-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathrun_mot.py
1101 lines (907 loc) · 50 KB
/
run_mot.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
'''
Adapted from https://github.com/kojima-takeshi188/zero_shot_cot
'''
import argparse
import os
import random
import torch
from utils import *
from sentence_transformers import SentenceTransformer
from InstructorEmbedding import INSTRUCTOR
import fitlog
from collections import Counter
import tqdm
os.makedirs('fitlog_dir', exist_ok=True)
fitlog.set_log_dir('fitlog_dir')
from utils import get_kmeans_clustered_idx
import time
from multi_thread_openai_api_call import MyThread
from openai_account_manager import get_account_manager, OpenAI_API_inp_Manager_MultiThread, call_openai_multi_thread
from transformers import AutoTokenizer
from data_process_utils import extract_premise_and_hypothesis
from lm_retrieval import retrieve_demos_by_lm
from fastNLP import cache_results
from evaluations.drop_f1 import pred_to_many_f1_metrics, pred_to_one_answer_f1_metrics
nli_dataset = ['anli_a2', 'anli_a3', 'anli_a1']
answer_format_prompt = "\nYour response must end with the format \"The answer is ...\". " \
"If my question is a multi-choice question and the answer is A, your response must end with \"The answer is A.\"" \
" If the answer is Bob, your response must end with \"The answer is Bob.\""
plan_prompt = "\nLet's first understand the problem and devise a plan to solve the problem. Then, let's carry out the plan to solve the problem step by step."
zero_shot_plan_cot_prompt = answer_format_prompt + plan_prompt
def main():
args = parse_arguments()
print('*****************************')
print(args)
print('*****************************')
task = args.dataset
if task == "aqua" or task == "last_letters":
num_demo = 4
elif task == "commonsensqa":
num_demo = 7
elif task in ["strategyqa", "strategyqa_small"]:
num_demo = 6
else:
num_demo = 8
if args.num_demo > 0:
num_demo = args.num_demo
# for k,v in args.__dict__.items():
# print('{}:{}'.format(k,type(v)))
# exit()
logger.info('fitlog.add_hyper start')
logger.info('args:\n{}'.format(args))
fitlog.add_hyper(args)
logger.info('fitlog.add_hyper end')
fitlog.add_best_metric({'tmp': 1})
logger.info('num_demo:{}'.format(num_demo))
logger.info('num_cluster:{}'.format(args.clustered_retrieval))
fix_seed(args.random_seed)
# print("OPENAI_API_KEY:")
# manager = get_manager()
# print(manager.[0:15] + '**********')
# Initialize decoder class (load model and tokenizer) ...
decoder = Decoder()
print("setup data loader ...")
dataloader = setup_data_loader(args)
dataset_in_loader = dataloader.dataset
print_now()
if args.method == "few_shot":
demo = create_demo_text(args, cot_flag=False)
elif args.method == "few_shot_cot" or args.method == "auto_cot":
demo = create_demo_text(args, cot_flag=True)
elif args.method in ['my_random_sample_few_shot_cot', 'my_random_sample_few_shot']:
if args.demo_pool_from == 'gt':
demo_pool = load_gt_demo_pool(args.dataset, args.direct_answer_trigger_for_fewshot, 'cot' in args.method)
else:
demo_pool = load_lm_inference_demo_pool(args.demo_pool_path, 'cot' in args.method)
sampled_demos = random.Random(args.demo_sampling_seed).sample(demo_pool, num_demo)
sampled_demos = list(map(lambda x: x['demonstration'], sampled_demos))
demo = concat_demos(sampled_demos)
else:
pass
if 'retrieval' in args.method:
if args.demo_pool_from == 'gt':
demo_pool = load_gt_demo_pool(args.dataset, args.direct_answer_trigger_for_fewshot, 'cot' in args.method)
else:
demo_pool = load_lm_inference_demo_pool(args.demo_pool_path, 'cot' in args.method)
# logger.info('demo_pool[0]')
# logger.info(demo_pool[0])
#
# logger.info('demo_pool[1]')
# logger.info(demo_pool[1])
#
# logger.info('demo_pool[2]')
# logger.info(demo_pool[2])
# exit()
if args.retrieval_hybrid_with_task_demos == 'random_sample_from_demo_pool':
sampled_demos = random.Random(args.demo_sampling_seed).sample(demo_pool, num_demo)
sampled_demos = list(map(lambda x: x['demonstration'], sampled_demos))
task_level_demo = concat_demos(sampled_demos)
elif args.retrieval_hybrid_with_task_demos == 'manual':
task_level_demo = create_demo_text(args, cot_flag=True)
elif args.retrieval_hybrid_with_task_demos is None or args.retrieval_hybrid_with_task_demos.lower() == 'none':
pass
else:
raise NotImplementedError
premise_to_demo_idxs_dict = {}
premise_list = []
if args.dataset in nli_dataset:
for idx, d in enumerate(tqdm.tqdm(demo_pool, desc='building premise index')):
premise = extract_premise_and_hypothesis(d['question'])['premise']
premise_list.append(premise)
if premise in premise_to_demo_idxs_dict:
premise_to_demo_idxs_dict[premise].append(idx)
else:
premise_to_demo_idxs_dict[premise] = [idx]
logger.info('nli demo pool\'s premises: {}'.format(len(premise_to_demo_idxs_dict)))
if 'instructor' not in args.retriever_name:
retriever = SentenceTransformer(args.retriever_name)
else:
retriever = INSTRUCTOR(args.retriever_name)
if args.query_encoding == 'x' and args.demo_encoding == 'x':
q_retrieval_instruction = 'Represent the question for retrieving duplicate questions: '
d_retrieval_instruction = 'Represent the question for retrieving duplicate questions: '
elif args.query_encoding == 'z' and args.demo_encoding == 'z':
q_retrieval_instruction = 'Represent the rationale for retrieving duplicate rationales: '
d_retrieval_instruction = 'Represent the rationale for retrieving duplicate rationales: '
elif args.query_encoding == 'x' and args.demo_encoding == 'z':
q_retrieval_instruction = 'Represent the question for retrieving relevant rationales: '
d_retrieval_instruction = 'Represent the rationale for retrieval: '
else:
logger.info(
'Invalid [query_encoding:{} and demo_encoding:{}]'.format(args.query_encoding, args.demo_encoding))
raise NotImplementedError
logger.info('q_retrieval_instruction:{}'.format(q_retrieval_instruction))
logger.info('d_retrieval_instruction:{}'.format(d_retrieval_instruction))
# assert args.query_encoding == 'x'
# assert args.demo_encoding == 'x'
if args.demo_encoding == 'x':
demo_text_to_encode_pool = list(map(lambda x: x['question'], demo_pool))
elif args.demo_encoding == 'z':
demo_text_to_encode_pool = list(map(lambda x: x['rationale'], demo_pool))
else:
raise NotImplementedError
if 'instructor' in args.retriever_name:
instruction_plus_demo_text = list(
map(lambda x: [d_retrieval_instruction, x], demo_text_to_encode_pool))
else:
instruction_plus_demo_text = demo_text_to_encode_pool
pool = retriever.start_multi_process_pool()
os.makedirs('embeddings_caches', exist_ok=True)
@cache_results(_cache_fp='./embeddings_caches/demo_embedding_{}_{}_{}_{}_{}_{}'.format(args.dataset,
args.retriever_name.replace(
'/', '_'),
args.demo_encoding,
args.query_encoding,
args.demo_pool_path.replace(
'/', '_'),
args.demo_pool_from),
_refresh=False)
def get_demo_embeddings():
demo_embeddings = retriever.encode_multi_process(instruction_plus_demo_text, pool, batch_size=128)
return demo_embeddings
demo_embeddings = get_demo_embeddings()
logger.info('demo_embeddings:{}'.format(demo_embeddings.shape))
if args.clustered_retrieval > 0:
assert args.clustered_retrieval >= num_demo
demo_cluster_idxs = get_kmeans_clustered_idx(demo_embeddings, args.clustered_retrieval).tolist()
else:
demo_cluster_idxs = None
logger.info('demo_cluster_idxs:\n{}'.format(demo_cluster_idxs))
counter = Counter(demo_cluster_idxs)
logger.info('clusters size:{}'.format(counter))
# import counter
# demo_embeddings =
fitlog.add_hyper({'demo_pool_size': len(demo_pool)})
else:
demo_pool = None
demo_embeddings = None
fitlog.add_hyper({'demo_pool_size': 0})
pass
gpt2_tokenizer = AutoTokenizer.from_pretrained('gpt2-large')
total = 0
correct_list = []
em_list_single_gold = []
f1_list_single_gold = []
em_list_multiple_gold = []
f1_list_multiple_gold = []
total_example_number = len(dataset_in_loader)
if 'retrieval' in args.method:
if args.query_encoding == 'x':
questions_text_to_encode = dataset_in_loader.questions
elif args.query_encoding == 'z':
assert args.dataset in ['gsm8k', 'aqua', 'strategyqa_small']
questions_text_to_encode = dataset_in_loader.rationales
else:
raise NotImplementedError
if 'instructor' in args.retriever_name:
instruction_plus_test_questions = list(
map(lambda x: [q_retrieval_instruction, x], questions_text_to_encode))
else:
instruction_plus_test_questions = questions_text_to_encode
@cache_results(_cache_fp='./embeddings_caches/query_embedding_{}_{}_{}_{}_{}_{}'.format(args.dataset,
args.retriever_name.replace(
'/', '_'),
args.demo_encoding,
args.query_encoding,
args.demo_pool_path.replace(
'/', '_'),
args.demo_pool_from),
_refresh=False)
def get_query_embeddings():
query_embeddings = retriever.encode_multi_process(instruction_plus_test_questions, pool)
return query_embeddings
query_embeddings = get_query_embeddings()
logger.info('query_embeddings:{}'.format(query_embeddings.shape))
retriever.stop_multi_process_pool(pool)
# second_start = time.time()
demos_num_clip_times = 0
datas = list(dataset_in_loader)
if args.method in ['lm_retrieval_few_shot_cot', 'lm_retrieval_few_shot',
'lm_retrieval_few_shot_cot_but_no_thinking']:
logger.info('method is lm_retrieval_few_shot_cot, so start retrieve demos by lm')
demos_group_s_for_gpt_to_decode = []
for i, data in enumerate(list(dataset_in_loader)):
x, y = data
if args.dataset in nli_dataset:
x = x + "\n" + "A:"
else:
x = "Q: " + x + "\n" + "A:"
y = y
if type(y) is str:
y = y.strip()
demo_scores = np.matmul(query_embeddings[i:i + 1], demo_embeddings.T)[0]
demo_scores = torch.from_numpy(demo_scores)
if args.dataset in nli_dataset:
tmp_premise = extract_premise_and_hypothesis(x)['premise']
if args.do_not_retrieve_same_premise_demo_with_test:
if tmp_premise in premise_to_demo_idxs_dict:
demo_scores[premise_to_demo_idxs_dict[tmp_premise]] = demo_scores[premise_to_demo_idxs_dict[
tmp_premise]] - 999
if args.clustered_retrieval == 0:
if args.how_to_divide_demos_for_retrieval == 'score_division':
# 1-10，11-20，21-30，31-40各一组去检索
_, demo_idxs = torch.topk(demo_scores, k=num_demo * 10)
demo_idxs_group = []
for j in range(num_demo):
demo_idxs_group.append(demo_idxs[j * 10: (j + 1) * 10])
elif args.how_to_divide_demos_for_retrieval == 'score_mod':
# 1，5，9，13……一组，2，6，10，14……一组
_, demo_idxs = torch.topk(demo_scores, k=num_demo * 10)
demo_idxs_group = []
for j in range(num_demo):
demo_idxs_group.append([])
for k in range(10):
demo_idxs_group[-1].append(demo_idxs[j + k * num_demo])
else:
raise NotImplementedError
elif args.clustered_retrieval > 0:
sorted_demo_idxs = torch.argsort(demo_scores, dim=0, descending=True).tolist()
demo_idxs_group = []
for j in range(args.num_demo):
demo_idxs_group.append([])
for j in range(len(sorted_demo_idxs)):
tmp_demo_idx = sorted_demo_idxs[j]
if len(demo_idxs_group[demo_cluster_idxs[tmp_demo_idx]]) < 10:
demo_idxs_group[demo_cluster_idxs[tmp_demo_idx]].append(tmp_demo_idx)
if all(list(map(lambda x: len(x) == 10, demo_idxs_group))):
break
else:
raise NotImplementedError
# demo_idxs_group
demos_group = []
for demo_idxs in demo_idxs_group:
demos_group.append([])
for tmp_demo_idx in demo_idxs:
demos_group[-1].append(demo_pool[tmp_demo_idx])
demos_group_s_for_gpt_to_decode.append([demos_group, x[3:-3]])
if (args.limit_dataset_size != 0) and ((i + 1) >= args.limit_dataset_size):
break
lm_retrieval_hyper_parameter = dict(model=args.model, n=1, top_p=1, temperature=0,
max_length=64)
logger.info('start retrieve_demos_by_lm')
os.makedirs('lm_r_cache', exist_ok=True)
os.makedirs('lm_r_cache/{}'.format(args.dataset), exist_ok=True)
if args.num_demo == 4:
@cache_results(
_cache_fp='lm_r_cache/{}/{}_{}_{}_{}'.format(args.dataset, args.retriever_name,
args.clustered_retrieval, args.model,
args.demo_pool_path.replace('/', '_')), _hash_param=True,_refresh=True)
def tmp():
lm_retrieval_result = retrieve_demos_by_lm(demos_group_s_for_gpt_to_decode,
lm_retrieval_hyper_parameter,
args.multi_thread, 1,
args.demos_for_retrieval_using_purely_question,
args.shuffle_demos_for_lm_retrieval,
args.lm_format_requirement_at_last)
return lm_retrieval_result
lm_retrieval_result = tmp()
else:
@cache_results(
_cache_fp='lm_r_cache/{}/{}_{}_{}_{}_{}'.format(args.dataset, args.retriever_name,
args.clustered_retrieval, args.model,
args.demo_pool_path.replace('/', '_'), args.num_demo),
_hash_param=True)
def tmp():
lm_retrieval_result = retrieve_demos_by_lm(demos_group_s_for_gpt_to_decode,
lm_retrieval_hyper_parameter,
args.multi_thread, 1,
args.demos_for_retrieval_using_purely_question,
args.shuffle_demos_for_lm_retrieval,
args.lm_format_requirement_at_last)
return lm_retrieval_result
lm_retrieval_result = tmp()
demos_for_every_x = lm_retrieval_result['retrieved_demos']
fitlog.add_best_metric({'parsing_error_p': lm_retrieval_result['parsing_error_p']}, name='lm_retrieval')
fitlog.add_best_metric(
{'actual_num_demos_for_retrieval_avg': lm_retrieval_result['actual_num_demos_for_retrieval_avg']},
name='lm_retrieval')
logger.info('method is lm_retrieval_few_shot_cot, retrieve demos by lm finish')
logger.info('demos_for_every_x:\n')
logger.info(demos_for_every_x[0])
demo_for_every_x_final = []
token_num_list_save_for_debug = []
demos_correct_p_for_every_x = []
for i, d_dict_s in enumerate(demos_for_every_x):
x, y = datas[i]
if args.dataset in nli_dataset:
x = x + "\n" + "A:"
else:
x = "Q: " + x + "\n" + "A:"
tmp_demos = []
# tmp_demos_correct_p = []
demos_correct_p_for_every_x.append(0)
for d_dict in d_dict_s:
if 'cot' in args.method:
tmp_demos.append(d_dict['demonstration'])
else:
tmp_demos.append(d_dict['demostration_without_rationale'])
demos_correct_p_for_every_x[-1]+=int(d_dict['gold_ans'] == d_dict['pred_ans'].replace('.',''))
while 1:
tmp_demo = concat_demos(tmp_demos)
if len(gpt2_tokenizer.tokenize(tmp_demo + x)) > 3600:
tmp_demos = tmp_demos[:-1]
else:
break
if len(tmp_demos) == 0:
break
if len(tmp_demos) < num_demo:
demos_num_clip_times += 1
tmp_demo = concat_demos(tmp_demos)
token_num_list_save_for_debug.append(tmp_demo + x)
demo_for_every_x_final.append(tmp_demo)
# with jsonlines.open('tmp_debug_1.jsonl','w') as out_f:
# for tmp_js in token_num_list_save_for_debug:
# out_f.write(tmp_js)
logger.info('lm_retrieval_few_shot_cot: demos_num_clip_times:{}'.format(demos_num_clip_times))
# exit()
logger.info('demos_correct_p_for_every_x count:{}'.format(Counter(demos_correct_p_for_every_x)))
x_list_to_decode = []
demos_num_clip_times = 0
with open(args.output_dir, "a") as wp:
logger.info('args.limit_dataset_size:{}'.format(args.limit_dataset_size))
for i, data in enumerate(dataset_in_loader):
if i < args.resume_id - 1:
# if i < 297:
continue
output_line = {}
# print('*************************')
# print("{}st data".format(i + 1))
# Prepare question template ...
x, y = data
if args.dataset in nli_dataset:
x = x
else:
x = "Q: " + x
x = x + "\n" + "A:"
# logger.info(i)
# logger.info('y:{}'.format(y))
y = y
if type(y) is str:
y = y.strip()
# logger.info('y[0]:{}'.format(y))
# print(x, y)
output_line["question"] = x
output_line["gold_ans"] = y
if args.method == "zero_shot":
x = x + " " + args.direct_answer_trigger_for_zeroshot
elif args.method == "zero_shot_cot":
x = x + " " + args.cot_trigger
elif args.method == "few_shot":
x = demo + x
elif args.method == "few_shot_cot":
x = demo + x
elif args.method == 'my_random_sample_few_shot':
x = demo + x
elif args.method == 'my_random_sample_few_shot_cot':
x = demo + x
elif args.method == "auto_cot":
x = demo + x + " " + args.cot_trigger
elif args.method == 'retrieval_few_shot_cot' or args.method == 'retrieval_few_shot':
if args.clustered_retrieval == 0:
demo_scores = np.matmul(query_embeddings[i:i + 1], demo_embeddings.T)[0]
demo_scores = torch.from_numpy(demo_scores)
if args.dataset in nli_dataset:
tmp_premise = extract_premise_and_hypothesis(x)['premise']
if args.do_not_retrieve_same_premise_demo_with_test:
if tmp_premise in premise_to_demo_idxs_dict:
demo_scores[premise_to_demo_idxs_dict[tmp_premise]] = demo_scores[
premise_to_demo_idxs_dict[
tmp_premise]] - 999
# demo_scores
pass
# print(demo_scores.size())
demos = []
if args.dataset in nli_dataset and args.do_not_retrieve_same_premise_demos:
sorted_demo_idxs = torch.argsort(demo_scores, dim=0, descending=True).tolist()
tmp_demo_premises = set()
for j, idx in enumerate(sorted_demo_idxs):
if len(demos) >= num_demo:
break
if premise_list[idx] not in tmp_demo_premises:
demos.append(demo_pool[idx]['demonstration'])
tmp_demo_premises.add(premise_list[idx])
else:
continue
else:
_, demo_idxs = torch.topk(demo_scores, k=num_demo)
# print(demo_idxs)
for idx in demo_idxs:
if 'cot' in args.method:
demos.append(demo_pool[idx]['demonstration'])
else:
demos.append(demo_pool[idx]['demostration_without_rationale'])
# print('demos:{}'.format(demos))
while 1:
tmp_demo = concat_demos(demos)
if len(gpt2_tokenizer.tokenize(tmp_demo + x)) > 3600:
demos = demos[:-1]
else:
break
if len(demos) == 0:
break
if len(demos) < num_demo:
demos_num_clip_times += 1
demo = concat_demos(demos)
if args.retrieval_hybrid_with_task_demos in ['manual', 'random_sample_from_demo_pool']:
#warn: if retireving the demos from zero_shot_plan_prompt, the task demo may not have the zero_shot_plan_prompt
demo = task_level_demo + demo
if zero_shot_plan_cot_prompt in demo:
x = x.replace('\nA:','')
x = x + zero_shot_plan_cot_prompt
x = x+ '\nA:'
x = demo + x
elif args.clustered_retrieval > 0:
demo_scores = np.matmul(query_embeddings[i:i + 1], demo_embeddings.T)[0]
demo_scores = torch.from_numpy(demo_scores)
if args.dataset in nli_dataset:
tmp_premise = extract_premise_and_hypothesis(x)['premise']
if args.do_not_retrieve_same_premise_demo_with_test:
if tmp_premise in premise_to_demo_idxs_dict:
demo_scores[premise_to_demo_idxs_dict[tmp_premise]] = demo_scores[
premise_to_demo_idxs_dict[
tmp_premise]] - 999
# print(demo_scores.size())
# _, demo_idxs = torch.topk(demo_scores, k=)
sorted_demo_idxs = torch.argsort(demo_scores, dim=0, descending=True).tolist()
top_cluster_idxs = []
for j in range(100):
top_cluster_idxs.append(demo_cluster_idxs[sorted_demo_idxs[j]])
# logger.info('top_cluster_idxs:{}'.format(top_cluster_idxs))
tmp_demo_idxs = []
demos = []
tmp_demo_cluster_idxs = []
for j, demo_idx in enumerate(sorted_demo_idxs):
if len(demos) >= num_demo:
break
if demo_cluster_idxs[demo_idx] not in tmp_demo_cluster_idxs:
tmp_demo_idxs.append(demo_idx)
demos.append(demo_pool[demo_idx]['demonstration'])
tmp_demo_cluster_idxs.append(demo_cluster_idxs[demo_idx])
else:
continue
while 1:
tmp_demo = concat_demos(demos)
if len(gpt2_tokenizer.tokenize(tmp_demo + x)) > 3600:
demos = demos[:-1]
else:
break
if len(demos) == 0:
break
if len(demos) < num_demo:
demos_num_clip_times += 1
demo = concat_demos(demos)
if zero_shot_plan_cot_prompt in demo:
x = x.replace('\nA:','')
x = x + zero_shot_plan_cot_prompt
x = x+ '\nA:'
x = demo + x
# print('demo_idxs:{}'.format(tmp_demo_idxs))
# print('demo_cluster_idxs:{}'.format(tmp_demo_cluster_idxs))
# while len(demos)<num_demo:
pass
elif args.method in ['lm_retrieval_few_shot_cot', 'lm_retrieval_few_shot',
'lm_retrieval_few_shot_cot_but_no_thinking']:
demo = demo_for_every_x_final[i]
if args.retrieval_hybrid_with_task_demos in ['manual', 'random_sample_from_demo_pool']:
demo = task_level_demo + demo
x = demo + x
if args.method == 'lm_retrieval_few_shot_cot_but_no_thinking':
x = x + ' I need you to straightly output the answer.'
x = [x, 'The answer is']
pass
else:
raise ValueError("method is not properly defined ...")
# Answer experiment by generating text ...
# response = decoder.decode(args, x, max_length)
x_list_to_decode.append(x)
if (args.limit_dataset_size != 0) and ((i + 1) >= args.limit_dataset_size):
break
# with jsonlines.open('tmp_debug_2.jsonl', 'w') as out_f:
# for tmp_js in x_list_to_decode:
# out_f.write(tmp_js)
for i in range(min(len(x_list_to_decode), 3)):
print('*' * 50)
logger.info('x_list_to_ddecode[{}]:'.format(i))
print(x_list_to_decode[i])
print('\n')
# exit()
# idx_x_list_to_decode = list(enumerate(x_list_to_decode))
if 'no_thinking' in args.method:
args.max_length_cot = args.max_length_direct
max_length = args.max_length_cot if "cot" in args.method else args.max_length_direct
logger.info('demos_num_clip_times:{} / {}'.format(demos_num_clip_times, len(x_list_to_decode)))
# exit()
if args.multi_thread < 2:
# manager = get_manager()
args.multi_thread_api = 0
responses_with_idx = []
for idx, x in tqdm.tqdm(list(enumerate(x_list_to_decode))):
response = decoder.decode(args, x, max_length)
responses_with_idx.append([idx, response])
else:
n = 1 if args.decoding_method == 'greedy' else args.self_consistency_paths
hyper_parameter = dict(model=args.model, n=n, top_p=args.top_p, temperature=args.temperature,
max_length=max_length)
logger.info('hyper_parameter:\n{}'.format(hyper_parameter))
assert (len(dataloader) == len(x_list_to_decode))
# start filtering too long input
tmp = list(zip(datas, x_list_to_decode))
def tmp_filter_func(x):
tokenized = gpt2_tokenizer.tokenize(x[1])
if len(tokenized) > 3650:
print('excessive token number:{}'.format(len(tokenized)))
return 0
else:
return 1
filtered_tmp = list(
filter(lambda x: tmp_filter_func(x), tqdm.tqdm(tmp, desc='length filtering')))
# filtered_tmp = list(
# filter(lambda x: 1, tqdm.tqdm(tmp, desc='length filtering')))
logger.info('examples before length filtering: {}'.format(len(tmp)))
logger.info('examples after length filtering: {}'.format(len(filtered_tmp)))
if len(filtered_tmp) < len(tmp):
logger.info(
'there are too long ones in x_list_to_decode, this makes the discrepancy of tested examples, so stop')
exit()
datas = list(map(lambda x: x[0], filtered_tmp))
x_list_to_decode = list(map(lambda x: x[1], filtered_tmp))
idx_x_list_to_decode = list(enumerate(x_list_to_decode))
if args.inference_split == 'train':
tmp_cache_fp = 'openai_result_caches/{}_{}_{}'.format(args.dataset, args.inference_split,
args.limit_dataset_size)
logger.info('tmp_cache_fp:{}'.format(tmp_cache_fp))
@cache_results(_cache_fp=tmp_cache_fp)
def tmp_get_response():
pass
responses = call_openai_multi_thread(x_list_to_decode, [hyper_parameter], args.multi_thread, 1,
args.turbo_system_message
)
return responses
responses = tmp_get_response()
else:
responses = call_openai_multi_thread(x_list_to_decode, [hyper_parameter], args.multi_thread, 1,
args.turbo_system_message
)
# args.multi_thread_api = 1
# inp_manager = OpenAI_API_inp_Manager_MultiThread(idx_x_list_to_decode)
#
# thread_list = []
# manager = get_account_manager(1)
# pbar = tqdm.tqdm(total=len(idx_x_list_to_decode))
# for i in range(args.multi_thread):
# thread_list.append(MyThread(i, args, max_length, manager, 1, pbar, inp_manager))
#
# for t in thread_list:
# t.start()
#
# for i, t in enumerate(thread_list):
# t.join()
# logger.info('thread {} finish'.format(t.thread_id))
#
# responses_with_idx = []
#
# for t in thread_list:
# responses_with_idx.extend(t.responses_with_idx)
#
# responses_with_idx.sort(key=lambda x: x[0])
assert (len(datas) == len(x_list_to_decode))
assert len(responses) == len(x_list_to_decode)
responses_with_idx = list(enumerate(responses))
logger.info('responses_with_idx: [{}]'.format(len(responses_with_idx)))
correct_list_for_every_demo_correct_p = {}
for i in range(5):
correct_list_for_every_demo_correct_p[i] = []
with open(args.output_dir, "a") as wp:
logger.info('output_num:{}'.format(len(list(zip(enumerate(datas), idx_x_list_to_decode,
responses_with_idx)))))
for (idx_1, data), (idx_2, inp), (idx_3, response) in zip(enumerate(datas), idx_x_list_to_decode,
responses_with_idx):
output_line = {}
# print('*************************')
# print("{}st data".format(idx_1 + 1))
x, y = data
if args.dataset in nli_dataset:
x = x + "\n" + "A:"
else:
x = "Q: " + x + "\n" + "A:"
y = y
if type(y) is str:
y = y.strip()
# print(x, y)
output_line["question"] = x
output_line["gold_ans"] = y
max_length = args.max_length_cot if "cot" in args.method else args.max_length_direct
# tqdm_bar.update(1)
# second_now = time.time()
# second_gap = second_now - second_start
# second_every_example = second_gap / (i+1)
# logger.info()
# output_line["rationale"] = z
if len(response['choices']) == 1:
# Answer extraction for zero-shot-cot ...
if 'turbo' in args.model:
z = response['choices'][0]['message']['content']
else:
z = response['choices'][0]['text']
if args.method == "zero_shot_cot":
z2 = x + z + " " + args.direct_answer_trigger_for_zeroshot_cot
max_length = args.max_length_direct
pred = decoder.decode(args, z2, max_length)
# print(z2 + pred)
else:
pred = z
# print(inp + pred)
# Clensing of predicted answer ...
if args.method == 'lm_retrieval_few_shot_cot_but_no_thinking':
pred = pred.split('.')[0]
pred = answer_cleansing(args, pred)
elif len(response['choices']) > 1:
preds_counter = Counter()
response_set = set()
if args.method == "zero_shot_cot":
for j, r in enumerate(response['choices']):
content = r['message']['content'] if 'turbo' in args.model else r['text']
response_set.add(content)
r_2_input = x + content + args.direct_answer_trigger_for_zeroshot_cot
tmp_pred = decoder.decode(args, r_2_input, max_length)
tmp_pred = answer_cleansing(args, tmp_pred, verbose=False)
preds_counter[tmp_pred] += 1
r['tmp_pred'] = tmp_pred
else:
for j, r in enumerate(response['choices']):
content = r['message']['content'] if 'turbo' in args.model else r['text']
if args.method == 'lm_retrieval_few_shot_cot_but_no_thinking':
content = content.split('.')[0]
response_set.add(content)
tmp_pred = answer_cleansing(args, content, verbose=False)
r['tmp_pred'] = tmp_pred
preds_counter[tmp_pred] += 1
pred = sorted(list(preds_counter.items()), key=lambda x: x[1], reverse=True)[0][0]
output_line['pred_set_size'] = len(preds_counter)
output_line['pred_count'] = sorted(list(preds_counter.items()), key=lambda x: x[1], reverse=True)
output_line['response_set_size'] = len(response_set)
print('response_set_size : {}'.format(len(response_set)))
print('pred_counter : {}'.format(preds_counter))
output_line['response'] = response
# print(r_2_input + pred)
output_line["pred_ans"] = pred
output_line["wrap_question"] = inp
output_json = json.dumps(output_line)
wp.write(output_json + '\n')
# Choose the most frequent answer from the list ...
print("pred : {}".format(pred))
print("GT : ", y)
print('*************************')
# Checking answer ...
if args.dataset == 'drop':
pass
# tmp_exact_match, tmp_f1 = (pred, y)
tmp_exact_match_single_gold, tmp_f1_single_gold = pred_to_one_answer_f1_metrics(pred, y[0],
numerically_strict=1)
tmp_exact_match_multiple_gold, tmp_f1_multiple_gold = pred_to_many_f1_metrics(pred, y,
numerically_strict=1)
em_list_single_gold.append(tmp_exact_match_single_gold)
f1_list_single_gold.append(tmp_f1_single_gold)
em_list_multiple_gold.append(tmp_exact_match_multiple_gold)
f1_list_multiple_gold.append(tmp_f1_multiple_gold)
# correct_list.append(tmp_exact_match)
# f1_list.append(tmp_f1)
total = idx_1 + 1
em_single_gold = (sum(em_list_single_gold) * 1.0 / total) * 100
f1_single_gold = (sum(f1_list_single_gold) * 1.0 / total) * 100
em_multiple_gold = (sum(em_list_multiple_gold) * 1.0 / total) * 100
f1_multiple_gold = (sum(f1_list_multiple_gold) * 1.0 / total) * 100
print('{}/{} exact match single gold: {}'.format(total, total_example_number, em_single_gold))
print('{}/{} f1 single gold: {}'.format(total, total_example_number, f1_single_gold))
print('')
print('{}/{} exact match multiple gold: {}'.format(total, total_example_number, em_multiple_gold))
print('{}/{} f1 multiple gold: {}'.format(total, total_example_number, f1_multiple_gold))
# exact_match_acc = (sum(correct_list) * 1.0 / total) * 100
# total_f1 = (sum(f1_list) * 1.0 / total) * 100
# print("{}/{} exact match : {}".format(total, total_example_number, exact_match_acc))
# print("{}/{} exact match : {}".format(total, total_example_number, total_f1))
elif args.dataset in ['hotpot_qa', 'qa_wikidata']:
tmp_exact_match_single_gold, tmp_f1_single_gold = pred_to_one_answer_f1_metrics(pred, y,
numerically_strict=1)
em_list_single_gold.append(tmp_exact_match_single_gold)
f1_list_single_gold.append(tmp_f1_single_gold)
total = idx_1 + 1
em_single_gold = (sum(em_list_single_gold) * 1.0 / total) * 100
f1_single_gold = (sum(f1_list_single_gold) * 1.0 / total) * 100
em_multiple_gold = (sum(em_list_multiple_gold) * 1.0 / total) * 100
f1_multiple_gold = (sum(f1_list_multiple_gold) * 1.0 / total) * 100
print('{}/{} exact match single gold: {}'.format(total, total_example_number, em_single_gold))
print('{}/{} f1 single gold: {}'.format(total, total_example_number, f1_single_gold))
else:
correct = (np.array([pred]) == np.array([y])).sum().item()
correct_list.append(correct)
correct_list_for_every_demo_correct_p[demos_correct_p_for_every_x[idx_1]].append(correct)
total += 1 # np.array([y]).size(0)
accuracy = (sum(correct_list) * 1.0 / total) * 100
print("{}/{} accuracy : {}".format(total, total_example_number, accuracy))
for j in range(5):
tmp_correct_list = correct_list_for_every_demo_correct_p[j]
print('demo_correct_p: {}, acc: {}'.format(j, (sum(tmp_correct_list) * 1.0 / (len(tmp_correct_list) if len(tmp_correct_list) > 0 else -1)) * 100))
tmp_correct_list = correct_list
print('demo_correct_p: {}, acc: {}'.format('all', (sum(tmp_correct_list) * 1.0 / len(tmp_correct_list)) * 100))
logger.info('demos_correct_p_for_every_x count:{}'.format(Counter(demos_correct_p_for_every_x)))
# logger.info('args.limit_dataset_size:{}'.format(args.limit_dataset_size))
# logger.info('(args.limit_dataset_size != 0) and ((i + 1) >= args.limit_dataset_size):{}'.format(
# (args.limit_dataset_size != 0) and ((i + 1) >= args.limit_dataset_size)))
# raise ValueError("Stop !!")
# Calculate accuracy ...
if args.dataset == 'drop':
pass
em_single_gold = (sum(em_list_single_gold) * 1.0 / total) * 100
f1_single_gold = (sum(f1_list_single_gold) * 1.0 / total) * 100
em_multiple_gold = (sum(em_list_multiple_gold) * 1.0 / total) * 100
f1_multiple_gold = (sum(f1_list_multiple_gold) * 1.0 / total) * 100
print('{}/{} exact match single gold: {}'.format(total, total_example_number, em_single_gold))
print('{}/{} f1 single gold: {}'.format(total, total_example_number, f1_single_gold))
print('')
print('{}/{} exact match multiple gold: {}'.format(total, total_example_number, em_multiple_gold))
print('{}/{} f1 multiple gold: {}'.format(total, total_example_number, f1_multiple_gold))
fitlog.add_best_metric({'em_s': em_single_gold})
fitlog.add_best_metric({'em_m': em_multiple_gold})
fitlog.add_best_metric({'f1_s': f1_single_gold})
fitlog.add_best_metric({'f1_m': f1_multiple_gold})
elif args.dataset in ['hotpot_qa', 'qa_wikidata']:
em_single_gold = (sum(em_list_single_gold) * 1.0 / total) * 100
f1_single_gold = (sum(f1_list_single_gold) * 1.0 / total) * 100
print('{}/{} exact match single gold: {}'.format(total, total_example_number, em_single_gold))
print('{}/{} f1 single gold: {}'.format(total, total_example_number, f1_single_gold))
fitlog.add_best_metric({'em_s': em_single_gold})
fitlog.add_best_metric({'f1_s': f1_single_gold})
else:
accuracy = (sum(correct_list) * 1.0 / total) * 100
print("accuracy : {}".format(accuracy))
fitlog.add_best_metric({'test_acc': accuracy})
fitlog.add_best_metric({'tmp': 2})
def parse_arguments():
parser = argparse.ArgumentParser(description="Zero-shot-CoT")
parser.add_argument('--filter_no_trigger', default=-1)
parser.add_argument('--demo_c', default=-1)
parser.add_argument('--entropy_threshold', default=-1)
parser.add_argument('--how_to_divide_demos_for_retrieval', required=True)
parser.add_argument('--lm_format_requirement_at_last', type=int, required=True)
parser.add_argument('--shuffle_demos_for_lm_retrieval', type=int, required=True)
parser.add_argument('--demos_for_retrieval_using_purely_question', required=True, type=int)
# parser.add_argument('--retrieval_lm_system_message',required=True)
parser.add_argument('--turbo_system_message', required=True)
parser.add_argument('--retrieval_hybrid_with_task_demos', required=True)
parser.add_argument('--do_not_retrieve_same_premise_demos', type=int, required=True)
# whether there are the same premises in retrieved demos
parser.add_argument('--do_not_retrieve_same_premise_demo_with_test', type=int, required=True)
# whether there are the same premise as query example's premise in retrieved demos
parser.add_argument('--limit_account_num', default=-1, type=int)
parser.add_argument('--exp_tag', default='None')
parser.add_argument('--demo_pool_path', )
parser.add_argument('--demo_pool_from', choices=['gt', 'lm_inference'], required=True)
parser.add_argument('--multi_thread', type=int, required=True)
parser.add_argument('--inference_split', required=True, choices=['test', 'train'])
parser.add_argument('--exp_name', required=True)
parser.add_argument('--num_demo', type=int, default=-1)
parser.add_argument('--demo_sampling_seed', type=int, required=True)
parser.add_argument('--retriever_name', )
parser.add_argument('--demo_encoding', choices=['x'], required=True)
parser.add_argument('--query_encoding', choices=['x'], required=True)
parser.add_argument('--clustered_retrieval', type=int, required=True)
parser.add_argument("--random_seed", type=int, default=1, help="random seed")
parser.add_argument(
"--dataset", type=str, default="multiarith",
choices=["aqua", "gsm8k", "commonsensqa", "addsub", "multiarith", "strategyqa", "svamp", "singleeq",
"coin_flip", "last_letters", "strategyqa_small", 'openbookqa', 'anli_a2', 'anli_a3', 'drop',
'elementary_math_qa', 'boolq', 'fact_checker', 'com_v', 'com_e', 'anli_a1', 'hotpot_qa',
'qa_wikidata'],
help="dataset used for experiment"
)
parser.add_argument(
"--demo_path", type=str, default="demos/multiarith", help="pre-generated demos used for experiment"
)