Source Code for Module mekano.ml.svm

  1  from __future__ import with_statement 
  2   
  3  import tempfile 
  4  import os 
  5   
  6  from ..atoms import AtomVector 
  7  from multiclassifier import MultiClassifier 
  8   
  9 -class SVMClassifier: 
 10      """SVM Wrapper 
 11   
 12          >>> svm = SVMClassifier(modelfile) 
 13          >>> svm.score(av) 
 14      """ 
 15   
 16 -    def __init__(self, modelfile = None): 
 17          self.w = AtomVector() 
 18          self.b = 0.0 
 19          self.sv = 0 
 20          self.c = 1.0 
 21          self.j = None 
 22          self.tmp = "/tmp/" 
 23          self.binary = "svm_perf_learn" 
 24          if modelfile is not None: 
 25              self.readmodelfile(modelfile) 
 26   
 27 -    def readmodelfile(self, modelfile): 
 28          with open(modelfile) as fin: 
 29              line = fin.readline() 
 30              assert "SVM" in line, "Not an SVM model file!" 
 31              [fin.readline() for i in xrange(8)] 
 32              line = fin.readline() 
 33              sv = int(line.split()[0]) 
 34              line = fin.readline() 
 35              b = float(line.split()[0]) 
 36              for i in xrange(sv-1): 
 37                  a = fin.readline().rstrip().split() 
 38                  alpha_y = float(a[0]) 
 39                  av = AtomVector() 
 40                  for pairs in a[1:]: 
 41                      if "#" in pairs: break 
 42                      a, v = map(float, pairs.split(":")) 
 43                      av[a] = v*alpha_y 
 44                  self.w.addvector(av) 
 45   
 46              self.sv = sv 
 47              self.b = b 
 48   
 49 -    def train(self, ds): 
 50          assert(ds.isBinary()) 
 51          fout = tempfile.NamedTemporaryFile(suffix='svm', dir=self.tmp) 
 52          ds.toSVM(fout) 
 53          fout.file.flush() 
 54          modelfilename = fout.name + ".model" 
 55          _run("%s %s %s %s > /dev/null 2>&1" % (self.binary, _svm_params(self), fout.name, modelfilename)) 
 56          fout.close() 
 57          self.readmodelfile(modelfilename) 
 58          os.remove(modelfilename) 
 59   
 60 -    def __repr__(self): 
 61          return "<SVMClassfier len(w)=%d  b=%7.4f  #sv=%d>" % (len(self.w), self.b, self.sv) 
 62   
 63 -    def score(self, av): 
 64          # keep the shorter vector on the left side for faster dot products! 
 65          return (av * self.w) - self.b 
 66   
 67 -class SVMMultiClassifier: 
 68 -    def __init__(self): 
 69          self.mc = MultiClassifier() 
 70          self.labelset = set() 
 71          self.c = 1.0 
 72          self.j = None 
 73          self.tmp = "/tmp/" 
 74          self.binary = "svm_perf_learn" 
 75           
 76 -    def train(self, ds): 
 77          bds = ds.binarize() 
 78          positions = None 
 79          print "SVMMultiClassifier: Training with %d docs, %d labels" % (len(ds.docs), len(bds)) 
 80          for label in bds: 
 81              if positions is None: 
 82                  fout = tempfile.NamedTemporaryFile(suffix='svm', dir=self.tmp) 
 83                  positions = bds[label].toSVM(fout) 
 84              else: 
 85                  bds[label].toSVMSubsequent(fout, positions) 
 86              fout.file.flush() 
 87              modelfilename = "%s-%s.model" % (fout.name, label) 
 88              _run("%s %s %s %s > /dev/null 2>&1" % (self.binary, _svm_params(self), fout.name, modelfilename)) 
 89              self.mc.add(label, SVMClassifier(modelfilename)) 
 90              os.remove(modelfilename) 
 91          fout.close() 
 92          self.labelset = set(bds) 
 93   
 94 -    def __repr__(self): 
 95          return "<SVMMultiClassfier: %d labels>" % len(self.labelset) 
 96   
 97 -    def score(self, av): 
 98          return self.mc.score(av) 
 99   
100 -def _run(cmd): 
101      print "running:", cmd 
102      retcode = os.system(cmd) 
103      assert retcode == 0 
104   
105   
106 -def _svm_params(classifier): 
107      """Extract .c and .j from classifier object and return cmd-line options for SVM. 
108       
109      For example: -c 1.0 -j 2.0""" 
110   
111      ret = "-c %f" % classifier.c 
112      if classifier.j is not None: 
113          ret += " -j %f" % classifier.j 
114      return ret 
115