Source Code for Module mvpa.clfs.libsvmc.sens

  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  """Provide sensitivity measures for libsvm's SVM.""" 
 10   
 11  __docformat__ = 'restructuredtext' 
 12   
 13  import numpy as N 
 14   
 15  from mvpa.base import warning 
 16  from mvpa.misc.state import StateVariable 
 17  from mvpa.misc.param import Parameter 
 18  from mvpa.measures.base import Sensitivity 
 19   
 20  if __debug__: 
 21      from mvpa.base import debug 
 22   
 23 -class LinearSVMWeights(Sensitivity): 
 24      """`SensitivityAnalyzer` for the LIBSVM implementation of a linear SVM. 
 25      """ 
 26   
 27      _ATTRIBUTE_COLLECTIONS = ['params'] 
 28   
 29      biases = StateVariable(enabled=True, 
 30                             doc="Offsets of separating hyperplanes") 
 31   
 32      split_weights = Parameter(False, allowedtype='bool', 
 33                    doc="If binary classification either to sum SVs per each " 
 34                        "class separately") 
 35   
 36 -    def __init__(self, clf, **kwargs): 
 37          """Initialize the analyzer with the classifier it shall use. 
 38   
 39          :Parameters: 
 40            clf: LinearSVM 
 41              classifier to use. Only classifiers sub-classed from 
 42              `LinearSVM` may be used. 
 43          """ 
 44          # init base classes first 
 45          Sensitivity.__init__(self, clf, **kwargs) 
 46   
 47   
 48 -    def _call(self, dataset, callables=[]): 
 49          # local bindings 
 50          clf = self.clf 
 51          model = clf.model 
 52          if clf.params.regression: 
 53              nr_class = None 
 54          else: 
 55              nr_class = model.nr_class 
 56   
 57          if not nr_class in [None, 2]: 
 58              warning("You are estimating sensitivity for SVM %s trained on %d" % 
 59                      (str(self.clf), self.clf.model.nr_class) + 
 60                      " classes. Make sure that it is what you intended to do" ) 
 61   
 62          svcoef = N.matrix(model.getSVCoef()) 
 63          svs = N.matrix(model.getSV()) 
 64          rhos = N.asarray(model.getRho()) 
 65   
 66          self.biases = rhos 
 67          if self.split_weights: 
 68              if nr_class != 2: 
 69                  raise NotImplementedError, \ 
 70                        "Cannot compute per-class weights for" \ 
 71                        " non-binary classification task" 
 72              # libsvm might have different idea on the ordering 
 73              # of labels, so we would need to map them back explicitely 
 74              svm_labels = model.getLabels() # labels as assigned by libsvm 
 75              ds_labels = list(dataset.uniquelabels) # labels in the dataset 
 76              senses = [None for i in ds_labels] 
 77              # first label is given positive value 
 78              for i, (c, l) in enumerate( [(svcoef > 0, lambda x: x), 
 79                                           (svcoef < 0, lambda x: x*-1)] ): 
 80                  # convert to array, and just take the meaningful dimension 
 81                  c_ = c.A[0] 
 82                  senses[ds_labels.index(svm_labels[i])] = \ 
 83                                  (l(svcoef[:, c_] * svs[c_, :])).A[0] 
 84              weights = N.array(senses) 
 85          else: 
 86              # XXX yoh: .mean() is effectively 
 87              # averages across "sensitivities" of all paired classifiers (I 
 88              # think). See more info on this topic in svm.py on how sv_coefs 
 89              # are stored 
 90              # 
 91              # First multiply SV coefficients with the actual SVs to get 
 92              # weighted impact of SVs on decision, then for each feature 
 93              # take mean across SVs to get a single weight value 
 94              # per feature 
 95              weights = svcoef * svs 
 96   
 97          if __debug__ and 'SVM' in debug.active: 
 98              if clf.params.regression: 
 99                  nsvs = model.getTotalNSV() 
100              else: 
101                  nsvs = model.getNSV() 
102              if clf.regression: 
103                  svm_type = clf._svm_impl # type of regression 
104              else: 
105                  svm_type = '%d-class SVM(%s)' % (nr_class, clf._svm_impl) 
106              debug('SVM', 
107                    "Extracting weights for %s: #SVs=%s, " % \ 
108                    (svm_type, nsvs) + \ 
109                    " SVcoefshape=%s SVs.shape=%s Rhos=%s." % \ 
110                    (svcoef.shape, svs.shape, rhos) + \ 
111                    " Result: min=%f max=%f" % (N.min(weights), N.max(weights))) 
112   
113          return N.asarray(weights.T) 
114   
115      _customizeDocInherit = True 
116