean(0) / dat1.std(0) * dat2.std(0) 
94   
95          return covar 
96