-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
1036 lines (876 loc) · 43.2 KB
/
utils.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 datetime
import json
import logging
import numpy as np
import os
import pytorch_lightning as pl
import random
import string
import sys
import torch
from collections import Counter, defaultdict
from torch.utils.data import Dataset, DataLoader
from torch.utils.data.dataset import ConcatDataset
from tqdm import tqdm
from transformers import AutoTokenizer
from copy import deepcopy
def sent_loader(task, split):
buf = []
with open(f'Few-NERD/data/{task}/{split}.txt') as f:
for line in f:
line = line.strip()
if line:
word, label = line.split('\t')
buf.append((word, label))
else:
words, labels = zip(*buf)
yield words, labels, {}
buf = []
#previous will lose last sent because there is not empty line
#but will not rerun all affected expriment
if buf:
words, labels = zip(*buf)
yield words, labels, {}
def episode_loader(task, split, n, k):
with open(f'Few-NERD/data/episode-data/{task}/{split}_{n}_{k}.jsonl') as f:
for line in f:
line = json.loads(line.strip())
query = line['query']
query = [(words, labels, {}) for words, labels in zip(query['word'], query['label'])]
support = line['support']
support = [(words, labels, {}) for words, labels in zip(support['word'], support['label'])]
types = line['types']
yield types, support, query
def truecase_episode_loader(mode, split, N, K):
text_2_sent = {}
for words, labels, addtional_info in sent_loader(mode, split):
sent_text = ' '.join([w.lower() for w in words])
text_2_sent[sent_text] = (words, labels)
for types, support, query in episode_loader(mode, split, N, K):
truecase_support, truecase_query = [], []
for words, labels, addtional_info in support:
sent_text = ' '.join([w for w in words])
assert sent_text in text_2_sent
truecase_words, truecase_labels = text_2_sent[sent_text]
assert len(truecase_words) == len(words)
assert all([tw.lower() == w for tw, w in zip(truecase_words, words)])
truecase_support.append((truecase_words, labels, {}))
for words, labels, addtional_info in query:
sent_text = ' '.join([w for w in words])
assert sent_text in text_2_sent
truecase_words, truecase_labels = text_2_sent[sent_text]
assert len(truecase_words) == len(words)
assert all([tw.lower() == w for tw, w in zip(truecase_words, words)])
truecase_query.append((truecase_words, labels, {}))
yield types, truecase_support, truecase_query
def get_io_spans(labels):
spans = []
span_count = Counter()
i = 0
while i < len(labels):
if type(labels[i]) == str:
if labels[i] != 'O' and labels[i] != '##NULL##':
start = i
current_label = labels[i]
i += 1
while i < len(labels) and labels[i] == current_label:
i += 1
spans.append((start, i - 1, current_label))
span_count[current_label] += 1
else:
i += 1
elif type(labels[i]) == int:
assert labels[i] >= -1
if labels[i] != 0 and labels[i] != -1:
start = i
current_label = labels[i]
i += 1
while i < len(labels) and labels[i] == current_label:
i += 1
spans.append((start, i - 1, current_label))
span_count[current_label] += 1
else:
i += 1
elif type(labels[i]) == tuple:
assert labels[i][0] >= -1
if labels[i][0] != 0 and labels[i][0] != -1:
start = i
current_label = labels[i]
i += 1
while i < len(labels) and labels[i][0] == current_label[0]:
i += 1
spans.append((start, i - 1, current_label))
span_count[current_label[0]] += 1
else:
i += 1
else:
raise Exception('Unknown label type')
return spans, span_count
class LabelEncoder:
def __init__(self, labels):
self.idx_to_item = {-1: '##NULL##', 0: 'O'}
self.item_to_idx = {'##NULL##': -1, 'O': 0}
for i, l in enumerate(labels):
self.add(l, i + 1)
def add(self, item, idx):
assert item not in self.item_to_idx
self.item_to_idx[item] = idx
self.idx_to_item[idx] = item
def get(self, idx):
if type(idx) in [list, tuple]:
return [self.get(i) for i in idx]
else:
return self.idx_to_item[idx]
def index(self, item):
if type(item) in [list, tuple]:
return [self.index(i) for i in item]
else:
return self.item_to_idx[item]
def __len__(self):
return len(self.item_to_idx)
def __str__(self):
return str(self.item_to_idx)
class Sentence:
def __init__(self, words, labels, **additional_info):
self.words = words
self.labels = labels
self.additional_info = additional_info
self.spans, self.span_count = get_io_spans(labels)
def sent_tokenize_split(self, bert_tokenizer, max_length):
token_ids = [bert_tokenizer.cls_token_id]
token_labels = ['##NULL##']
word_masks = [0]
atten_masks = [1]
for i, (w, l) in enumerate(zip(self.words, self.labels)):
subword = bert_tokenizer.tokenize(w)
assert len(subword) <= max_length - 2, subword
# todo: filter too long word, find subword seq >= 96 in ontonote or roberta, too old old, no fix
# if len(subword) >= 10:
# print(w,'long subword', len(subword))
if len(subword) == 0:
# for FEW-NERD and bert-base-(un)cased, all zero subword words are O and not in test set
# so remove it should not make a difference for test and is OK
# for SNIPS, one zero subword word in test are non-O(playlist), we need to use bio tag at eval so add
# zero_subword_remove to make log removal position and padding O afterwards
if 'zero_subword_remove' not in self.additional_info:
self.additional_info['zero_subword_remove'] = [(i,l)]
else:
self.additional_info['zero_subword_remove'].append((i, l))
print('empty subword', [ord(c) for c in w], l, self.additional_info)
continue
#todo: should be max_length - 1 since cls is also counted but this bug exist too long so no fix
if len(token_ids) + len(subword) > max_length - 1:
token_ids.append(bert_tokenizer.sep_token_id)
token_labels.append('##NULL##')
word_masks.append(0)
atten_masks.append(1)
# if sum(word_masks) <= 5:
# print(str(self),' '.join(subword_arr), 'little word sentence')
assert len(token_ids) == len(token_labels) == len(word_masks) == len(atten_masks)
yield token_ids, token_labels, word_masks, atten_masks
token_ids = [bert_tokenizer.cls_token_id]
token_labels = ['##NULL##']
word_masks = [0]
atten_masks = [1]
#old code will drop the first word, we only rerun the evaluation due to time constraint
token_ids.extend(bert_tokenizer.convert_tokens_to_ids(subword))
token_labels.extend([l] * len(subword))
word_masks.extend([1] + [0] * (len(subword) - 1))
atten_masks.extend([1] * len(subword))
else:
token_ids.extend(bert_tokenizer.convert_tokens_to_ids(subword))
token_labels.extend([l] * len(subword))
word_masks.extend([1] + [0] * (len(subword) - 1))
atten_masks.extend([1] * len(subword))
if len(token_ids) > 0:
token_ids.append(bert_tokenizer.sep_token_id)
token_labels.append('##NULL##')
word_masks.append(0)
atten_masks.append(1)
# if sum(word_masks) <= 5:
# print(str(self),' '.join(subword_arr), 'little word sentence')
assert len(token_ids) == len(token_labels) == len(word_masks) == len(atten_masks)
yield token_ids, token_labels, word_masks, atten_masks
def split_tokenize(self, bert_tokenizer, max_length):
if not hasattr(self, 'samples'):
self.samples = []
for ret in self.sent_tokenize_split(bert_tokenizer, max_length):
self.samples.append(ret)
return self.samples
else:
return self.samples
def json_repr(self):
return self.words, self.labels, self.additional_info
def __str__(self):
rep = []
for w, l in zip(self.words, self.labels):
if l == 'O':
rep.append(w)
else:
rep.append(f'{w}[{l}]')
return ' '.join(rep)
class SentenceDataset(Dataset):
def __init__(self, loader, backbone_name, max_length):
self.bert_tokenizer = AutoTokenizer.from_pretrained(backbone_name, use_fast=False)
self.max_length = max_length
self.types = set()
self.samples = []
for words, labels, additional_info in tqdm(loader):
self.types.update(labels)
self.samples.append(Sentence(words, labels, **additional_info))
self.types.remove('O')
self.types = sorted(list(self.types))
self.label_encoder = LabelEncoder(self.types)
def __getitem__(self, index):
dataset = {'token_id': [], 'label': [], 'atten_mask': [], 'word_mask': [], 'types': self.types}
sent = self.samples[index]
token_ids, token_labels, word_masks, atten_masks = zip(
*sent.split_tokenize(self.bert_tokenizer, self.max_length))
dataset['token_id'].extend(token_ids)
dataset['label'].extend(self.label_encoder.index(token_labels))
dataset['word_mask'].extend(word_masks)
dataset['atten_mask'].extend(atten_masks)
# total_num = 0
# for word_mask in word_masks:
# total_num += sum(word_mask)
# if total_num+len(sent.additional_info.get('zero_subword_remove',[])) != len(sent.words):
# print(total_num, len(sent.words), sent.json_repr())
# raise Exception('Mismatch label token num')
dataset['json'] = json.dumps(sent.json_repr())
return dataset
def __len__(self):
return len(self.samples)
def concat_dataset_until(dataset, target_size):
list_of_dataset = []
list_size = 0
while list_size < target_size:
list_of_dataset.append(dataset)
list_size += len(dataset)
extended_dataset = ConcatDataset(list_of_dataset)
return extended_dataset
class SentenceSpanDataset(Dataset):
def __init__(self, sent_dataset, neg_rate, max_len, length_limited_full_span):
self.sent_dataset = sent_dataset
# print(f'original dataset len:{self.sent_dataset_len}')
self.neg_rate = neg_rate
self.max_len = max_len
assert self.neg_rate > 0
self.length_limited_full_span = length_limited_full_span
def __getitem__(self, index):
dataset = self.sent_dataset[index]
random.seed(index)
assert 'span' not in dataset
dataset['span'], dataset['word_len'], dataset['subword_len'] = [], [], []
for token_labels, word_mask in zip(dataset['label'], dataset['word_mask']):
valid_pos = [i for i, flag in enumerate(word_mask) if flag != 0]
valid_label = [token_labels[idx] for idx in valid_pos]
spans, span_count = get_io_spans(valid_label)
valid_pos_span = [(start, end, label) for start, end, label in spans if end - start + 1 <= self.max_len]
valid_reject_set = set([(start, end) for start, end, label in spans])
valid_neg_span = []
for i in range(len(valid_label)):
for j in range(i, len(valid_label)):
if j - i + 1 <= self.max_len and (i, j) not in valid_reject_set:
valid_neg_span.append((i, j, 0))
if len(valid_neg_span) > 0:
neg_num = int(len(valid_label) * self.neg_rate) + 1
sample_num = min(neg_num, len(valid_neg_span))
sampled_valid_neg_span = random.sample(valid_neg_span, sample_num)
else:
sampled_valid_neg_span = valid_neg_span
full_span = np.ones((len(word_mask), len(word_mask), 2), dtype=np.int64) * -1
for i, j, l in valid_pos_span:
word_i, word_j = valid_pos[i], valid_pos[j]
full_span[word_i, word_j, 0] = l
if self.length_limited_full_span:
for i, j, l in valid_neg_span:
word_i, word_j = valid_pos[i], valid_pos[j]
assert l == 0
full_span[word_i, word_j, 0] = l
else:
for i in range(len(valid_label)):
for j in range(i, len(valid_label)):
if (i, j) not in valid_reject_set:
word_i, word_j = valid_pos[i], valid_pos[j]
full_span[word_i, word_j, 0] = 0
# Sample mask
full_span[:, :, 1] = 0
for i, j, l in valid_pos_span + sampled_valid_neg_span:
word_i, word_j = valid_pos[i], valid_pos[j]
assert full_span[word_i, word_j, 0] == l
full_span[word_i, word_j, 1] = 1
dataset['span'].append(full_span)
dataset['word_len'].append(len(valid_label))
dataset['subword_len'].append(len(token_labels) - 2)
return dataset
def __len__(self):
return len(self.sent_dataset)
class Episode:
def __init__(self, query, support, types):
self.types = sorted(types)
self.query_counter = Counter()
self.query = query
for sent in self.query:
self.query_counter += sent.span_count
self.support_counter = Counter()
self.support = support
for sent in self.support:
self.support_counter += sent.span_count
def json_repr(self):
support = [sent.json_repr() for sent in self.query]
query = [sent.json_repr() for sent in self.query]
return {'types':self.types, 'support':support, 'query':query}
def __str__(self):
rep = []
rep.append('{0:<30}{1:<30}{2:<30}'.format('label', 'support', 'query'))
for l in self.types:
rep.append(f'{l:<30}{self.support_counter[l]:<30}{self.query_counter[l]:<30}')
rep.append('{0:<30}{1:<30}{2:<30}'.format('total', len(self.support), len(self.query)))
return '\n'.join(rep)
def to_data(self, bert_tokenizer, max_length):
label_encoder = LabelEncoder(self.types)
dataset = {'token_id': [], 'label': [], 'atten_mask': [], 'word_mask': [], 'query_id': [], 'sent_flag': [],
'types': self.types}
for sent in self.support:
token_ids, token_labels, word_masks, atten_masks = zip(*sent.split_tokenize(bert_tokenizer, max_length))
dataset['token_id'].extend(token_ids)
dataset['label'].extend(label_encoder.index(token_labels))
dataset['word_mask'].extend(word_masks)
dataset['atten_mask'].extend(atten_masks)
dataset['sent_flag'].extend([0 for j in range(len(token_ids))])
last_sent_id = 0
for sent in self.query:
token_ids, token_labels, word_masks, atten_masks = zip(*sent.split_tokenize(bert_tokenizer, max_length))
dataset['token_id'].extend(token_ids)
dataset['label'].extend(label_encoder.index(token_labels))
dataset['word_mask'].extend(word_masks)
dataset['atten_mask'].extend(atten_masks)
dataset['query_id'].extend([last_sent_id for j in range(len(token_ids))])
dataset['sent_flag'].extend([1 for j in range(len(token_ids))])
last_sent_id += 1
dataset['json'] = json.dumps(self.json_repr())
return dataset
class EpisodeDataset(Dataset):
def __init__(self, loader, backbone_name, max_length):
self.bert_tokenizer = AutoTokenizer.from_pretrained(backbone_name, use_fast=False)
assert self.bert_tokenizer.pad_token_id == 0, 'padding token is not 0'
self.max_length = max_length
self.samples = []
self.types = set()
for types, support, query in tqdm(loader):
self.add_sample(types, support, query)
self.types.update(types)
self.types = sorted(list(self.types))
self.label_encoder = LabelEncoder(self.types)
def add_sample(self, types, support, query):
support_sent = [Sentence(words, labels, **additional_info) for words, labels, additional_info in support]
query_sent = [Sentence(words, labels, **additional_info) for words, labels, additional_info in query]
episode = Episode(query_sent, support_sent, types)
self.samples.append(episode)
def __getitem__(self, index):
# print(index, torch.utils.data.get_worker_info().id)
episode = self.samples[index]
assert len(episode.support) > 0 and len(episode.query) > 0, 'empty support or query'
dataset = episode.to_data(self.bert_tokenizer, self.max_length)
dataset['index'] = index
dataset['global_label'] = []
label_encoder = LabelEncoder(dataset['types'])
for token_labels in dataset['label']:
token_str_labels = label_encoder.get(token_labels)
token_global_labels = self.label_encoder.index(token_str_labels)
dataset['global_label'].append(token_global_labels)
return dataset
def __len__(self):
return len(self.samples)
class EpisodeSamplingDataset(Dataset):
def __init__(self, N, K, Q, loader, backbone_name, max_length):
self.K = K
self.N = N
self.Q = Q
self.bert_tokenizer = AutoTokenizer.from_pretrained(backbone_name, use_fast=False)
assert self.bert_tokenizer.pad_token_id == 0, 'padding token is not 0'
self.max_length = max_length
self.samples = []
self.sample_counter = Counter()
self.class_to_idx = defaultdict(lambda: set())
self.types = []
for words, labels, additional_info in tqdm(loader):
self.add_sample(Sentence(words, labels, **additional_info))
self.types = sorted(list(self.class_to_idx.keys()))
self.label_encoder = LabelEncoder(self.types)
self.relax_k = 2
def add_sample(self, sample):
self.sample_counter += sample.span_count
self.samples.append(sample)
for k in sample.span_count:
self.types.append(k)
self.class_to_idx[k].add(len(self.samples) - 1)
def valid(self, sample, set_counter, target_classes, k_shot):
if len(sample.span_count) == 0:
# NO EMPTY IN SAMPLING
return False
isvalid = False
for c in sample.span_count:
if c not in target_classes:
return False
if sample.span_count[c] + set_counter[c] > self.relax_k * k_shot:
return False
if set_counter[c] < k_shot:
isvalid = True
return isvalid
@staticmethod
def finish(set_counter, n_way, k_shot):
if len(set_counter) < n_way:
return False
for k in set_counter:
if set_counter[k] < k_shot:
return False
return True
def gen_candidate_idx(self, target_classes):
candidate_idx = set()
# for c in self.types:
# if c in target_classes:
# candidate_idx = candidate_idx | self.class_to_idx[c]
# else:
# candidate_idx = candidate_idx - self.class_to_idx[c]
# old code not fully elimate impossible sentence since it depends on self.types's order
for c in self.types:
if c in target_classes:
candidate_idx = candidate_idx | self.class_to_idx[c]
for c in self.types:
if c not in target_classes:
candidate_idx = candidate_idx - self.class_to_idx[c]
return sorted(list(candidate_idx))
def gen_episode(self):
support_counter = Counter()
support_idx = []
query_counter = Counter()
query_idx = []
target_classes = random.sample(self.types, self.N)
candidates = self.gen_candidate_idx(target_classes)
while not candidates:
target_classes = random.sample(self.types, self.N)
candidates = self.gen_candidate_idx(target_classes)
while not self.finish(support_counter, self.N, self.K):
index = random.choice(candidates)
if index not in support_idx:
if self.valid(self.samples[index], support_counter, target_classes, self.K):
support_counter += self.samples[index].span_count
support_idx.append(index)
while not self.finish(query_counter, self.N, self.Q):
index = random.choice(candidates)
if index not in query_idx and index not in support_idx:
if self.valid(self.samples[index], query_counter, target_classes, self.Q):
query_counter += self.samples[index].span_count
query_idx.append(index)
# print(support_idx, query_idx)
support_samples = [self.samples[idx] for idx in support_idx]
query_samples = [self.samples[idx] for idx in query_idx]
return target_classes, support_samples, query_samples
def __getitem__(self, index):
random.seed(index)
# print(index, torch.utils.data.get_worker_info().id)
target_classes, support_samples, query_samples = self.gen_episode()
episode = Episode(query_samples, support_samples, target_classes)
assert len(episode.support) > 0 and len(episode.query) > 0, 'empty support or query'
dataset = episode.to_data(self.bert_tokenizer, self.max_length)
dataset['index'] = index
dataset['global_label'] = []
label_encoder = LabelEncoder(dataset['types'])
for token_labels in dataset['label']:
token_str_labels = label_encoder.get(token_labels)
token_global_labels = self.label_encoder.index(token_str_labels)
dataset['global_label'].append(token_global_labels)
return dataset
def __len__(self):
return 1000000
class EpisodeSpanDataset(Dataset):
def __init__(self, episode_dataset, neg_rate, max_len, length_limited_full_span):
self.episode_dataset = episode_dataset
self.neg_rate = neg_rate
self.max_len = max_len
self.length_limited_full_span = length_limited_full_span
def __getitem__(self, index):
dataset = self.episode_dataset[index]
random.seed(index)
assert 'span' not in dataset
dataset['span'], dataset['word_len'], dataset['subword_len'] = [], [], []
for token_labels, token_global_labels, word_mask in zip(dataset['label'], dataset['global_label'],
dataset['word_mask']):
valid_pos = [i for i, flag in enumerate(word_mask) if flag != 0]
valid_label = [(token_labels[idx], token_global_labels[idx]) for idx in valid_pos]
spans, span_count = get_io_spans(valid_label)
valid_pos_span = [(start, end, label, global_label) for start, end, (label, global_label) in spans
if end - start + 1 <= self.max_len]
valid_reject_set = set([(start, end) for start, end, (label, global_label) in spans])
valid_neg_span = []
for i in range(len(valid_label)):
for j in range(i, len(valid_label)):
if j - i + 1 <= self.max_len and (i, j) not in valid_reject_set:
valid_neg_span.append((i, j, 0, 0))
if len(valid_neg_span) > 0 and self.neg_rate > 0:
neg_num = int(len(valid_label) * self.neg_rate) + 1
sample_num = min(neg_num, len(valid_neg_span))
sampled_valid_neg_span = random.sample(valid_neg_span, sample_num)
else:
sampled_valid_neg_span = valid_neg_span
full_span = np.ones((len(word_mask), len(word_mask), 3), dtype=np.int64) * -1
for i, j, l, gl in valid_pos_span:
word_i, word_j = valid_pos[i], valid_pos[j]
full_span[word_i, word_j, 0] = l
full_span[word_i, word_j, 1] = gl
if self.length_limited_full_span:
for i, j, l, gl in valid_neg_span:
word_i, word_j = valid_pos[i], valid_pos[j]
assert l == 0 and gl == 0
full_span[word_i, word_j, 0] = l
full_span[word_i, word_j, 1] = gl
else:
for i in range(len(valid_label)):
for j in range(i, len(valid_label)):
if (i, j) not in valid_reject_set:
word_i, word_j = valid_pos[i], valid_pos[j]
full_span[word_i, word_j, 0] = 0
full_span[word_i, word_j, 1] = 0
# Sample mask
full_span[:, :, 2] = 0
for i, j, l, gl in valid_pos_span + sampled_valid_neg_span:
word_i, word_j = valid_pos[i], valid_pos[j]
assert full_span[word_i, word_j, 0] == l
assert full_span[word_i, word_j, 1] == gl, (full_span[word_i, word_j, 1], l, gl)
full_span[word_i, word_j, 2] = 1
dataset['span'].append(full_span)
dataset['word_len'].append(len(valid_label))
dataset['subword_len'].append(len(token_labels) - 2)
return dataset
def __len__(self):
return len(self.episode_dataset)
def padding(list_of_list, pad_sign):
sizes = [len(outer_list) for outer_list in list_of_list]
max_size = max(sizes)
pad_sizes = [max_size - size for size in sizes]
return [outer_list + [pad_sign] * pad_size for outer_list, pad_size in zip(list_of_list, pad_sizes)]
def label_padding(list_of_tensor, max_dim, pad_sign):
padded_tensor_list = []
for i, tensor in enumerate(list_of_tensor):
padded_tensor = np.ones((max_dim, max_dim, tensor.shape[-1]), dtype=np.int64) * pad_sign
padded_tensor[:tensor.shape[0], :tensor.shape[1]] = tensor
padded_tensor_list.append(padded_tensor)
return np.stack(padded_tensor_list)
def episode_collate_fn(data):
dataset = {'token_id': [], 'label': [], 'global_label': [], 'atten_mask': [], 'word_mask': [], 'query_id': [],
'sent_flag': [], 'batch_id': [], 'index': [], 'batch_size': len(data), 'types': [], 'jsons':[]}
last_batch_id = 0
for episode_data in data:
dataset['token_id'].extend(episode_data['token_id'])
dataset['label'].extend(episode_data['label'])
dataset['global_label'].extend(episode_data['global_label'])
dataset['atten_mask'].extend(episode_data['atten_mask'])
dataset['word_mask'].extend(episode_data['word_mask'])
dataset['sent_flag'].extend(episode_data['sent_flag'])
dataset['query_id'].append(episode_data['query_id'])
dataset['types'].append(episode_data['types'])
dataset['jsons'].append(episode_data['json'])
dataset['index'].append(episode_data['index'])
dataset['batch_id'].extend([last_batch_id] * len(episode_data['sent_flag']))
last_batch_id += 1
dataset['token_id'] = torch.LongTensor(padding(dataset['token_id'], 0))
dataset['label'] = torch.LongTensor(padding(dataset['label'], -1))
dataset['global_label'] = torch.LongTensor(padding(dataset['global_label'], -1))
dataset['atten_mask'] = torch.LongTensor(padding(dataset['atten_mask'], 0))
dataset['word_mask'] = torch.LongTensor(padding(dataset['word_mask'], 0))
dataset['sent_flag'] = torch.LongTensor(dataset['sent_flag'])
dataset['batch_id'] = torch.LongTensor(dataset['batch_id'])
has_span = 'span' in data[0]
if has_span:
assert all(['span' in episode_data for episode_data in data])
seq_num, max_token_len = dataset['token_id'].size()
full_span = []
word_len = []
subword_len = []
for episode_data in data:
full_span.extend(episode_data['span'])
word_len.extend(episode_data['word_len'])
subword_len.extend(episode_data['subword_len'])
assert len(full_span) == dataset['token_id'].size(0)
dataset['full_span'] = torch.from_numpy(label_padding(full_span, max_token_len, -1))
dataset['word_len'] = torch.LongTensor(word_len)
dataset['subword_len'] = torch.LongTensor(subword_len)
return dataset
def sentence_collate_fn(data):
dataset = {'token_id': [], 'label': [], 'atten_mask': [], 'word_mask': [], 'sent_id': [], 'batch_size': len(data),
'jsons':[]}
last_sent_id = 0
for sent_data in data:
dataset['token_id'].extend(sent_data['token_id'])
dataset['label'].extend(sent_data['label'])
dataset['atten_mask'].extend(sent_data['atten_mask'])
dataset['word_mask'].extend(sent_data['word_mask'])
dataset['sent_id'].extend([last_sent_id] * len(sent_data['token_id']))
dataset['jsons'].append(sent_data['json'])
last_sent_id += 1
assert last_sent_id == len(data)
dataset['token_id'] = torch.LongTensor(padding(dataset['token_id'], 0))
dataset['label'] = torch.LongTensor(padding(dataset['label'], -1))
dataset['atten_mask'] = torch.LongTensor(padding(dataset['atten_mask'], 0))
dataset['word_mask'] = torch.LongTensor(padding(dataset['word_mask'], 0))
has_span = 'span' in data[0]
if has_span:
assert all(['span' in sent_data for sent_data in data])
batch_size, max_token_len = dataset['token_id'].size()
full_span = []
word_len = []
subword_len = []
for episode_data in data:
full_span.extend(episode_data['span'])
word_len.extend(episode_data['word_len'])
subword_len.extend(episode_data['subword_len'])
assert len(full_span) == dataset['token_id'].size(0)
dataset['full_span'] = torch.from_numpy(label_padding(full_span, max_token_len, -1))
dataset['word_len'] = torch.LongTensor(word_len)
dataset['subword_len'] = torch.LongTensor(subword_len)
return dataset
def validate_bio(ys):
tags, labels = ['O'], ['O']
for y in ys:
if y == 'O':
tags.append(y)
labels.append(y)
else:
tags.append(y[0])
labels.append(y[2:])
for i in range(len(ys)):
assert (tags[i], tags[i+1]) != ('O', 'I')
if (tags[i], tags[i+1]) in [('I', 'I'), ('B', 'I')]:
assert labels[i] == labels[i+1]
return ys
def convert_bio_io(bio_ys):
validate_bio(bio_ys)
io_ys = deepcopy(bio_ys)
for i, y in enumerate(io_ys):
if y == 'O':
continue
else:
tag, label = y[0], y[2:]
assert '-' not in label
io_ys[i] = label
assert len(io_ys) == len(bio_ys)
cmp_bio_io_span(io_ys, bio_ys)
return io_ys
def cmp_bio_io_span(io_ys, bio_ys):
io_spans, io_span_count = get_io_spans(io_ys)
bio_spans, bio_span_count = get_bio_spans(bio_ys)
if io_spans != bio_spans:
for i,j,l in bio_spans:
if (i,j,l) not in io_spans:
print(f'({i},{j},{l}) missing')
for i,j,l in io_spans:
if (i,j,l) not in bio_spans:
print(f'({i},{j},{l}) new')
def SNIPS_CV_loader(split, cv_i, K):
split_2_types = {
'AddToPlaylist': {'playlist', 'music_item', 'entity_name', 'playlist_owner', 'artist'},
'RateBook': {'best_rating', 'rating_value', 'object_type', 'object_select', 'object_part_of_series_type',
'rating_unit', 'object_name'},
'SearchScreeningEvent': {'timeRange', 'movie_type', 'object_type', 'object_location_type', 'spatial_relation',
'movie_name', 'location_name'},
'BookRestaurant': {'party_size_number', 'party_size_description', 'sort', 'spatial_relation', 'state',
'timeRange', 'cuisine', 'poi', 'facility', 'city', 'restaurant_name', 'country',
'restaurant_type', 'served_dish'},
'SearchCreativeWork': {'object_type', 'object_name'},
'PlayMusic': {'playlist', 'genre', 'music_item', 'service', 'album', 'track', 'year', 'sort', 'artist'},
'GetWeather': {'spatial_relation', 'state', 'timeRange', 'condition_description', 'condition_temperature',
'current_location', 'city', 'country', 'geographic_poi'}
}
if K == 5:
with open(f'FewShotTagging/ACL2020data/xval_snips_shot_5/snips-{split}-{cv_i}-shot-5.json') as f:
data = json.load(f)
named_episodes = {}
for name, episodes in data.items():
named_episodes[name] = []
for episode in episodes:
support = episode['support']
support = list(zip(support['seq_ins'], support['seq_outs']))
support_types = set([tag[2:] if tag != 'O' else 'O' for words, labels in support for tag in labels])
support = [(words, convert_bio_io(labels), {'bio_labels':labels}) for words, labels in support]
query = episode['batch']
query = list(zip(query['seq_ins'], query['seq_outs']))
query_types = set([tag[2:] if tag != 'O' else 'O' for words, labels in query for tag in labels])
query = [(words, convert_bio_io(labels), {'bio_labels':labels}) for words, labels in query]
assert len(query_types - support_types) == 0, (support_types, query_types)
named_episodes[name].append((split_2_types[name], support, query))
return named_episodes
elif K == 1:
with open(f'FewShotTagging/ACL2020data/xval_snips/snips_{split}_{cv_i}.json') as f:
data = json.load(f)
named_episodes = {}
for name, episodes in data.items():
named_episodes[name] = []
for episode in episodes:
support = episode['support']
support = list(zip(support['seq_ins'], support['seq_outs']))
support_types = set([tag[2:] if tag != 'O' else 'O' for words, labels in support for tag in labels])
support = [(words, convert_bio_io(labels), {'bio_labels':labels}) for words, labels in support]
query = episode['batch']
query = list(zip(query['seq_ins'], query['seq_outs']))
query_types = set([tag[2:] if tag != 'O' else 'O' for words, labels in query for tag in labels])
query = [(words, convert_bio_io(labels), {'bio_labels':labels}) for words, labels in query]
assert len(query_types - support_types) == 0, (support_types, query_types)
named_episodes[name].append((split_2_types[name], support, query))
return named_episodes
def get_bio_spans(labels):
spans = []
span_count = Counter()
i = 0
while i < len(labels):
assert type(labels[i]) == str, 'Unknown label type'
assert labels[i] != '##NULL##', 'Null label'
if labels[i] == 'O':
tag, label = labels[i], labels[i]
else:
#B-creative-work does not work for labels[i].split('-')
assert labels[i][0] in ['B','I'] and labels[i][1] == '-', labels
tag, label = labels[i][0], labels[i][2:]
assert label != ''
if tag == 'B':
start = i
current_label = label
i += 1
while i < len(labels):
if labels[i] == 'O':
tag, label = labels[i], labels[i]
else:
assert labels[i][0] in ['B', 'I'] and labels[i][1] == '-'
tag, label = labels[i][0], labels[i][2:]
assert label != ''
if tag == 'I' and label == current_label:
i += 1
else:
break
spans.append((start, i - 1, current_label))
span_count[current_label] += 1
else:
i += 1
return spans, span_count
def SNIPS_sent_loader(subset_names):
split_2_types = {
'AddToPlaylist':{'playlist', 'music_item', 'entity_name', 'playlist_owner', 'artist'},
'RateBook':{'best_rating', 'rating_value', 'object_type', 'object_select', 'object_part_of_series_type', 'rating_unit', 'object_name'},
'SearchScreeningEvent':{'timeRange', 'movie_type', 'object_type', 'object_location_type', 'spatial_relation', 'movie_name', 'location_name'},
'BookRestaurant':{'party_size_number', 'party_size_description', 'sort', 'spatial_relation', 'state', 'timeRange', 'cuisine', 'poi', 'facility', 'city', 'restaurant_name', 'country', 'restaurant_type', 'served_dish'},
'SearchCreativeWork':{'object_type', 'object_name'},
'PlayMusic':{'playlist', 'genre', 'music_item', 'service', 'album', 'track', 'year', 'sort', 'artist'},
'GetWeather':{'spatial_relation', 'state', 'timeRange', 'condition_description', 'condition_temperature', 'current_location', 'city', 'country', 'geographic_poi'}
}
sents = []
for split in ['train', 'valid', 'test']:
with open(f'SlotGated-SLU/data/snips/{split}/seq.in') as f:
words = [line.strip().split() for line in f]
with open(f'SlotGated-SLU/data/snips/{split}/seq.out') as f:
labels = [line.strip().split() for line in f]
with open(f'SlotGated-SLU/data/snips/{split}/label') as f:
intents = [line.strip() for line in f]
sents.extend(zip(intents, words, labels))
for intent, words, labels in sents:
assert intent in split_2_types
assert len(words) == len(labels)
if intent in subset_names:
yield words, convert_bio_io(labels), {'bio_labels':labels}
def load_ontonote(split):
from fastNLP.io.loader.conll import OntoNotesNERLoader
root_path = 'ontonote/'
train_path = root_path + '/train.txt'
dev_path = root_path + '/dev.txt'
test_path = root_path + '/test.txt'
for instance in OntoNotesNERLoader()._load(f'{root_path}/{split}.txt'):
words, labels = instance['raw_words'], instance['target']
labels = list(map(bio_2_io, labels))
yield words, convert_bio_io(labels), {'bio_labels':labels}
def load_conll(split):
from fastNLP.io.loader.conll import Conll2003NERLoader
root_path = 'conll/'
train_path = root_path + '/train.txt'
dev_path = root_path + '/dev.txt'
test_path = root_path + '/test.txt'
for instance in Conll2003NERLoader()._load(f'{root_path}/{split}.txt'):
words, labels = instance['raw_words'], instance['target']
yield words, convert_bio_io(labels), {'bio_labels':labels}
class PerClassEpisodeSamplingDataset(EpisodeSamplingDataset):
def __init__(self, N, K, Q, loader, backbone_name, max_length):
super().__init__(N, K, Q, loader, backbone_name, max_length)
self.relax_k = 2
def gen_candidate_idx(self, target_classes):
candidate_idx = set()
per_class_cand = []
for c in self.types:
if c in target_classes:
candidate_idx = candidate_idx | self.class_to_idx[c]
for c in self.types:
if c not in target_classes:
candidate_idx = candidate_idx - self.class_to_idx[c]
for c in target_classes:
valid_cand_idx = list(self.class_to_idx[c] & candidate_idx)
per_class_cand.append((valid_cand_idx, c, len(valid_cand_idx)))
per_class_cand = sorted(per_class_cand, key=lambda x: x[2])
return per_class_cand
def select(self, set_counter, set_idx, n_way, k_shot, target_classes, candidates, select_class):
if set_counter[select_class] >= k_shot:
return True
new_cand = []
total_available = 0
for index in random.sample(candidates, k=len(candidates)):
if index not in set_idx and self.valid(self.samples[index], set_counter, target_classes, k_shot):
new_cand.append(index)
total_available += self.samples[index].span_count[select_class]
if set_counter[select_class] + total_available < k_shot:
return False
for i, index in enumerate(new_cand):
set_counter += self.samples[index].span_count
set_idx.append(index)
if set_counter[select_class] >= k_shot:
return True
else:
if self.select(set_counter, set_idx, n_way, k_shot, target_classes, new_cand[i + 1:], select_class):
return True
else:
set_counter -= self.samples[index].span_count
set_idx.pop()
return False
def gen_episode(self):
support_counter = Counter()
support_idx = []
query_counter = Counter()
query_idx = []
target_classes = random.sample(self.types, self.N)
per_class_candidates = self.gen_candidate_idx(target_classes)