1#!/usr/bin/env python 2""" 3LDfeatureselect.py - 4LD (Lang-Domain) feature extractor 5Marco Lui November 2011 6 7Based on research by Marco Lui and Tim Baldwin. 8 9Copyright 2011 Marco Lui <saffsd@gmail.com>. All rights reserved. 10 11Redistribution and use in source and binary forms, with or without modification, are 12permitted provided that the following conditions are met: 13 14 1. Redistributions of source code must retain the above copyright notice, this list of 15 conditions and the following disclaimer. 16 17 2. Redistributions in binary form must reproduce the above copyright notice, this list 18 of conditions and the following disclaimer in the documentation and/or other materials 19 provided with the distribution. 20 21THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER ``AS IS'' AND ANY EXPRESS OR IMPLIED 22WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND 23FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR 24CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 25CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 26SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON 27ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING 28NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF 29ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 30 31The views and conclusions contained in the software and documentation are those of the 32authors and should not be interpreted as representing official policies, either expressed 33or implied, of the copyright holder. 34""" 35 36###### 37# Default values 38# Can be overriden with command-line options 39###### 40FEATURES_PER_LANG = 300 # number of features to select for each language 41 42import os, sys, argparse 43import csv 44import marshal 45import numpy 46import multiprocessing as mp 47from collections import defaultdict 48 49from common import read_weights, Enumerator, write_features 50 51def select_LD_features(ig_lang, ig_domain, feats_per_lang, ignore_domain=False): 52 """ 53 @param ignore_domain boolean to indicate whether to use domain weights 54 """ 55 assert (ig_domain is None) or (len(ig_lang) == len(ig_domain)) 56 num_lang = len(ig_lang.values()[0]) 57 num_term = len(ig_lang) 58 59 term_index = defaultdict(Enumerator()) 60 61 62 ld = numpy.empty((num_lang, num_term), dtype=float) 63 64 for term in ig_lang: 65 term_id = term_index[term] 66 if ignore_domain: 67 ld[:, term_id] = ig_lang[term] 68 else: 69 ld[:, term_id] = ig_lang[term] - ig_domain[term] 70 71 terms = sorted(term_index, key=term_index.get) 72 # compile the final feature set 73 selected_features = dict() 74 for lang_id, lang_w in enumerate(ld): 75 term_inds = numpy.argsort(lang_w)[-feats_per_lang:] 76 selected_features[lang_id] = [terms[t] for t in term_inds] 77 78 return selected_features 79 80if __name__ == "__main__": 81 parser = argparse.ArgumentParser() 82 parser.add_argument("-o","--output", metavar="OUTPUT_PATH", help = "write selected features to OUTPUT_PATH") 83 parser.add_argument("--feats_per_lang", type=int, metavar='N', help="select top N features for each language", default=FEATURES_PER_LANG) 84 parser.add_argument("--per_lang", action="store_true", default=False, help="produce a list of features selecter per-language") 85 parser.add_argument("--no_domain_ig", action="store_true", default=False, help="use only per-langugage IG in LD calculation") 86 parser.add_argument("model", metavar='MODEL_DIR', help="read index and produce output in MODEL_DIR") 87 args = parser.parse_args() 88 89 lang_w_path = os.path.join(args.model, 'IGweights.lang.bin') 90 domain_w_path = os.path.join(args.model, 'IGweights.domain') 91 feature_path = args.output if args.output else os.path.join(args.model, 'LDfeats') 92 93 # display paths 94 print "model path:", args.model 95 print "lang weights path:", lang_w_path 96 print "domain weights path:", domain_w_path 97 print "feature output path:", feature_path 98 99 lang_w = read_weights(lang_w_path) 100 domain_w = read_weights(domain_w_path) if not args.no_domain_ig else None 101 102 features_per_lang = select_LD_features(lang_w, domain_w, args.feats_per_lang, ignore_domain=args.no_domain_ig) 103 if args.per_lang: 104 with open(feature_path + '.perlang', 'w') as f: 105 writer = csv.writer(f) 106 for i in range(len(features_per_lang)): 107 writer.writerow(map(repr,features_per_lang[i])) 108 109 110 final_feature_set = reduce(set.union, map(set, features_per_lang.values())) 111 print 'selected %d features' % len(final_feature_set) 112 113 write_features(sorted(final_feature_set), feature_path) 114 print 'wrote features to "%s"' % feature_path 115 116