-
Notifications
You must be signed in to change notification settings - Fork 13
/
nli_attack.py
1429 lines (1233 loc) · 59.1 KB
/
nli_attack.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
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
import sys
import pickle
import argparse
import os
from pathlib import Path
import numpy as np
np.random.seed(1234)
from scipy.special import softmax
import fnmatch
import criteria
import string
import pickle
import random
random.seed(0)
import csv
from fuzzywuzzy import fuzz
from InferSent.models import NLINet, InferSent
from esim.model import ESIM
from esim.data import Preprocessor
from esim.utils import correct_predictions
from collections import defaultdict
import tensorflow.compat.v1 as tf
#To make tf 2.0 compatible with tf1.0 code, we disable the tf2.0 functionalities
tf.compat.v1.disable_eager_execution()
import tensorflow_hub as hub
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader, SequentialSampler, TensorDataset
from BERT.tokenization import BertTokenizer
from BERT.modeling import BertForSequenceClassification, BertConfig
class NLI_infer_InferSent(nn.Module):
def __init__(self,
pretrained_file,
embedding_path,
data,
batch_size=32):
super(NLI_infer_InferSent, self).__init__()
# self.device = torch.device("cuda:{}".format(local_rank) if local_rank > -1 else "cpu")
# torch.cuda.set_device(local_rank)
# Retrieving model parameters from checkpoint.
config_nli_model = {
'word_emb_dim': 300,
'enc_lstm_dim': 2048,
'n_enc_layers': 1,
'dpout_model': 0.,
'dpout_fc': 0.,
'fc_dim': 512,
'bsize': batch_size,
'n_classes': 3,
'pool_type': 'max',
'nonlinear_fc': 0,
'encoder_type': 'InferSent',
'use_cuda': True,
'use_target': False,
'version': 1,
}
params_model = {'bsize': 64, 'word_emb_dim': 200, 'enc_lstm_dim': 2048,
'pool_type': 'max', 'dpout_model': 0.0, 'version': 1}
print("\t* Building model...")
self.model = NLINet(config_nli_model).cuda()
print("Reloading pretrained parameters...")
self.model.load_state_dict(torch.load(os.path.join("savedir/", "model.pickle")))
# construct dataset loader
print('Building vocab and embeddings...')
self.dataset = NLIDataset_InferSent(embedding_path, data=data, batch_size=batch_size)
def text_pred(self, text_data):
# Switch the model to eval mode.
self.model.eval()
# transform text data into indices and create batches
data_batches = self.dataset.transform_text(text_data)
# Deactivate autograd for evaluation.
probs_all = []
with torch.no_grad():
for batch in data_batches:
# Move input and output data to the GPU if one is used.
(s1_batch, s1_len), (s2_batch, s2_len) = batch
s1_batch, s2_batch = s1_batch.cuda(), s2_batch.cuda()
logits = self.model((s1_batch, s1_len), (s2_batch, s2_len))
probs = nn.functional.softmax(logits, dim=-1)
probs_all.append(probs)
return torch.cat(probs_all, dim=0)
class NLI_infer_ESIM(nn.Module):
def __init__(self,
pretrained_file,
worddict_path,
local_rank=-1,
batch_size=32):
super(NLI_infer_ESIM, self).__init__()
self.batch_size = batch_size
self.device = torch.device("cuda:{}".format(local_rank) if local_rank > -1 else "cuda")
checkpoint = torch.load(pretrained_file)
# Retrieving model parameters from checkpoint.
vocab_size = checkpoint['model']['_word_embedding.weight'].size(0)
embedding_dim = checkpoint['model']['_word_embedding.weight'].size(1)
hidden_size = checkpoint['model']['_projection.0.weight'].size(0)
num_classes = checkpoint['model']['_classification.4.weight'].size(0)
print("\t* Building model...")
self.model = ESIM(vocab_size,
embedding_dim,
hidden_size,
num_classes=num_classes,
device=self.device).to(self.device)
self.model.load_state_dict(checkpoint['model'])
# construct dataset loader
self.dataset = NLIDataset_ESIM(worddict_path)
def text_pred(self, text_data):
# Switch the model to eval mode.
self.model.eval()
device = self.device
# transform text data into indices and create batches
self.dataset.transform_text(text_data)
dataloader = DataLoader(self.dataset, shuffle=False, batch_size=self.batch_size)
# Deactivate autograd for evaluation.
probs_all = []
with torch.no_grad():
for batch in dataloader:
# Move input and output data to the GPU if one is used.
premises = batch['premise'].to(device)
premises_lengths = batch['premise_length'].to(device)
hypotheses = batch['hypothesis'].to(device)
hypotheses_lengths = batch['hypothesis_length'].to(device)
_, probs = self.model(premises,
premises_lengths,
hypotheses,
hypotheses_lengths)
probs_all.append(probs)
return torch.cat(probs_all, dim=0)
class NLI_infer_BERT(nn.Module):
def __init__(self,
pretrained_dir,
max_seq_length=128,
batch_size=32):
super(NLI_infer_BERT, self).__init__()
self.model = BertForSequenceClassification.from_pretrained(pretrained_dir, num_labels=3).cuda()
# construct dataset loader
self.dataset = NLIDataset_BERT(pretrained_dir, max_seq_length=max_seq_length, batch_size=batch_size)
def text_pred(self, text_data):
# Switch the model to eval mode.
self.model.eval()
# transform text data into indices and create batches
dataloader = self.dataset.transform_text(text_data)
probs_all = []
# for input_ids, input_mask, segment_ids in tqdm(dataloader, desc="Evaluating"):
for input_ids, input_mask, segment_ids in dataloader:
input_ids = input_ids.cuda()
input_mask = input_mask.cuda()
segment_ids = segment_ids.cuda()
with torch.no_grad():
logits = self.model(input_ids, segment_ids, input_mask)
probs = nn.functional.softmax(logits, dim=-1)
probs_all.append(probs)
return torch.cat(probs_all, dim=0)
class USE(object):
def __init__(self, cache_path):
super(USE, self).__init__()
os.environ['TFHUB_CACHE_DIR'] = cache_path
module_url = "https://tfhub.dev/google/universal-sentence-encoder-large/3"
self.embed = hub.Module(module_url)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
self.sess = tf.Session(config=config)
self.build_graph()
self.sess.run([tf.global_variables_initializer(), tf.tables_initializer()])
def build_graph(self):
self.sts_input1 = tf.placeholder(tf.string, shape=(None))
self.sts_input2 = tf.placeholder(tf.string, shape=(None))
sts_encode1 = tf.nn.l2_normalize(self.embed(self.sts_input1), axis=1)
sts_encode2 = tf.nn.l2_normalize(self.embed(self.sts_input2), axis=1)
self.cosine_similarities = tf.reduce_sum(tf.multiply(sts_encode1, sts_encode2), axis=1)
clip_cosine_similarities = tf.clip_by_value(self.cosine_similarities, -1.0, 1.0)
self.sim_scores = 1.0 - tf.acos(clip_cosine_similarities)
def semantic_sim(self, sents1, sents2):
scores = self.sess.run(
[self.sim_scores],
feed_dict={
self.sts_input1: sents1,
self.sts_input2: sents2,
})
return scores
def pick_most_similar_words_batch(src_words, sim_mat, idx2word, ret_count=10, threshold=0.):
"""
embeddings is a matrix with (d, vocab_size)
"""
sim_order = np.argsort(-sim_mat[src_words, :])[:, 1:1 + ret_count]
sim_words, sim_values = [], []
for idx, src_word in enumerate(src_words):
sim_value = sim_mat[src_word][sim_order[idx]]
mask = sim_value >= threshold
sim_word, sim_value = sim_order[idx][mask], sim_value[mask]
sim_word = [idx2word[id] for id in sim_word]
sim_words.append(sim_word)
sim_values.append(sim_value)
return sim_words, sim_values
def read_data(filepath, data_size, target_model='infersent', lowercase=False, ignore_punctuation=False, stopwords=[]):
"""
Read the premises, hypotheses and labels from some NLI dataset's
file and return them in a dictionary. The file should be in the same
form as SNLI's .txt files.
Args:
filepath: The path to a file containing some premises, hypotheses
and labels that must be read. The file should be formatted in
the same way as the SNLI (and MultiNLI) dataset.
Returns:
A dictionary containing three lists, one for the premises, one for
the hypotheses, and one for the labels in the input data.
"""
if target_model == 'bert':
labeldict = {"contradiction": 0,
"entailment": 1,
"neutral": 2}
else:
labeldict = {"entailment": 0,
"neutral": 1,
"contradiction": 2}
with open(filepath, 'r', encoding='utf8') as input_data:
premises, hypotheses, labels = [], [], []
# Translation tables to remove punctuation from strings.
punct_table = str.maketrans({key: ' '
for key in string.punctuation})
for idx, line in enumerate(input_data):
if idx >= data_size:
break
line = line.strip().split('\t')
# Ignore sentences that have no gold label.
if line[0] == '-':
continue
premise = line[1]
hypothesis = line[2]
if lowercase:
premise = premise.lower()
hypothesis = hypothesis.lower()
if ignore_punctuation:
premise = premise.translate(punct_table)
hypothesis = hypothesis.translate(punct_table)
# Each premise and hypothesis is split into a list of words.
premises.append([w for w in premise.rstrip().split()
if w not in stopwords])
hypotheses.append([w for w in hypothesis.rstrip().split()
if w not in stopwords])
labels.append(labeldict[line[0]])
return {"premises": premises,
"hypotheses": hypotheses,
"labels": labels}
class NLIDataset_ESIM(Dataset):
"""
Dataset class for Natural Language Inference datasets.
The class can be used to read preprocessed datasets where the premises,
hypotheses and labels have been transformed to unique integer indices
(this can be done with the 'preprocess_data' script in the 'scripts'
folder of this repository).
"""
def __init__(self,
worddict_path,
padding_idx=0,
bos="_BOS_",
eos="_EOS_"):
"""
Args:
data: A dictionary containing the preprocessed premises,
hypotheses and labels of some dataset.
padding_idx: An integer indicating the index being used for the
padding token in the preprocessed data. Defaults to 0.
max_premise_length: An integer indicating the maximum length
accepted for the sequences in the premises. If set to None,
the length of the longest premise in 'data' is used.
Defaults to None.
max_hypothesis_length: An integer indicating the maximum length
accepted for the sequences in the hypotheses. If set to None,
the length of the longest hypothesis in 'data' is used.
Defaults to None.
"""
self.bos = bos
self.eos = eos
self.padding_idx = padding_idx
# build word dict
with open(worddict_path, 'rb') as pkl:
self.worddict = pickle.load(pkl)
def __len__(self):
return self.num_sequences
def __getitem__(self, index):
return {
"premise": self.data["premises"][index],
"premise_length": min(self.premises_lengths[index],
self.max_premise_length),
"hypothesis": self.data["hypotheses"][index],
"hypothesis_length": min(self.hypotheses_lengths[index],
self.max_hypothesis_length)
}
def words_to_indices(self, sentence):
"""
Transform the words in a sentence to their corresponding integer
indices.
Args:
sentence: A list of words that must be transformed to indices.
Returns:
A list of indices.
"""
indices = []
# Include the beggining of sentence token at the start of the sentence
# if one is defined.
if self.bos:
indices.append(self.worddict["_BOS_"])
for word in sentence:
if word in self.worddict:
index = self.worddict[word]
else:
# Words absent from 'worddict' are treated as a special
# out-of-vocabulary word (OOV).
index = self.worddict['_OOV_']
indices.append(index)
# Add the end of sentence token at the end of the sentence if one
# is defined.
if self.eos:
indices.append(self.worddict["_EOS_"])
return indices
def transform_to_indices(self, data):
"""
Transform the words in the premises and hypotheses of a dataset, as
well as their associated labels, to integer indices.
Args:
data: A dictionary containing lists of premises, hypotheses
and labels, in the format returned by the 'read_data'
method of the Preprocessor class.
Returns:
A dictionary containing the transformed premises, hypotheses and
labels.
"""
transformed_data = {"premises": [],
"hypotheses": []}
for i, premise in enumerate(data['premises']):
# Ignore sentences that have a label for which no index was
# defined in 'labeldict'.
indices = self.words_to_indices(premise)
transformed_data["premises"].append(indices)
indices = self.words_to_indices(data["hypotheses"][i])
transformed_data["hypotheses"].append(indices)
return transformed_data
def transform_text(self, data):
# # standardize data format
# data = defaultdict(list)
# for hypothesis in hypotheses:
# data['premises'].append(premise)
# data['hypotheses'].append(hypothesis)
# transform data into indices
data = self.transform_to_indices(data)
self.premises_lengths = [len(seq) for seq in data["premises"]]
self.max_premise_length = max(self.premises_lengths)
self.hypotheses_lengths = [len(seq) for seq in data["hypotheses"]]
self.max_hypothesis_length = max(self.hypotheses_lengths)
self.num_sequences = len(data["premises"])
self.data = {
"premises": torch.ones((self.num_sequences,
self.max_premise_length),
dtype=torch.long) * self.padding_idx,
"hypotheses": torch.ones((self.num_sequences,
self.max_hypothesis_length),
dtype=torch.long) * self.padding_idx}
for i, premise in enumerate(data["premises"]):
end = min(len(premise), self.max_premise_length)
self.data["premises"][i][:end] = torch.tensor(premise[:end])
hypothesis = data["hypotheses"][i]
end = min(len(hypothesis), self.max_hypothesis_length)
self.data["hypotheses"][i][:end] = torch.tensor(hypothesis[:end])
class NLIDataset_InferSent(Dataset):
"""
Dataset class for Natural Language Inference datasets.
The class can be used to read preprocessed datasets where the premises,
hypotheses and labels have been transformed to unique integer indices
(this can be done with the 'preprocess_data' script in the 'scripts'
folder of this repository).
"""
def __init__(self,
embedding_path,
data,
word_emb_dim=300,
batch_size=32,
bos="<s>",
eos="</s>"):
"""
Args:
data: A dictionary containing the preprocessed premises,
hypotheses and labels of some dataset.
padding_idx: An integer indicating the index being used for the
padding token in the preprocessed data. Defaults to 0.
max_premise_length: An integer indicating the maximum length
accepted for the sequences in the premises. If set to None,
the length of the longest premise in 'data' is used.
Defaults to None.
max_hypothesis_length: An integer indicating the maximum length
accepted for the sequences in the hypotheses. If set to None,
the length of the longest hypothesis in 'data' is used.
Defaults to None.
"""
self.bos = bos
self.eos = eos
self.word_emb_dim = word_emb_dim
self.batch_size = batch_size
# build word dict
self.word_vec = self.build_vocab(data['premises']+data['hypotheses'], embedding_path)
def build_vocab(self, sentences, embedding_path):
word_dict = self.get_word_dict(sentences)
word_vec = self.get_embedding(word_dict, embedding_path)
print('Vocab size : {0}'.format(len(word_vec)))
return word_vec
def get_word_dict(self, sentences):
# create vocab of words
word_dict = {}
for sent in sentences:
for word in sent:
if word not in word_dict:
word_dict[word] = ''
word_dict['<s>'] = ''
word_dict['</s>'] = ''
word_dict['<oov>'] = ''
return word_dict
def get_embedding(self, word_dict, embedding_path):
# create word_vec with glove vectors
word_vec = {}
word_vec['<oov>'] = np.random.normal(size=(self.word_emb_dim))
with open(embedding_path) as f:
for line in f:
word, vec = line.split(' ', 1)
if word in word_dict:
word_vec[word] = np.array(list(map(float, vec.split())))
print('Found {0}(/{1}) words with embedding vectors'.format(
len(word_vec), len(word_dict)))
return word_vec
def get_batch(self, batch, word_vec, emb_dim=300):
# sent in batch in decreasing order of lengths (bsize, max_len, word_dim)
lengths = np.array([len(x) for x in batch])
max_len = np.max(lengths)
# print(max_len)
embed = np.zeros((max_len, len(batch), emb_dim))
for i in range(len(batch)):
for j in range(len(batch[i])):
if batch[i][j] in word_vec:
embed[j, i, :] = word_vec[batch[i][j]]
else:
embed[j, i, :] = word_vec['<oov>']
# embed[j, i, :] = np.random.normal(size=(emb_dim))
return torch.from_numpy(embed).float(), lengths
def transform_text(self, data):
# transform data into seq of embeddings
premises = data['premises']
hypotheses = data['hypotheses']
# add bos and eos
premises = [['<s>'] + premise + ['</s>'] for premise in premises]
hypotheses = [['<s>'] + hypothese + ['</s>'] for hypothese in hypotheses]
batches = []
for stidx in range(0, len(premises), self.batch_size):
# prepare batch
s1_batch, s1_len = self.get_batch(premises[stidx:stidx + self.batch_size],
self.word_vec, self.word_emb_dim)
s2_batch, s2_len = self.get_batch(hypotheses[stidx:stidx + self.batch_size],
self.word_vec, self.word_emb_dim)
batches.append(((s1_batch, s1_len), (s2_batch, s2_len)))
return batches
class InputFeatures(object):
"""A single set of features of data."""
def __init__(self, input_ids, input_mask, segment_ids):
self.input_ids = input_ids
self.input_mask = input_mask
self.segment_ids = segment_ids
class NLIDataset_BERT(Dataset):
"""
Dataset class for Natural Language Inference datasets.
The class can be used to read preprocessed datasets where the premises,
hypotheses and labels have been transformed to unique integer indices
(this can be done with the 'preprocess_data' script in the 'scripts'
folder of this repository).
"""
def __init__(self,
pretrained_dir,
max_seq_length=128,
batch_size=32):
"""
Args:
data: A dictionary containing the preprocessed premises,
hypotheses and labels of some dataset.
padding_idx: An integer indicating the index being used for the
padding token in the preprocessed data. Defaults to 0.
max_premise_length: An integer indicating the maximum length
accepted for the sequences in the premises. If set to None,
the length of the longest premise in 'data' is used.
Defaults to None.
max_hypothesis_length: An integer indicating the maximum length
accepted for the sequences in the hypotheses. If set to None,
the length of the longest hypothesis in 'data' is used.
Defaults to None.
"""
self.tokenizer = BertTokenizer.from_pretrained(pretrained_dir, do_lower_case=True)
self.max_seq_length = max_seq_length
self.batch_size = batch_size
def _truncate_seq_pair(self, tokens_a, tokens_b, max_length):
"""Truncates a sequence pair in place to the maximum length."""
# This is a simple heuristic which will always truncate the longer sequence
# one token at a time. This makes more sense than truncating an equal percent
# of tokens from each, since if one sequence is very short then each token
# that's truncated likely contains more information than a longer sequence.
while True:
total_length = len(tokens_a) + len(tokens_b)
if total_length <= max_length:
break
if len(tokens_a) > len(tokens_b):
tokens_a.pop()
else:
tokens_b.pop()
def convert_examples_to_features(self, examples, max_seq_length, tokenizer):
"""Loads a data file into a list of `InputBatch`s."""
features = []
for (ex_index, (text_a, text_b)) in enumerate(examples):
tokens_a = tokenizer.tokenize(' '.join(text_a))
tokens_b = None
if text_b:
tokens_b = tokenizer.tokenize(' '.join(text_b))
# Modifies `tokens_a` and `tokens_b` in place so that the total
# length is less than the specified length.
# Account for [CLS], [SEP], [SEP] with "- 3"
self._truncate_seq_pair(tokens_a, tokens_b, max_seq_length - 3)
else:
# Account for [CLS] and [SEP] with "- 2"
if len(tokens_a) > max_seq_length - 2:
tokens_a = tokens_a[:(max_seq_length - 2)]
tokens = ["[CLS]"] + tokens_a + ["[SEP]"]
segment_ids = [0] * len(tokens)
if tokens_b:
tokens += tokens_b + ["[SEP]"]
segment_ids += [1] * (len(tokens_b) + 1)
input_ids = tokenizer.convert_tokens_to_ids(tokens)
# The mask has 1 for real tokens and 0 for padding tokens. Only real
# tokens are attended to.
input_mask = [1] * len(input_ids)
# Zero-pad up to the sequence length.
padding = [0] * (max_seq_length - len(input_ids))
input_ids += padding
input_mask += padding
segment_ids += padding
assert len(input_ids) == max_seq_length
assert len(input_mask) == max_seq_length
assert len(segment_ids) == max_seq_length
features.append(
InputFeatures(input_ids=input_ids,
input_mask=input_mask,
segment_ids=segment_ids))
return features
def transform_text(self, data):
# transform data into seq of embeddings
eval_features = self.convert_examples_to_features(list(zip(data['premises'], data['hypotheses'])),
self.max_seq_length, self.tokenizer)
all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=self.batch_size)
return eval_dataloader
# It calculates semantic similarity between two text inputs.
# text_ls (list): First text input either original text input or previous text.
# new_texts (list): Updated text inputs.
# idx (int): Index of the word that has been changed.
# sim_score_window (int): The number of words to consider around idx. If idx = -1 consider the whole text.
def calc_sim(text_ls, new_texts, idx, sim_score_window, sim_predictor):
len_text = len(text_ls)
half_sim_score_window = (sim_score_window - 1) // 2
# Compute the starting and ending indices of the window.
if idx >= half_sim_score_window and len_text - idx - 1 >= half_sim_score_window:
text_range_min = idx - half_sim_score_window
text_range_max = idx + half_sim_score_window + 1
elif idx < half_sim_score_window and len_text - idx - 1 >= half_sim_score_window:
text_range_min = 0
text_range_max = sim_score_window
elif idx >= half_sim_score_window and len_text - idx - 1 < half_sim_score_window:
text_range_min = len_text - sim_score_window
text_range_max = len_text
else:
text_range_min = 0
text_range_max = len_text
if text_range_min < 0:
text_range_min = 0
if text_range_max > len_text:
text_range_max = len_text
if idx == -1:
text_rang_min = 0
text_range_max = len_text
semantic_sims = \
sim_predictor.semantic_sim([' '.join(text_ls[text_range_min:text_range_max])],
list(map(lambda x: ' '.join(x[text_range_min:text_range_max]), new_texts)))[0]
return semantic_sims
# Returns the hard label prediction of the target model.
# new_text (list): Text to be fed to target model.
# predictor: Target Model.
# orig_label (int): Original label.
# batch_size (int): Batch size.
def get_attack_result(hypotheses, premise, predictor, orig_label, batch_size):
new_probs = predictor({'premises': [premise] * len(hypotheses), 'hypotheses': hypotheses})
pr=(orig_label != torch.argmax(new_probs, dim=-1)).data.cpu().numpy()
return pr
# It changes the inputt text at the specified index.
# rand_idx (int): Index to be mutated.
# text_ls (list): Original text.
# pos_ls (list): POS tage list.
# new_attack (list): The changed text during genetic optimization.
# best_attack (list): The best attack until now.
# remaining_indices (list): The indices in text input different from original input.
# synonyms_dict (dict): Synonym dict for each word.
# orig_label (int): Original prediction of the target model.
# sim_score_window (int): The number of words to consider around idx.
# predictor: Target model.
# sim_predictor: USE to compute semantic similarity.
# batch_size (int): batch size.
def mutate(rand_idx, text_ls, pos_ls, premise, new_attack, best_attack, remaining_indices,
synonyms_dict, old_syns, orig_label, sim_score_window,
predictor, sim_predictor, batch_size):
# Calculates the semantic similarity before mutation.
random_text = new_attack[:]
syns = synonyms_dict[text_ls[rand_idx]]
prev_semantic_sims = calc_sim(text_ls, [best_attack], rand_idx, sim_score_window, sim_predictor)
# Gives Priority to Original Word
orig_word = 0
if random_text[rand_idx] != text_ls[rand_idx]:
temp_text = random_text[:]
temp_text[rand_idx] = text_ls[rand_idx]
pr = get_attack_result([temp_text], premise, predictor, orig_label, batch_size)
semantic_sims = calc_sim(text_ls, [temp_text], rand_idx, sim_score_window, sim_predictor)
if np.sum(pr) > 0:
orig_word = 1
return temp_text, 1 #(updated_text, queries_taken)
# If replacing with original word does not yield adversarial text, then try to replace with other synonyms.
if orig_word == 0:
final_mask = []
new_texts = []
final_texts = []
# Replace with synonyms.
for syn in syns:
# Ignore the synonym already present at position rand_idx.
if syn == best_attack[rand_idx]:
final_mask.append(0)
else:
final_mask.append(1)
temp_text = random_text[:]
temp_text[rand_idx] = syn
new_texts.append(temp_text[:])
# Filter out mutated texts that: (1) are not having same POS tag of the synonym, (2) lowers Semantic Similarity and (3) Do not satisfy adversarial criteria.
synonyms_pos_ls = [criteria.get_pos(new_text[max(rand_idx - 4, 0):rand_idx + 5])[min(4, rand_idx)]
if len(new_text) > 10 else criteria.get_pos(new_text)[rand_idx] for new_text in new_texts]
pos_mask = np.array(criteria.pos_filter(pos_ls[rand_idx], synonyms_pos_ls))
semantic_sims = calc_sim(text_ls, new_texts, rand_idx, sim_score_window, sim_predictor)
pr = get_attack_result(new_texts, premise, predictor, orig_label, batch_size)
final_mask = np.asarray(final_mask)
sem_filter = semantic_sims >= prev_semantic_sims[0]
prediction_filter = pr > 0
final_mask = final_mask*sem_filter
final_mask = final_mask*prediction_filter
final_mask = final_mask*pos_mask
sem_vals = final_mask*semantic_sims
for i in range(len(sem_vals)):
if sem_vals[i] > 0:
final_texts.append((new_texts[i], sem_vals[i]))
# Return mutated text with best semantic similarity.
final_texts.sort(key = lambda x : x[1])
final_texts.reverse()
if len(final_texts) > 0:
#old_syns[rand_idx].append(final_texts[0][0][rand_idx])
return final_texts[0][0], len(new_texts)
else:
return [], len(new_texts)
# It generates children texts from the parent texts using crossover.
# population_size (int): Size of population used.
# population (list): The population currently in the optimization process.
# parent1_idx (int): The index of parent text input 1.
# parent2_idx (int): The index of parent text input 2.
# text_ls (list): Original text.
# best_attack (list): The best attack until now in the optimization.
# max_changes (int): The number of words substituted in the best_attack.
# changed_indices (list): The indices in text input different from original input.
# sim_score_window (int): The number of words to consider around idx.
# predictor: Target model.
# sim_predictor: USE to compute semantic similarity.
# orig_label (int): Original prediction of the target model.
# batch_size (int): batch size.
def crossover(population_size, population, parent1_idx, parent2_idx,
text_ls, best_attack, max_changes, changed_indices,
sim_score_window, sim_predictor,
predictor, orig_label, batch_size):
childs = []
changes = []
# Do crossover till population_size-1.
for i in range(population_size-1):
# Generates new child.
p1 = population[parent1_idx[i]]
p2 = population[parent2_idx[i]]
assert len(p1) == len(p2)
new_child = []
for j in range(len(p1)):
if np.random.uniform() < 0.5:
new_child.append(p1[j])
else:
new_child.append(p2[j])
change = 0
cnt = 0
mismatches = 0
# Filter out crossover child which (1) Do not improve semantic similarity, (2) Have number of words substituted
# more than the current best_attack.
for k in range(len(changed_indices)):
j = changed_indices[k]
if new_child[j] == text_ls[j]:
change+=1
cnt+=1
elif new_child[j] == best_attack[j]:
change+=1
cnt+=1
elif new_child[j] != best_attack[j]:
change+=1
prev_semantic_sims = calc_sim(text_ls, [best_attack], j, sim_score_window, sim_predictor)
semantic_sims = calc_sim(text_ls, [new_child], j, sim_score_window, sim_predictor)
if semantic_sims[0] >= prev_semantic_sims[0]:
mismatches+=1
cnt+=1
if cnt==change and mismatches<=max_changes:
childs.append(new_child)
changes.append(change)
if len(childs) == 0:
return [], 0
# Filter out childs whoch do not satisfy the adversarial criteria.
pr = get_attack_result(childs, predictor, orig_label, batch_size)
final_childs = [childs[i] for i in range(len(pr)) if pr[i] > 0]
return final_childs, len(final_childs)
def attack(fuzz_val, top_k_words, qrs, sample_index, hypotheses, premise, true_label,
predictor, stop_words_set, word2idx, idx2word, cos_sim, sim_predictor=None,
import_score_threshold=-1., sim_score_threshold=0.5, sim_score_window=15, synonym_num=50,
batch_size=32):
# first check the prediction of the original text
orig_probs = predictor({'premises': [premise], 'hypotheses': [hypotheses]}).squeeze() #predictor(premise,hypothese).squeeze()
orig_label = torch.argmax(orig_probs)
orig_prob = orig_probs.max()
if true_label != orig_label:
return '', 0, 0, orig_label, orig_label, 0, 0, 0
else:
text_ls = hypotheses[:]
pos_ls = criteria.get_pos(text_ls)
len_text = len(text_ls)
if len_text < sim_score_window:
sim_score_threshold = 0.1 # shut down the similarity thresholding function
half_sim_score_window = (sim_score_window - 1) // 2
num_queries = 1
rank = {}
# get the pos and verb tense info
words_perturb = []
pos_ls = criteria.get_pos(text_ls)
pos_pref = ["ADJ", "ADV", "VERB", "NOUN"]
for pos in pos_pref:
for i in range(len(pos_ls)):
if pos_ls[i] == pos and len(text_ls[i]) > 2:
words_perturb.append((i, text_ls[i]))
random.shuffle(words_perturb)
# find synonyms and make a dict of synonyms of each word.
words_perturb = words_perturb[:top_k_words]
words_perturb_idx = [word2idx[word] for idx, word in words_perturb if word in word2idx]
synonym_words,synonym_values=[],[]
for idx in words_perturb_idx:
res = list(zip(*(cos_sim[idx])))
temp=[]
for ii in res[1]:
temp.append(idx2word[ii])
synonym_words.append(temp)
temp=[]
for ii in res[0]:
temp.append(ii)
synonym_values.append(temp)
synonyms_all = []
synonyms_dict = defaultdict(list)
for idx, word in words_perturb:
if word in word2idx:
synonyms = synonym_words.pop(0)
if synonyms:
synonyms_all.append((idx, synonyms))
synonyms_dict[word] = synonyms
# STEP 1: Random initialisation.
qrs = 0
num_changed = 0
flag = 0
th = 0
# Try substituting a random index with its random synonym.
while qrs < len(text_ls):
random_text = text_ls[:]
for i in range(len(synonyms_all)):
idx = synonyms_all[i][0]
syn = synonyms_all[i][1]
random_text[idx] = random.choice(syn)
if i >= th:
break
pr = get_attack_result([random_text], premise, predictor, orig_label, batch_size)
qrs+=1
th +=1
if th > len_text:
break
if np.sum(pr)>0:
flag = 1
break
old_qrs = qrs
# If adversarial text is not yet generated try to substitute more words than 30%.
while qrs < old_qrs + 2500 and flag == 0:
random_text = text_ls[:]
for j in range(len(synonyms_all)):
idx = synonyms_all[j][0]
syn = synonyms_all[j][1]
random_text[idx] = random.choice(syn)
if j >= len_text:
break
pr = get_attack_result([random_text], premise, predictor, orig_label, batch_size)
qrs+=1
if np.sum(pr)>0:
flag = 1
break
if flag == 1:
#print("Found "+str(sample_index))
changed = 0
for i in range(len(text_ls)):
if text_ls[i]!=random_text[i]:
changed+=1
print(changed)
# STEP 2: Search Space Reduction i.e. Move Sample Close to Boundary
while True:
choices = []
# For each word substituted in the original text, change it with its original word and compute
# the change in semantic similarity.
for i in range(len(text_ls)):
if random_text[i] != text_ls[i]:
new_text = random_text[:]
new_text[i] = text_ls[i]
semantic_sims = calc_sim(text_ls, [new_text], -1, sim_score_window, sim_predictor)
qrs+=1
pr = get_attack_result([new_text], premise, predictor, orig_label, batch_size)
if np.sum(pr) > 0:
choices.append((i,semantic_sims[0]))
# Sort the relacements by semantic similarity and replace back the words with their original
# counterparts till text remains adversarial.
if len(choices) > 0:
choices.sort(key = lambda x: x[1])
choices.reverse()
for i in range(len(choices)):
new_text = random_text[:]
new_text[choices[i][0]] = text_ls[choices[i][0]]
pr = get_attack_result([new_text], premise, predictor, orig_label, batch_size)
qrs+=1
if pr[0] == 0:
break
random_text[choices[i][0]] = text_ls[choices[i][0]]