Source Code for Module classifiers.q_classifier_mm

 1  """Subroutines for q_classifier.py that implement a multi-mu model. 
 2  That is, one where all the classes share a common covariance matrix. 
 3  """ 
 4   
 5  import math 
 6  import multivariance as M 
 7  import q_classifier as QC 
 8  from gmisclib import dictops 
 9   
10  HUGE = 1e30 
11   
12 -def count_the_classes(data): 
13          """How many examples of each class do we have?""" 
14          counts = {} 
15          for datum in data: 
16                  dictops.add_doc(counts, datum.classid, 1) 
17          return counts 
18   
19   
20   
21 -def build_classifiers(data, N, modelchoice=M.multi_mu): 
22          """Builds a set of quadratic classifiers. 
23          Data is a list of datum_tr . 
24          It returns a list: [ (evaluation_result, classifier_description), ... ] .""" 
25   
26          assert modelchoice==M.multi_mu or modelchoice==M.multi_mu_diag 
27   
28          cprms = M.meanvar(data, N, modelchoice) 
29          assert isinstance(cprms[0], M.multi_mu_with_numbers)    \ 
30                  or isinstance(cprms[0], M.multi_mu_diag_with_numbers) 
31   
32          classlist = cprms[0].desc.int_to_id.values() 
33   
34          if modelchoice==M.multi_mu_diag: 
35                  details = M.diag_quadratic(ndim=cprms[0].desc.ndim()) 
36          elif modelchoice==M.multi_mu: 
37                  details = M.quadratic(ndim=cprms[0].desc.ndim()) 
38          else: 
39                  RuntimeError, 'Cannot happen' 
40   
41          classcounts = count_the_classes(data) 
42          # Now, we build a bunch of classifiers 
43          classifiers = [] 
44          for tmp in cprms: 
45                  qc = QC.classifier() 
46                  for c in classlist: 
47                          # Bias is the Bayesean prior: classes with more training examples 
48                          # are more likely. 
49                          bias = math.log(classcounts[c]) 
50                          if modelchoice==M.multi_mu_diag: 
51                                  q = M.diag_quadratic_with_numbers(tmp.mu[c], tmp.invsigma, \ 
52                                                                                  details, bias) 
53                          elif modelchoice==M.multi_mu: 
54                                  q = M.quadratic_with_numbers(tmp.mu[c], tmp.invsigma, 
55                                                                                  details, bias) 
56                          else: 
57                                  RuntimeError, 'Cannot happen' 
58                          qc.addQuad(c, q) 
59                  classifiers.append(qc) 
60          return classifiers 
61