ocroex-gmmtree-build

#!/usr/bin/python
import code,pickle,sys,os,re,cPickle
from pylab import *
from optparse import OptionParser
import traceback
from ocrolib import gmmtree,dbtables,improc

parser = OptionParser("""
usage: %prog [options] chars.db clusters.gmmtree

""")

parser.add_option("-D","--display",help="display chars",action="store_true")
parser.add_option("-S","--reg",help="minimum value for sigma",type=float,default=0.001)
parser.add_option("-t","--table",help="database table to use for clustering",default="chars")
parser.add_option("-b","--branch",help="branch values",default="500 100")
parser.add_option("-s","--split",help="split thresholds",default="100000 10000")
parser.add_option("-v","--verbose",help="verbose output",action="store_true")
parser.add_option("-n","--maxrecords",help="max number of images per class for clustering",type=int,
                  default=1500000)
parser.add_option("-M","--mincount",help="min number of elements in a cluster",type=int,default=3)
parser.add_option("-F","--epsfactor",help="factor in estimating epsilon",type=float,default=1.5)
parser.add_option("-N","--epsnest",help="# samples to use for estimating epsilon",type=int,default=2000)
parser.add_option("-R","--norandom",help="don't randomize",action="store_true")
parser.add_option("-r","--reject",help="also train reject classes",action="store_true")
parser.add_option("-m","--multi",help="also train multi-character sequences",action="store_true")
parser.add_option("-u","--unlabeled",help="also train unlabeled chars",action="store_true")

(options,args) = parser.parse_args()

options.branch = [int(s) for s in options.branch.split()]
options.split = [int(s) for s in options.split.split()]

if len(args)==0:
    parser.print_help()
    sys.exit(0)

dbfile = args[0]
output = args[1]

table = dbtables.Table(dbfile,options.table,read_only=1)
table.verbose = options.verbose
table.converter("image",dbtables.SmallImage())
table.open(image="blob",count="integer",cls="text",classes="text")

if options.display:
    ion()
    show()
    ginput(1,timeout=0.1)

tree = gmmtree.GmmTree(branch=options.branch,ndata=options.split,mincount=options.mincount,reg=options.reg)
tree.display = options.display

r = 32
print "loading rows"
rows = table.get(random_=(not options.norandom))

print "collecting data"
count = 0
data = []
values = []
for row in rows:
    cls = row.cls
    if cls is None: cls = "_"
    if cls=="": cls = "_"
    if not options.unlabeled and cls=="_": continue
    if not options.reject and cls=="~": continue
    if not options.multi and len(cls)!=1: continue
    if count>options.maxrecords:
        print "reached",count,"records"
        break
    if count%1000==0: print count
    count += 1
    image = array(row.image,'f')
    image /= amax(image)
    image = improc.center_maxsize(image,32)
    data.append(image)
    values.append(cls)

data = array(data)
print "data",data.shape
data.shape = (len(data),prod(data.shape[1:]))
values = array(values)
print "data",data.shape

tree = gmmtree.GmmTree()
tree.build(data,values)

with open(output,"wb") as stream:
    cPickle.dump(tree,stream,2)