ocroex-gmmtree-test

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

ion()
clf()

print "starting"

n = 100000

# load data if it's not already loaded (convenient for Emacs interaction)

if "data" not in dir() or len(data)!=n:
    dbfile = "chars-1.db"
    table = dbtables.Table(dbfile,"chars")
    table.verbose = 1
    table.converter("image",dbtables.SmallImage())
    table.create(image="blob",count="integer",cls="text",classes="text",key="text")
    rows = table.get(random_=0)
    r = 32
    data = zeros((n,r*r),'f')
    values = [None]*n
    i = 0
    for row in rows:
        if i>=n: break
        cls = row.cls
        if cls is None: continue
        if len(cls)!=1: continue
        if cls=="~" or cls=="_": continue
        if i%1000==0: print i
        image = array(row.image,'f')
        image /= amax(image)
        image = improc.center_maxsize(image,32)
        image /= sqrt(sum(image**2))
        data[i,:] = image[:].ravel()
        values[i] = cls
        i += 1
        
# tree = gmmtree.GmmTree()
# tree.build(data,values)

# means,sigmas = gmmtree.gmm_em(data,int(sqrt(len(data))))

k = 25

all_means = []
all_sigmas = []

for c in sorted(list(set(values))):
    indexes = array([i for i in range(len(values)) if values[i]==c],'i')
    rows = data[indexes,:]
    print "===",c,len(rows)
    if len(rows)<1000: continue
    means,sigmas,counts = gmmtree.gmm_em(rows,k)
    counts = array(counts,'i')
    order = argsort(-array(counts))
    means = means[order,:]
    sigmas = sigmas[order,:]
    counts = counts[order]
    print counts
    print "outliers",len(rows)-sum(counts),"::",len(rows),sum(counts)
    all_means.append(means)
    all_sigmas.append(sigmas)
    gmmtree.showgrid3(sigmas,sigmas,means,d=int(sqrt(k)))
    ginput(1,timeout=2.0)

all_means = row_stack(all_means)
all_sigmas = row_stack(all_sigmas)