ocropus-econf

#!/usr/bin/python
# -*- coding: utf-8 -*-

import warnings,numpy,argparse,re,sys,os,os.path,unicodedata,multiprocessing,codecs
from collections import Counter
import ocrolib
from pylab import *
from scipy.ndimage import filters
from ocrolib import edist

# disable rank warnings from polyfit
warnings.simplefilter('ignore',numpy.RankWarning) 

parser = argparse.ArgumentParser(description = """
Compute the edit distances between ground truth and recognizer output.
Run with the ground truth files as arguments, and it will find the
corresponnding recognizer output files using the given extension (-x).
Missing output files are handled as empty strings, unless the -s
option is given.
""")
parser.add_argument("files",default=[],nargs='*',help="input lines")
parser.add_argument("-x","--extension",default=".txt",help="extension for recognizer output, default: %(default)s")

parser.add_argument("-k","--kind",default="exact",help="kind of comparison (exact, nospace, letdig, letters, digits, lnc), default: %(default)s")
parser.add_argument("-s","--skipmissing",action="store_true",help="don't use missing or empty output files in the calculation")

parser.add_argument("-c","--confusion",default=10,type=int,help="output this many top confusion")
parser.add_argument("-a","--allconf",default=None,help="output all confusions to this file")
parser.add_argument("-e","--perfile",default=None,help="output per-file errors to this file")
parser.add_argument("-C","--context",type=int,default=0,help="context for confusion matrix")
parser.add_argument("-Q","--parallel",type=int,default=multiprocessing.cpu_count())
args = parser.parse_args()
args.files = ocrolib.glob_all(args.files)

if not ".gt." in args.files[0]:
    sys.stderr.write("warning: compare on .gt.txt files, not .txt files\n")


def process1(fname):
    # fgt = ocrolib.allsplitext(fname)[0]+args.gtextension
    counts = Counter()
    gt = ocrolib.project_text(ocrolib.read_text(fname),kind=args.kind)
    ftxt = ocrolib.allsplitext(fname)[0]+args.extension
    missing = 0
    if os.path.exists(ftxt):
        txt = ocrolib.project_text(ocrolib.read_text(ftxt),kind=args.kind)
    else:
        missing = len(gt)
        txt = ""
    if args.confusion>0 or args.allconf is not None:
        err,cs = edist.xlevenshtein(txt,gt,context=args.context)
        for u,v in cs:
            counts[(u,v)] += 1
    else:
        err = edist.levenshtein(txt,gt)
    #assert err==xerr
    return fname,err,len(gt),missing,counts

outputs = ocrolib.parallel_map(process1,args.files,parallel=args.parallel,chunksize=10)
outputs = sorted(list(outputs))

perfile = None
if args.perfile is not None:
    perfile = codecs.open(args.perfile,"w","utf-8")

allconf = None
if args.allconf is not None:
    allconf = codecs.open(args.allconf,"w","utf-8")

errs = 0
total = 0
missing = 0
counts = Counter()
for fname,e,t,m,c in outputs:
    errs += e
    total += t
    missing += m
    counts += c
    if perfile is not None:
        perfile.write("%6d\t%6d\t%s\n"%(e,t,fname))
    if allconf is not None:
        for (a,b),v in c.most_common(1000):
            allconf.write("%s\t%s\t%s\n"%(a,b,fname))

if perfile is not None: perfile.close()
if allconf is not None: allconf.close()

sys.stderr.write("errors    %8d\n"%errs)
sys.stderr.write("missing   %8d\n"%missing)
sys.stderr.write("total     %8d\n"%total)
sys.stderr.write("err       %8.3f %%\n"%(errs*100.0/total))
sys.stderr.write("errnomiss %8.3f %%\n"%((errs-missing)*100.0/total))

if args.confusion>0:
    for (a,b),v in counts.most_common(args.confusion):
        print "%d\t%s\t%s"%(v,a,b)

print errs*1.0/total