Source Code for Module gmisclib.multivariance_classes

  1  """Support module for multivariance.py""" 
  2   
  3  from gmisclib import gpkmisc 
  4  from gmisclib import Num 
  5  from gmisclib import die 
  6   
  7   
  8  BAYES_PRIOR_SPACE = 0   # 1 for a flat prior in sigma, 
  9  HUGE = 1e30 
 10   
 11 -class QuadraticNotNormalizable(ValueError): 
 12 -        def __init__(self, s=''): 
 13                  ValueError.__init__(self, s) 
 14   
 15   
 16   
 17   
 18   
 19 -def vec_inv_variance(start): 
 20          vv = gpkmisc.vec_variance(start) 
 21          assert Num.alltrue(Num.greater(vv, 0.0)) 
 22          return 1.0/vv 
 23   
 24   
 25 -def diag_inv_variance(start): 
 26          return gpkmisc.make_diag(vec_inv_variance(start)) 
 27   
 28   
 29   
 30 -class modeldesc: 
 31          __doc__ = """Virtual base class for a description of a model 
 32                          of a particular size.""" 
 33   
 34          LF = (1.0/3.0)**2 
 35   
 36 -        def __init__(self, ndim): 
 37                  assert ndim > 0 
 38                  self.dim = ndim 
 39   
 40 -        def ndim(self): 
 41                  """This returns the dimensionality of the data.""" 
 42                  return self.dim 
 43   
 44 -        def modeldim(self): 
 45                  """This gives the dimensionality of the model, 
 46                  i.e. the number of parameters required to specify 
 47                  the means and covariance matrix(ces).""" 
 48                  raise RuntimeError, 'Virtual method' 
 49   
 50 -        def unpack(self, prms): 
 51                  """This returns some subclass of model.""" 
 52                  raise RuntimeError, 'Virtual method' 
 53   
 54 -        def new(self, mu, invsigma): 
 55                  """Creates a model that contains data.""" 
 56                  raise RuntimeError, 'Virtual method' 
 57   
 58 -        def start(self, dataset): 
 59                  """Selects a random starting point from the dataset.""" 
 60                  raise RuntimeError, 'Virtual method' 
 61   
 62   
 63   
 64 -class model_with_numbers: 
 65          __doc__ = """Virtual base class for adding in the functions 
 66                  you need when you create a model with known parameters 
 67                  (like mu and sigma).""" 
 68   
 69 -        def __init__(self, details, bias): 
 70                  """Bias is an overall shift of the log probability up or 
 71                  down.  In a classifier, it is used to bias things toward 
 72                  one class or another.""" 
 73                  self.desc = details 
 74                  self._offset = None 
 75                  self.bias = bias 
 76   
 77 -        def logp(self, datum): 
 78                  raise RuntimeError, 'Virtual method' 
 79   
 80 -        def pack(self): 
 81                  """Returns a vector of parameters.""" 
 82                  raise RuntimeError, 'Virtual method' 
 83   
 84 -        def ndim(self): 
 85                  return self.desc.ndim() 
 86   
 87 -        def unpack(self, prms): 
 88                  return self.desc.unpack(prms) 
 89   
 90 -        def new(self, mu, invsigma): 
 91                  return self.desc.new(mu, invsigma) 
 92   
 93 -        def start(self, dataset): 
 94                  return self.desc.start(dataset) 
 95   
 96 -        def offset(self): 
 97                  if self._offset is None: 
 98                          self.addoff() 
 99                  return self._offset 
100   
101 -        def addoff(self):       # Should not be called if _offset is not None 
102                  raise RuntimeError, 'Virtual Function' 
103   
104   
105   
106   
107   
108   
109 -def _q_addoff(self):    # Will not be called if _offset is not None 
110          phasespacefac = BAYES_PRIOR_SPACE+1 
111          determinantfac = 1 - phasespacefac 
112          trace = Num.sum(Num.diagonal(self.invsigma)) 
113          if trace <= 0.0: 
114                  raise QuadraticNotNormalizable, "Input trace is nonpositive." 
115          try: 
116                  self.ev = Num.LA.Heigenvalues(self.invsigma) 
117          except Num.LA.LinAlgError, x: 
118                  die.warn('While computing the volume of the probability distribution: %s' % str(x)) 
119                  raise QuadraticNotNormalizable, x 
120          if not Num.alltrue(self.ev > 0.0): 
121                  raise QuadraticNotNormalizable, "Some eigenvalues are zero or negative." 
122          assert abs(Num.sum(self.ev)-trace) < 1e-10 + 1e-6*trace, "Bad Eigenvalues run: trace not matched: %g to %g" % (trace, Num.sum(self.ev)) 
123          assert Num.sum(self.ev) > 0.0, "Input trace is nonpositive." 
124          self._offset = Num.sum(Num.log(self.ev))*determinantfac 
125   
126   
127   
128 -def _d_addoff(self):    # Will not be called if _offset is not None 
129          phasespacefac = BAYES_PRIOR_SPACE+1 
130          determinantfac = 1 - phasespacefac 
131          self.ev = Num.array(self.invsigma, copy=True) 
132          if not Num.alltrue(self.ev > 0.0): 
133                  raise QuadraticNotNormalizable, "Some eigenvalues are zero or negative." 
134          self._offset = Num.sum(Num.log(self.ev))*determinantfac 
135