mosesdecoder/scripts/training/flexibility_score.py
Jeroen Vermeulen ef028446f3 Add license notices to scripts.
This is not pleasant to read (and much, much less pleasant to write!) but
sort of necessary in an open project.  Right now it's quite hard to figure
out what is licensed how, which doesn't matter much to most people but can
suddenly become very important when people want to know what they're being
allowed to do.

I kept the notices as short as I could.  As far as I could see, everything
without a clear license notice is LGPL v2.1 or later.
2015-05-29 18:30:26 +07:00

193 lines
5.8 KiB
Python
Executable File

#!/usr/bin/env python
# -*- coding: utf-8 -*-
# author: Rico Sennrich
#
# This file is part of moses. Its use is licensed under the GNU Lesser General
# Public License version 2.1 or, at your option, any later version.
"""Add flexibility scores to a phrase table half.
You usually don't have to call this script directly; to add flexibility
scores to your model, run train-model.perl with the option
"--flexibility-score" (will only affect steps 5 and 6).
Usage:
python flexibility_score.py extract.context(.inv).sorted \
[--Inverse] [--Hierarchical] < phrasetable > output_file
"""
from __future__ import division
from __future__ import unicode_literals
import sys
import gzip
from collections import defaultdict
class FlexScore:
def __init__(self, inverted, hierarchical):
self.inverted = inverted
self.hierarchical = hierarchical
def store_pt(self, obj):
"""Store line in dictionary.
If we work with inverted phrase table, swap the two phrases.
"""
src, target = obj[0], obj[1]
if self.inverted:
src, target = target, src
self.phrase_pairs[src][target] = obj
def update_contextcounts(self, obj):
"""count the number of contexts a phrase pair occurs in"""
src, target = obj[0], obj[1]
self.context_counts[src][target] += 1
if obj[-1].startswith(b'<'):
self.context_counts_l[src][target] += 1
elif obj[-1].startswith(b'>'):
self.context_counts_r[src][target] += 1
elif obj[-1].startswith(b'v'):
self.context_counts_d[src][target] += 1
else:
sys.stderr.write(
b"\nERROR in line: {0}\n".format(b' ||| '.join(obj)))
sys.stderr.write(
b"ERROR: expecting one of '<, >, v' as context marker "
"in context extract file.\n")
raise ValueError
def traverse_incrementally(self, phrasetable, flexfile):
"""Traverse phrase table and phrase extract file (with context
information) incrementally without storing all in memory.
"""
increment = b''
old_increment = 1
stack = [''] * 2
# which phrase to use for sorting
sort_pt = 0
if self.inverted:
sort_pt = 1
while old_increment != increment:
old_increment = increment
self.phrase_pairs = defaultdict(dict)
self.context_counts = defaultdict(lambda: defaultdict(int))
self.context_counts_l = defaultdict(lambda: defaultdict(int))
self.context_counts_r = defaultdict(lambda: defaultdict(int))
self.context_counts_d = defaultdict(lambda: defaultdict(int))
if stack[0]:
self.store_pt(stack[0])
stack[0] = b''
if stack[1]:
self.update_contextcounts(stack[1])
stack[1] = b''
for line in phrasetable:
line = line.rstrip().split(b' ||| ')
if line[sort_pt] != increment:
increment = line[sort_pt]
stack[0] = line
break
else:
self.store_pt(line)
for line in flexfile:
line = line.rstrip().split(b' ||| ')
if line[0] + b' |' <= old_increment + b' |':
self.update_contextcounts(line)
else:
stack[1] = line
break
yield 1
def main(self, phrasetable, flexfile, output_object):
i = 0
sys.stderr.write(
"Incrementally loading phrase table "
"and adding flexibility score...")
for block in self.traverse_incrementally(phrasetable, flexfile):
self.flexprob_l = normalize(self.context_counts_l)
self.flexprob_r = normalize(self.context_counts_r)
self.flexprob_d = normalize(self.context_counts_d)
# TODO: Why this lambda? It doesn't affect sorting, does it?
sortkey = lambda x: x + b' |'
for src in sorted(self.phrase_pairs, key=sortkey):
for target in sorted(self.phrase_pairs[src], key=sortkey):
if i % 1000000 == 0:
sys.stderr.write('.')
i += 1
outline = self.write_phrase_table(src, target)
output_object.write(outline)
sys.stderr.write('done\n')
def write_phrase_table(self, src, target):
line = self.phrase_pairs[src][target]
flexscore_l = b"{0:.6g}".format(self.flexprob_l[src][target])
flexscore_r = b"{0:.6g}".format(self.flexprob_r[src][target])
line[3] += b' ' + flexscore_l + b' ' + flexscore_r
if self.hierarchical:
try:
flexscore_d = b"{0:.6g}".format(self.flexprob_d[src][target])
except KeyError:
flexscore_d = b"1"
line[3] += b' ' + flexscore_d
return b' ||| '.join(line) + b'\n'
def normalize(d):
out_dict = defaultdict(dict)
for src in d:
total = sum(d[src].values())
for target in d[src]:
out_dict[src][target] = d[src][target] / total
return out_dict
if __name__ == '__main__':
if len(sys.argv) < 1:
sys.stderr.write(
"Usage: "
"python flexibility_score.py extract.context(.inv).sorted "
"[--Inverse] [--Hierarchical] < phrasetable > output_file\n")
exit()
flexfile = sys.argv[1]
if '--Inverse' in sys.argv:
inverted = True
else:
inverted = False
if '--Hierarchical' in sys.argv:
hierarchical = True
else:
hierarchical = False
FS = FlexScore(inverted, hierarchical)
FS.main(sys.stdin, gzip.open(flexfile, 'r'), sys.stdout)