From 68b05cdc63d49dbe26bfb2a99a2dd680ab70c209 Mon Sep 17 00:00:00 2001 From: "trevor.cohn" Date: Tue, 20 Jul 2010 18:37:04 +0000 Subject: Cleaned up scripts git-svn-id: https://ws10smt.googlecode.com/svn/trunk@336 ec762483-ff6d-05da-a07a-a48fb63a330f --- gi/evaluation/conditional_entropy.py | 61 ++++++++++++++++++ gi/evaluation/confusion_matrix.py | 118 +++++++++++++++++++++++++++++++++++ gi/evaluation/entropy.py | 38 +++++++++++ gi/evaluation/evaluate_entropy.py | 116 ---------------------------------- 4 files changed, 217 insertions(+), 116 deletions(-) create mode 100644 gi/evaluation/conditional_entropy.py create mode 100644 gi/evaluation/confusion_matrix.py create mode 100644 gi/evaluation/entropy.py delete mode 100644 gi/evaluation/evaluate_entropy.py diff --git a/gi/evaluation/conditional_entropy.py b/gi/evaluation/conditional_entropy.py new file mode 100644 index 00000000..356d3b1d --- /dev/null +++ b/gi/evaluation/conditional_entropy.py @@ -0,0 +1,61 @@ +#!/usr/bin/env python + +import sys, math, itertools, getopt + +def usage(): + print >>sys.stderr, 'Usage:', sys.argv[0], '[-s slash_threshold] input-1 input-2' + sys.exit(0) + +optlist, args = getopt.getopt(sys.argv[1:], 'hs:') +slash_threshold = None +for opt, arg in optlist: + if opt == '-s': + slash_threshold = int(arg) + else: + usage() +if len(args) != 2: + usage() + +ginfile = open(args[0]) +pinfile = open(args[1]) + +# evaluating: H(G | P) = sum_{g,p} p(g,p) log { p(p) / p(g,p) } +# = sum_{g,p} c(g,p)/N { log c(p) - log N - log c(g,p) + log N } +# = 1/N sum_{g,p} c(g,p) { log c(p) - log c(g,p) } +# where G = gold, P = predicted, N = number of events + +N = 0 +gold_frequencies = {} +predict_frequencies = {} +joint_frequencies = {} + +for gline, pline in itertools.izip(ginfile, pinfile): + gparts = gline.split('||| ')[1].split() + pparts = pline.split('||| ')[1].split() + assert len(gparts) == len(pparts) + + for gpart, ppart in zip(gparts, pparts): + gtag = gpart.split(':',1)[1] + ptag = ppart.split(':',1)[1] + + if slash_threshold == None or gtag.count('/') + gtag.count('\\') <= slash_threshold: + joint_frequencies.setdefault((gtag, ptag), 0) + joint_frequencies[gtag,ptag] += 1 + + predict_frequencies.setdefault(ptag, 0) + predict_frequencies[ptag] += 1 + + gold_frequencies.setdefault(gtag, 0) + gold_frequencies[gtag] += 1 + + N += 1 + +hg2p = 0 +hp2g = 0 +for (gtag, ptag), cgp in joint_frequencies.items(): + hp2g += cgp * (math.log(predict_frequencies[ptag], 2) - math.log(cgp, 2)) + hg2p += cgp * (math.log(gold_frequencies[gtag], 2) - math.log(cgp, 2)) +hg2p /= N +hp2g /= N + +print 'H(P|G)', hg2p, 'H(G|P)', hp2g, 'VI', hg2p + hp2g diff --git a/gi/evaluation/confusion_matrix.py b/gi/evaluation/confusion_matrix.py new file mode 100644 index 00000000..c5e2a379 --- /dev/null +++ b/gi/evaluation/confusion_matrix.py @@ -0,0 +1,118 @@ +#!/usr/bin/env python + +import sys, math, itertools, getopt + +def usage(): + print >>sys.stderr, 'Usage:', sys.argv[0], '[-s slash_threshold] [-p output] [-m] input-1 input-2' + sys.exit(0) + +optlist, args = getopt.getopt(sys.argv[1:], 'hs:') +slash_threshold = None +output_fname = None +show_matrix = False +for opt, arg in optlist: + if opt == '-s': + slash_threshold = int(arg) + elif opt == '-p': + output_fname = arg + elif opt == '-m': + show_matrix = True + else: + usage() +if len(args) != 2 or (not show_matrix and not output_fname): + usage() + +ginfile = open(args[0]) +pinfile = open(args[1]) + +if output_fname: + try: + import Image, ImageDraw + except ImportError: + print >>sys.stderr, "Error: Python Image Library not available. Did you forget to set your PYTHONPATH environment variable?" + sys.exit(1) + +N = 0 +gold_frequencies = {} +predict_frequencies = {} +joint_frequencies = {} + +for gline, pline in itertools.izip(ginfile, pinfile): + gparts = gline.split('||| ')[1].split() + pparts = pline.split('||| ')[1].split() + assert len(gparts) == len(pparts) + + for gpart, ppart in zip(gparts, pparts): + gtag = gpart.split(':',1)[1] + ptag = ppart.split(':',1)[1] + + if slash_threshold == None or gtag.count('/') + gtag.count('\\') <= slash_threshold: + joint_frequencies.setdefault((gtag, ptag), 0) + joint_frequencies[gtag,ptag] += 1 + + predict_frequencies.setdefault(ptag, 0) + predict_frequencies[ptag] += 1 + + gold_frequencies.setdefault(gtag, 0) + gold_frequencies[gtag] += 1 + + N += 1 + +# find top tags +gtags = gold_frequencies.items() +gtags.sort(lambda x,y: x[1]-y[1]) +gtags.reverse() +#gtags = gtags[:50] + +preds = predict_frequencies.items() +preds.sort(lambda x,y: x[1]-y[1]) +preds.reverse() + +if show_matrix: + print '%7s %7s' % ('pred', 'cnt'), + for gtag, gcount in gtags: print '%7s' % gtag, + print + print '=' * 80 + + for ptag, pcount in preds: + print '%7s %7d' % (ptag, pcount), + for gtag, gcount in gtags: + print '%7d' % joint_frequencies.get((gtag, ptag), 0), + print + + print '%7s %7d' % ('total', N), + for gtag, gcount in gtags: print '%7d' % gcount, + print + +if output_fname: + offset=10 + + image = Image.new("RGB", (len(preds), len(gtags)), (255, 255, 255)) + #hsl(hue, saturation%, lightness%) + + # re-sort preds to get a better diagonal + if False: + ptags = [] + remaining = set(predict_frequencies.keys()) + for y, (gtag, gcount) in enumerate(gtags): + best = (None, 0) + for ptag in remaining: + #pcount = predict_frequencies[ptag] + p = joint_frequencies.get((gtag, ptag), 0)# / float(pcount) + if p > best[1]: best = (ptag, p) + ptags.append(ptag) + remaining.remove(ptag) + if not remaining: break + + draw = ImageDraw.Draw(image) + for x, ptag in enumerate(ptags): + pcount = predict_frequencies[ptag] + minval = math.log(offset) + maxval = math.log(pcount + offset) + for y, (gtag, gcount) in enumerate(gtags): + f = math.log(offset + joint_frequencies.get((gtag, ptag), 0)) + z = int(240. * (maxval - f) / float(maxval - minval)) + #print x, y, z, f, maxval + draw.point([(x,y)], fill='hsl(%d, 100%%, 50%%)' % z) + del draw + image.save(output_fname) diff --git a/gi/evaluation/entropy.py b/gi/evaluation/entropy.py new file mode 100644 index 00000000..cef0dbb4 --- /dev/null +++ b/gi/evaluation/entropy.py @@ -0,0 +1,38 @@ +#!/usr/bin/env python + +import sys, math, itertools, getopt + +def usage(): + print >>sys.stderr, 'Usage:', sys.argv[0], '[-s slash_threshold] input file' + sys.exit(0) + +optlist, args = getopt.getopt(sys.argv[1:], 'hs:') +slash_threshold = None +for opt, arg in optlist: + if opt == '-s': + slash_threshold = int(arg) + else: + usage() +if len(args) != 1: + usage() + +infile = open(args[0]) +N = 0 +frequencies = {} + +for line in infile: + + for part in line.split('||| ')[1].split(): + tag = part.split(':',1)[1] + + if slash_threshold == None or tag.count('/') + tag.count('\\') <= slash_threshold: + frequencies.setdefault(gtag, 0) + frequencies[gtag] += 1 + N += 1 + +h = 0 +for tag, c in frequencies.items(): + h -= c * (math.log(c, 2) - math.log(N, 2)) +h /= N + +print 'entropy', h diff --git a/gi/evaluation/evaluate_entropy.py b/gi/evaluation/evaluate_entropy.py deleted file mode 100644 index e4980ccf..00000000 --- a/gi/evaluation/evaluate_entropy.py +++ /dev/null @@ -1,116 +0,0 @@ -#!/usr/bin/env python - -import sys, math, itertools - -ginfile = open(sys.argv[1]) -pinfile = open(sys.argv[2]) -if len(sys.argv) > 3: - slash_threshold = int(sys.argv[3]) - #print >>sys.stderr, 'slash threshold', slash_threshold -else: - slash_threshold = 99999 - -# evaluating: H(G | P) = sum_{g,p} p(g,p) log { p(p) / p(g,p) } -# = sum_{g,p} c(g,p)/N { log c(p) - log N - log c(g,p) + log N } -# = 1/N sum_{g,p} c(g,p) { log c(p) - log c(g,p) } -# where G = gold, P = predicted, N = number of events - -N = 0 -gold_frequencies = {} -predict_frequencies = {} -joint_frequencies = {} - -for gline, pline in itertools.izip(ginfile, pinfile): - gparts = gline.split('||| ')[1].split() - pparts = pline.split('||| ')[1].split() - assert len(gparts) == len(pparts) - - for gpart, ppart in zip(gparts, pparts): - gtag = gpart.split(':',1)[1] - ptag = ppart.split(':',1)[1] - - if gtag.count('/') + gtag.count('\\') <= slash_threshold: - joint_frequencies.setdefault((gtag, ptag), 0) - joint_frequencies[gtag,ptag] += 1 - - predict_frequencies.setdefault(ptag, 0) - predict_frequencies[ptag] += 1 - - gold_frequencies.setdefault(gtag, 0) - gold_frequencies[gtag] += 1 - - N += 1 - -hg2p = 0 -hp2g = 0 -for (gtag, ptag), cgp in joint_frequencies.items(): - hp2g += cgp * (math.log(predict_frequencies[ptag], 2) - math.log(cgp, 2)) - hg2p += cgp * (math.log(gold_frequencies[gtag], 2) - math.log(cgp, 2)) -hg2p /= N -hp2g /= N - -hg = 0 -for gtag, c in gold_frequencies.items(): - hg -= c * (math.log(c, 2) - math.log(N, 2)) -hg /= N - -print 'H(P|G)', hg2p, 'H(G|P)', hp2g, 'VI', hg2p + hp2g, 'H(G)', hg - -# find top tags -gtags = gold_frequencies.items() -gtags.sort(lambda x,y: x[1]-y[1]) -gtags.reverse() -#gtags = gtags[:50] - -print '%7s %7s' % ('pred', 'cnt'), -for gtag, gcount in gtags: print '%7s' % gtag, -print -print '=' * 80 - -preds = predict_frequencies.items() -preds.sort(lambda x,y: x[1]-y[1]) -preds.reverse() -for ptag, pcount in preds: - print '%7s %7d' % (ptag, pcount), - for gtag, gcount in gtags: - print '%7d' % joint_frequencies.get((gtag, ptag), 0), - print - -print '%7s %7d' % ('total', N), -for gtag, gcount in gtags: print '%7d' % gcount, -print - -if len(sys.argv) > 4: - # needs Python Image Library (PIL) - import Image, ImageDraw - - offset=10 - - image = Image.new("RGB", (len(preds), len(gtags)), (255, 255, 255)) - #hsl(hue, saturation%, lightness%) - - # resort preds to get a better diagonal - ptags = [] - remaining = set(predict_frequencies.keys()) - for y, (gtag, gcount) in enumerate(gtags): - best = (None, 0) - for ptag in remaining: - #pcount = predict_frequencies[ptag] - p = joint_frequencies.get((gtag, ptag), 0)# / float(pcount) - if p > best[1]: best = (ptag, p) - ptags.append(ptag) - remaining.remove(ptag) - if not remaining: break - - draw = ImageDraw.Draw(image) - for x, ptag in enumerate(ptags): - pcount = predict_frequencies[ptag] - minval = math.log(offset) - maxval = math.log(pcount + offset) - for y, (gtag, gcount) in enumerate(gtags): - f = math.log(offset + joint_frequencies.get((gtag, ptag), 0)) - z = int(240. * (maxval - f) / float(maxval - minval)) - #print x, y, z, f, maxval - draw.point([(x,y)], fill='hsl(%d, 100%%, 50%%)' % z) - del draw - image.save(sys.argv[4]) -- cgit v1.2.3