Source Code for Module mvpa.mappers.svd

 1  # emacs: -*- mode: python; py-indent-offset: 4; indent-tabs-mode: nil -*- 
 2  # vi: set ft=python 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  """Singular-value decomposition mapper""" 
10   
11  __docformat__ = 'restructuredtext' 
12   
13  import numpy as N 
14   
15  from mvpa.base.dochelpers import enhancedDocString 
16  from mvpa.mappers.base import ProjectionMapper 
17  from mvpa.featsel.helpers import ElementSelector 
18   
19  if __debug__: 
20      from mvpa.base import debug 
21   
22   
23 -class SVDMapper(ProjectionMapper): 
24      """Mapper to project data onto SVD components estimated from some dataset. 
25      """ 
26 -    def __init__(self, **kwargs): 
27          """Initialize the SVDMapper 
28   
29          :Parameters: 
30            **kwargs: 
31              All keyword arguments are passed to the ProjectionMapper 
32              constructor. 
33   
34              Note, that for the 'selector' argument this class also supports 
35              passing a `ElementSelector` instance, which will be used to 
36              determine the to be selected features, based on the singular 
37              values of each component. 
38          """ 
39          ProjectionMapper.__init__(self, **kwargs) 
40   
41          self._sv = None 
42          """Singular values of the training matrix.""" 
43   
44      __doc__ = enhancedDocString('SVDMapper', locals(), ProjectionMapper) 
45   
46   
47 -    def _train(self, dataset): 
48          """Determine the projection matrix onto the SVD components from 
49          a 2D samples x feature data matrix. 
50          """ 
51          X = N.asmatrix(dataset.samples) 
52          X = self._demeanData(X) 
53   
54          # singular value decomposition 
55          U, SV, Vh = N.linalg.svd(X, full_matrices=0) 
56   
57          # store the final matrix with the new basis vectors to project the 
58          # features onto the SVD components. And store its .H right away to 
59          # avoid computing it in forward() 
60          self._proj = Vh.H 
61   
62          # also store singular values of all components 
63          self._sv = SV 
64   
65          if __debug__: 
66              debug("MAP", "SVD was done on %s and obtained %d SVs " % 
67                    (dataset, len(SV)) + " (%d non-0, max=%f)" % 
68                    (len(SV.nonzero()), SV[0])) 
69              # .norm might be somewhat expensive to compute 
70              if "MAP_" in debug.active: 
71                  debug("MAP_", "Mixing matrix has %s shape and norm=%f" % 
72                        (self._proj.shape, N.linalg.norm(self._proj))) 
73   
74   
75 -    def selectOut(self, outIds): 
76          """Choose a subset of SVD components (and remove all others).""" 
77          # handle ElementSelector operating on SV (base class has no idea about) 
78          # XXX think about more generic interface, where some 'measure' is assigned 
79          # per each projection dimension, like in _sv in case of SVD. 
80          # May be selector could be parametrized with an instance + attribute as literal 
81          # so later on it could extract necessary values? 
82          if isinstance(self._selector, ElementSelector): 
83              ProjectionMapper.selectOut(self, self._selector(self._sv)) 
84          else: 
85              ProjectionMapper.selectOut(self, outIds) 
86   
87 -    def _computeRecon(self): 
88          """Since singular vectors are orthonormal, sufficient to take hermitian 
89          """ 
90          return self._proj.H 
91   
92      sv = property(fget=lambda self: self._sv, doc="Singular values") 
93