/
insert.py
269 lines (230 loc) · 8.87 KB
/
insert.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
from pycocoevalcap.bleu.bleu import Bleu
from pycocoevalcap.rouge.rouge import Rouge
from pycocoevalcap.cider.cider import Cider
from pycocoevalcap.meteor.meteor import Meteor
from pycocoevalcap.spice.spice import Spice
import numpy as np
import json
import operator
import argparse
import stop_words
import re
import spacy
import tqdm
from nltk.tokenize import word_tokenize
from collections import deque, defaultdict
def open_json(path):
with open(path, "r") as f:
return json.load(f)
def score(ref, hypo):
scorers = [
(Bleu(4), ["Bleu_1", "Bleu_2", "Bleu_3", "Bleu_4"]),
(Meteor(), "METEOR"),
(Rouge(), "ROUGE_L"),
(Cider(), "CIDEr"),
(Spice(), "Spice")
]
final_scores = {}
all_scores = {}
for scorer, method in scorers:
score, scores = scorer.compute_score(ref, hypo)
if type(score) == list:
for m, s in zip(method, score):
final_scores[m] = s
for m, s in zip(method, scores):
all_scores[m] = s
else:
final_scores[method] = score
all_scores[method] = scores
return final_scores, all_scores
def evaluate(ref, cand, get_scores=True):
# make dictionary
hypo = {}
for i, caption in enumerate(cand):
hypo[i] = [caption]
truth = {}
for i, caption in enumerate(ref):
truth[i] = [caption]
# compute bleu score
final_scores = score(truth, hypo)
# print out scores
print 'Bleu_1:\t ;', final_scores[0]['Bleu_1']
print 'Bleu_2:\t ;', final_scores[0]['Bleu_2']
print 'Bleu_3:\t ;', final_scores[0]['Bleu_3']
print 'Bleu_4:\t ;', final_scores[0]['Bleu_4']
print 'METEOR:\t ;', final_scores[0]['METEOR']
print 'ROUGE_L: ;', final_scores[0]['ROUGE_L']
print 'CIDEr:\t ;', final_scores[0]['CIDEr']
print 'Spice:\t', final_scores[0]['Spice']
if get_scores:
return final_scores
def organize_ner(ner):
new = defaultdict(list)
for k, v in ner.iteritems():
value = ' '.join(k.split())
if value not in stopwords:
new[v].append(value)
return new
def fill_random(cap, ner_dict):
assert cap != list
filled = []
for c in cap:
if c.split('_')[0] in named_entities and c.isupper():
ent = c.split('_')[0]
if ner_dict[ent]:
ner = np.random.choice(ner_dict[ent])
filled.append(ner)
else:
filled.append(c)
else:
filled.append(c)
return filled
def rank_sentences(cap, sent):
# make them unicode, spacy accepts only unicode
cap = unicode(cap)
sent = [unicode(s) for s in sent]
# feed them to spacy to get the vectors
cap = nlp(cap)
list_sent = [nlp(s) for s in sent]
compare = [s.similarity(cap) for s in list_sent]
# we sort the article sentences according to their similarity to produced caption
similarity = sorted([(s, c) for s, c in zip(list_sent, compare)], key=lambda x: x[1], reverse=True)
return similarity
def ner_finder(ranked_sen, score_sen, word):
for sen, sc in zip(ranked_sen, score_sen):
beg = sen.find(word)
if beg is not -1:
end = beg + len(word)
return sen[beg:end], sc
else:
return None, None
def fill_word2vec(cap, ner_dict, ner_articles, return_ners=False):
assert cap != list
filled = []
similarity = rank_sentences(' '.join(cap), ner_articles)
ranked_sen = [s[0].text for s in similarity]
score_sen = [s[1] for s in similarity]
if return_ners: ners = []
new = {}
for key, values in ner_dict.iteritems():
temp = {}
for word in values:
found, sc1 = ner_finder(ranked_sen, score_sen, re.sub('[^A-Za-z0-9]+', ' ', word))
found2, sc2 = ner_finder(ranked_sen, score_sen, word)
if found:
temp[word] = sc1
elif ner_finder(ranked_sen, score_sen, word):
temp[word] = sc2
else:
temp[word] = 0
new[key] = temp
new = {k: deque([i for i, _ in sorted(v.items(), key=operator.itemgetter(1), reverse=True)]) for k, v in
new.iteritems()}
for c in cap:
if c.split('_')[0] in named_entities and c.isupper():
ent = c.split('_')[0]
if ner_dict[ent]:
ner = new[ent].popleft()
# append it again, we might need to reuse some entites.
new[ent].append(ner)
filled.append(ner)
if return_ners: ners.append((ner, ent))
else:
filled.append(c)
else:
filled.append(c)
if return_ners:
return filled, ners
else:
return filled
def insert_word(ner_test, sen_att, ix, ner_dict, return_ner=False):
if ner_test in named_entities:
for ii in sen_att[ix]:
if ii < len(article['sentence']):
art_sen = article['sentence'][ii]
temp = [(art_sen.find(ner), ner) for ner in ner_dict[ner_test] if art_sen.find(ner) != -1]
temp = sorted(temp, key=lambda x: x[0])
if temp and return_ner: return temp[0][1], ner_test
if temp: return temp[0][1], None
else:
return ner_test, None
else:
return ner_test, None
def insert(cap, sen_att, ner_dict, return_ners=False):
new_sen = ''
words = []
if return_ners: ners = []
for ix, c in enumerate(cap):
ner_test = c.split('_')[0]
word, ner = insert_word(ner_test, sen_att, ix, ner_dict, return_ners)
if ner:
ners.append((word, ner))
words.append(word)
# new_sen += ' ' +
if return_ners:
return ' '.join(words), ners
else:
return ' '.join(words)
if __name__=='__main__':
parser = argparse.ArgumentParser()
# Input paths
parser.add_argument('--output', type=str, default='./vis/vis_show_attend_tell_default.json',
help='path to model to evaluate')
parser.add_argument('--insertion_method',type=list, default=['ctx', 'rand', 'att'],
help='rand: random insertion, ctx: context/word2vec/glove insertion, att: attention insertion')
parser.add_argument('--dump', type=bool, default=True,
help='Save the inserted captions in a json file')
test_compact = open_json('./data/test.json')
article_dataset = open_json('./data/article.json')
stopwords = stop_words.get_stop_words('en')
named_entities = ['PERSON', 'NORP', 'FAC', 'ORG', 'GPE', 'LOC', 'PRODUCT', 'EVENT', 'WORK_OF_ART', 'LANGUAGE',
'DATE', 'TIME', 'PERCENT', 'MONEY', 'QUANTITY', 'ORDINAL', 'CARDINAL']
nlp = spacy.load('en_core_web_lg', disable=['parser', 'tagger', 'ner'])
# Start the insertion process
opt = parser.parse_args()
output = open_json(opt.output)
id_to_key = {h['image_id']: h['image_path'].split('/')[1].split('_')[0] for h in output}
id_to_index = {h['cocoid']: i for i, h in enumerate(test_compact)}
ref = []
for h in tqdm.tqdm(output):
imgId = h['image_id']
index = id_to_index[imgId]
ref.append(test_compact[index]['sentences_full'][0]['raw'])
for method in opt.insertion_method:
hypo = []
if method == 'att':
att_sen = []
for h in tqdm.tqdm(output):
imgId = h['image_id']
# cap = compact_NE(h['caption'])
cap = word_tokenize(h['caption'])
key = id_to_key[imgId]
# index = id_to_index[imgId]
# ref.append(test_compact[index]['sentences_full'][0]['raw'])
ner_articles = article_dataset[key]['sentence_ner']
ner_dict = article_dataset[key]['ner']
ner_dict = organize_ner(ner_dict)
# fill the caption with named entities
if method=='ctx':
cap = fill_word2vec(cap, ner_dict, ner_articles)
cap = ' '.join(cap)
hypo.append(' '.join(cap.split()))
elif method=='rand':
cap = fill_random(cap, ner_dict)
cap = ' '.join(cap)
hypo.append(' '.join(cap.split()))
elif method=='att':
sen_att = np.array(h['sen_att']).squeeze(axis=2)
sorted_sen_att = [s.argsort()[-55:][::-1] for s in sen_att]
att_sen.append(sorted_sen_att)
article = article_dataset[key]
index = id_to_index[imgId]
ner_dict = article_dataset[key]['ner']
ner_dict = organize_ner(ner_dict)
sen, name = insert(cap, sorted_sen_att, ner_dict, True)
hypo.append(sen)
# retrieve the reference sentences
if opt.dump:
json.dump(hypo, open('./vis/%s.json' % method, 'wb'))
print('Insertion Method: %s' % method)
sc, scs = evaluate(ref, hypo)