-
Notifications
You must be signed in to change notification settings - Fork 0
/
run.py
1948 lines (1730 loc) · 100 KB
/
run.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
#!/usr/bin/env python3
#-*- coding: utf-8 -*-
"""
@file: run.py
@author: ImKe at 2022/6/11
@email: tuisaac163@gmail.com
@feature: #Enter features here
"""
import datetime, math, os, sys, json, argparse, time, re, copy
import numpy as np
import torch
import torch.nn as nn
from torch.utils.data import DataLoader, SequentialSampler
from tensorboardX import SummaryWriter
sys.path.append("./")
from logger import Logger
from transformers import (
GPT2Tokenizer,
GPT2Config,
GPT2LMHeadModel,
BertTokenizer,
BertConfig,
T5Tokenizer,
T5Config,
AdamW,
get_linear_schedule_with_warmup,
get_cosine_schedule_with_warmup)
from torch.optim import Adam
from tqdm import tqdm
from utils import model_utils
from model.base import GPT2ModelForMultiModal, GPT2LMforMMDialog
from model.base_bert import BertForMMDialog
from model.base_t5 import T5ForMMDialogGeneration
from data_utils.dialog_reader import MMTensorizer, MMDialogDataset
from utils.eval_utils import evaluate_on_wow_multimodal_training
configs = json.load(open("config.json"))
cache_dir = configs['CACHE_DIR']
data_dir = configs['DATA_DIR']
MIN_PPL_THR=10
FIXED_GPT_LEN=250
# episodes, entity_turn, caption_turn, entities_img_feats, turn_img_feats, history_in_context, max_episode_length=1
parser = argparse.ArgumentParser()
## data preparation
parser.add_argument("--dataset", default='WOW', type=str, choices=["WOW", "DD"],
help="Training dataset. WOW for Wizard of Wikipedia, DD for DailyDialog")
parser.add_argument("--episodes_train_file", default='./wow/train.json', type=str)
parser.add_argument("--episodes_val_file", default='./wow/valid_topic_split.json', type=str)
parser.add_argument("--episodes_test_file", default='./wow/test_topic_split.json', type=str)
parser.add_argument("--trainloader_file", default='./wow/processed_data/IP-1turn.json', type=str)
parser.add_argument("--valloader_file", default='./wow/processed_data/IP-valid_topic-1turn.json', type=str)
parser.add_argument("--testloader_file", default='./wow/processed_data/IP-test_topic-1turn.json', type=str)
parser.add_argument("--model_type", default="t5", type=str, choices=['gpt', 't5', 'unilm'])
parser.add_argument("--unilm_cache", default='/usr/.cache/unilm-base', type=str)
parser.add_argument("--t5_model_name", default='t5-base', type=str)
parser.add_argument("--gpt_model_name", default='gpt2-medium',
type=str, choices=['gpt2-medium', 'microsoft/DialoGPT-medium'])
parser.add_argument("--load", default=None, type=str)
parser.add_argument("--history_in_context", action='store_true', help="Use one dialog session as multiple training sample")
parser.add_argument("--max_episode_length", default=50, type=int)
parser.add_argument("--output_dir", default='output/', type=str,
help="The output directory to save checkpoint and test results.")
parser.add_argument("--max_seq_length", default=185, type=int,
help="The maximum total input sequence length after tokenization. "
"Sequences longer than this will be truncated, "
"sequences shorter will be padded.")
parser.add_argument("--max_seq_a_length", default=30, type=int,
help="The maximum sequence length for caption.")
parser.add_argument("--knowledge_len", default=0, type=int,
help="The maximum sequence length for knowledge document for Wizard of Wikipedia dataset.")
parser.add_argument("--do_train", action='store_true', help="Whether to run training.")
parser.add_argument("--do_sample", action='store_true', help="Whether to do sampling.")
parser.add_argument("--do_test", action='store_true', help="Whether to run inference.")
parser.add_argument("--do_eval", action='store_true', help="Whether to run evaluation.")
parser.add_argument("--mask_prob", default=0.0, type=float,
help= "Probability to mask input sentence during training.")
parser.add_argument("--mask_response_prob", default=0.8, type=float,
help= "Probability to mask input sentence during training.")
parser.add_argument("--max_masked_response_tokens", type=int, default=30,
help="The max number of masked tokens of response.")
parser.add_argument("--max_masked_context_tokens", type=int, default=0,
help="The max number of masked tokens of context and caption.")
parser.add_argument("--drop_out", default=0.1, type=float, help="Drop out in the model.")
parser.add_argument("--label_smoothing", default=0, type=float, help="Label smoothing")
parser.add_argument("--alpha", default=0.5, type=float)
parser.add_argument("--beta", default=1.0, type=float)
parser.add_argument("--max_turn_img_length", default=5, type=int,
help="The maximum image number turn level image.")
parser.add_argument("--per_max_turn_img_seq_length", default=10, type=int,
help="The maximum image number turn level image.")
parser.add_argument("--max_ent_img_seq_length", default=8, type=int,
help="The maximum total input entity level image sequence length.")
parser.add_argument("--img_feature_dim", default=512, type=int,
help="The Image Feature Dimension default is 512 for CLIP-base. 1024 for CLIP-large")
parser.add_argument("--type_vocab_size", default=3, type=int,
help="segment embedding vocabulary size.")
parser.add_argument("--img_feature_type", default='clip', type=str,
help="Image feature type.")
parser.add_argument("--img_layer_norm_eps", default=1e-4, type=float,
help="segment embedding vocabulary size.")
parser.add_argument("--img_add_pos", default="concat", type=str,
help="image adding method",
choices=['concat', 'enc_cross_attn', 'dec_cross_attn', 'cross_attn'])
parser.add_argument("--use_img_layernorm", action='store_true')
parser.add_argument('--no_gpu', action='store_true')
parser.add_argument('--gpu', nargs='+', type=int, default=[0])
parser.add_argument("--per_gpu_train_batch_size", default=42, type=int,
help="Batch size per GPU/CPU for training.")
parser.add_argument("--per_gpu_eval_batch_size", default=5, type=int,
help="Batch size per GPU/CPU for evaluation.")
parser.add_argument("--workers", default=0, type=int,
help="Dataloader worker.")
parser.add_argument("--early_stop", default=5, type=int,
help="Early stopping iteration.")
parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
help="Number of updates steps to accumulate before backward.")
parser.add_argument("--learning_rate", default=3e-5, type=float, help="The initial learning rate.")
parser.add_argument("--weight_decay", default=1e-6, type=float, help="Weight decay")
parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam.")
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
parser.add_argument("--warm_up_ratio", default=0.1, type=float)
parser.add_argument("--warmup_steps", default=1000, type=int, help="Linear warmup.")
parser.add_argument("--scheduler", default='linear', type=str, help="linear or consine", choices=["linear", "consine"])
parser.add_argument("--num_workers", default=4, type=int, help="Workers in dataloader.")
parser.add_argument("--num_train_epochs", default=None, type=int,
help="Total number of training epochs to perform.")
parser.add_argument("--vit_version", default="base", type=str, choices=["base", "large"])
parser.add_argument("--iterations", default=30000, type=int,
help="Total number of training iterations to perform.")
parser.add_argument("--log_epoch", default=2, type=int)
parser.add_argument("--test_iter", default=10, type=int)
parser.add_argument("--save_per_test_iter", default=0, type=int)
parser.add_argument("--max_val_batches", default=200, type=int)
parser.add_argument("--max_test_batches", default=200, type=int)
parser.add_argument("--evaluate_during_training", action='store_true',
help="Run evaluation during training at each save_steps.")
parser.add_argument("--evaluate_testset_during_training", action='store_true',
help="Run evaluation testset during training at each save_steps.")
parser.add_argument('--seed', type=int, default=42, help="random seed for initialization.")
parser.add_argument('--ent_img_num', type=int, default=1)
parser.add_argument('--add_response_mlm', action='store_true', help='Add masked response prediction')
parser.add_argument('--add_cap_bias', action='store_true', help='Add caption bias to logits')
parser.add_argument('--add_ent_bias', action='store_true', help='Add entity bias to logits')
parser.add_argument('--add_ent_matching', action='store_true')
parser.add_argument('--add_textual_cap', action='store_true',
help="Add captions if available for training")
parser.add_argument('--add_textual_ent', action='store_true',
help="Add entity texts if available for training")
parser.add_argument('--t5_mlm', action='store_true')
parser.add_argument('--eda', action='store_true', help='Use easy data augmentation for encoder input')
parser.add_argument('--eda_pos', type=str, default='res', choices=['res', 'conx'],
help='eda position for T5 based model, context or response')
parser.add_argument('--no_turn_vis', action='store_true',
help='Train the model without Turn-level Visual Knowledge')
parser.add_argument('--no_ent_vis', action='store_true',
help='Train the model without Entity-level Visual Knowledge')
parser.add_argument('--last_res', action='store_true',
help='Train the model with only the last response of dialogue data')
parser.add_argument('--bleu_eval', action='store_true',
help='Use BLEU-1 as the evaluation metric')
# for generation
parser.add_argument("--eval_model_dir", type=str, default="test_output",
help="Model directory for evaluation.")
parser.add_argument("--evaluate_cache", type=str, default="",
help="Model directory storing ckpt for model testing and evaluation.")
parser.add_argument('--max_gen_length', type=int, default=40,
help="max length of generated sentences")
parser.add_argument('--output_hidden_states', action='store_true',
help="Turn on for fast decoding")
parser.add_argument('--num_return_sequences', type=int, default=1,
help="repeating times per image")
parser.add_argument('--num_beams', type=int, default=5, help="beam search width")
parser.add_argument('--num_keep_best', type=int, default=1,
help="number of hypotheses to keep in beam search")
parser.add_argument('--temperature', type=float, default=1,
help="temperature in softmax for sampling")
parser.add_argument('--top_k', type=int, default=10,
help="filter distribution for sampling")
parser.add_argument('--top_p', type=float, default=0.10,
help="filter distribution for sampling")
parser.add_argument('--repetition_penalty', type=float, default=3.0,
help="repetition penalty from CTRL paper (https://arxiv.org/abs/1909.05858)")
parser.add_argument('--length_penalty', type=float, default=0.3,
help="beam search length penalty")
parser.add_argument('--add_cap_labels', action='store_true',
help="Add captions if available for generation")
parser.add_argument('--add_ent_labels', action='store_true',
help="Add entity texts if available for generation")
parser.add_argument('--dp', action='store_true')
# for Constrained Beam Search
parser.add_argument('--use_cbs', action='store_true',
help='Use constrained beam search for decoding')
parser.add_argument('--min_constraints_to_satisfy', type=int, default=2,
help="minimum number of constraints to satisfy")
def test(args, test_dataloader, gen_dataloader, model, tokenizer, predict_file,
logging, max_test_batches, generate, device):
model.eval()
test_step = min(len(test_dataloader), max_test_batches)
gen_step = min(len(gen_dataloader), max_test_batches)
with torch.no_grad():
val_losses, val_masked_token_losses, val_masked_res_losses, \
val_ent_matching_losses = [], [], [], []
for step, data_dict in tqdm(enumerate(test_dataloader),
ncols=100, desc="Testing Forward: ",
total=test_step):
img_attention_mask = data_dict["attention_mask"][1] if args.img_add_pos == "cross_attn" else None
## remove padding tokens
masked_ids = data_dict["masked_ids"]
masked_ids = masked_ids[masked_ids != tokenizer.pad_token_id]
if args.add_response_mlm:
masked_res = data_dict["masked_res"][:, 1:] # remove BOS id at the beginning
masked_res = masked_res[masked_res != tokenizer.pad_token_id]
input_ids_rmask = data_dict["input_ids_rmask"]
else:
masked_res = None
input_ids_rmask = None
inputs = {"input_ids": data_dict["input_ids"],
"input_ids_rmask": input_ids_rmask,
"turn_img_feats": data_dict["turn_img_feats"],
"ent_img_feats": data_dict["ent_img_feats"],
"attention_mask": data_dict["attention_mask"][0],
"img_attention_mask": img_attention_mask,
"token_type_ids": data_dict["segment_ids"],
"masked_ids": masked_ids,
"masked_pos": data_dict["masked_pos"],
"masked_res": masked_res,
"masked_res_pos": data_dict["masked_res_pos"],
"add_ent_bias": args.add_ent_bias,
"add_cap_bias": args.add_cap_bias,
"cap_pos": data_dict["cap_pos"],
"ent_pos": data_dict["ent_pos"],
"add_ent_matching": args.add_ent_matching,
"ent_img_pos": data_dict["ent_img_pos"],
"ent_ids": data_dict["ent_ids"] if (not args.no_ent_vis and args.add_textual_ent) else None,
}
model_output = model(**inputs)
masked_token_loss, masked_res_loss, ent_matching_loss, class_logits, batch_masked_score, \
batch_masked_res_score, batch_ent_score = model_output[:7]
loss = args.alpha * masked_token_loss + (1 - args.alpha) * masked_res_loss + args.beta * ent_matching_loss
val_losses.append(loss.item())
val_masked_token_losses.append(masked_token_loss.item())
val_masked_res_losses.append(masked_res_loss.item())
val_ent_matching_losses.append(ent_matching_loss.item())
def gen_rows():
time_meter = 0
with torch.no_grad():
for step, batch in tqdm(enumerate(gen_dataloader),
ncols=100, desc="Testing Forward: ",
total=gen_step):
# batch = tuple(t.to(device) for t in batch)
img_attention_mask = batch["attention_mask"][1] if args.img_add_pos == "cross_attn" else None
inputs = {'is_decode': True,
'input_ids': batch['input_ids'],
'attention_mask': batch['attention_mask'][0],
'token_type_ids': batch['segment_ids'],
'masked_pos': batch['masked_pos'],
'turn_img_feats': batch['turn_img_feats'],
'ent_img_feats': batch['ent_img_feats'],
'img_attention_mask': img_attention_mask,
'gth_ids': batch['gth_ids'],
'do_sample': args.do_sample,
'bos_token_id': tokenizer.bos_token_id,
'pad_token_id': tokenizer.pad_token_id,
'eos_token_ids': tokenizer.eos_token_id,
'mask_token_id': tokenizer.mask_token_id,
# for adding entity labels
'add_ent_labels': args.add_ent_labels,
'add_cap_labels': args.add_cap_labels,
## since we separately pad response, context, caption, entity tokens
## their starting positions are the same across the same batch
'ent_labels_start_posid': batch['ent_labels_start_posid'][0],
'cap_start_posid': batch['cap_start_posid'][0],
'context_start_posid': batch['context_start_posid'][0],
'context_end_posid': batch['context_end_posid'][0],
# hyperparameters of beam search
'max_length': args.max_seq_a_length,
'num_beams': args.num_beams,
"temperature": args.temperature,
"top_k": args.top_k,
"top_p": args.top_p,
"repetition_penalty": args.repetition_penalty,
"length_penalty": args.length_penalty,
"num_return_sequences": args.num_return_sequences,
"num_keep_best": args.num_keep_best,
}
if args.use_cbs:
inputs.update({'use_cbs': True,
'fsm': batch[5],
'num_constraints': batch[6],
'min_constraints_to_satisfy': args.min_constraints_to_satisfy,
})
tic = time.time()
# captions, logprobs
outputs = model(**inputs)
time_meter += time.time() - tic
all_caps = outputs[0] # [batch_size, num_keep_best, max_len]
all_confs = torch.exp(outputs[1])
all_responses = batch["response"]
for caps, confs, response in zip(all_caps, all_confs, all_responses):
res = []
for cap, conf in zip(caps, confs):
cap = tokenizer.decode(cap.tolist(), skip_special_tokens=True)
cap = cap.split('.')[0]
res.append({'response_pred': cap, 'response': response, 'conf': conf.item()})
yield res
# return res
if step >= test_step:
break
logging.info("Inference model computing time: {} seconds per batch".format(time_meter / (step + 1)))
if generate:
model_utils.tsv_writer(gen_rows(), predict_file)
logging.info("========================= Validate Status =========================")
logging.info("Avg. Val Loss : {:.4f}".format(np.mean(val_losses)))
logging.info("Avg. Val Masked LM Loss : {:.4f}".format(np.mean(val_masked_token_losses)))
logging.info("Avg. Val Masked PPL : {:.4f}".format(math.exp(min(np.mean(val_masked_token_losses), MIN_PPL_THR))))
logging.info("Avg. Val Masked Res Loss : {:.4f}".format(np.mean(val_masked_res_losses)))
logging.info("Avg. Val Ent Matching : {:.4f}".format(np.mean(val_ent_matching_losses)))
return np.mean(val_losses), np.mean(val_masked_token_losses), np.mean(val_masked_res_losses), np.mean(val_ent_matching_losses)
def train_UniLM(args):
now = datetime.datetime.now()
# GPU
# if not torch.cuda.is_available(): args.no_gpu = True
gpu = not args.no_gpu
if gpu:
print("There are ", torch.cuda.device_count(), " available GPUs!")
print('Current single GPU: {}'.format(torch.cuda.current_device()))
args.n_gpu = len(args.gpu)
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
else:
args.train_batch_size = args.per_gpu_train_batch_size
args.eval_batch_size = args.per_gpu_eval_batch_size
device = torch.device(args.gpu[0] if gpu else "cpu")
# randomness
np.random.seed(args.seed)
prng = np.random.RandomState()
torch.random.manual_seed(args.seed)
if gpu: torch.cuda.manual_seed(args.seed); torch.cuda.manual_seed_all(args.seed)
loading = True if args.load is not None else False
if args.no_ent_vis and args.no_turn_vis:
ex_appended = f"novis_{now.month}.{now.day}"
elif args.no_ent_vis:
ex_appended = f"no_evis_{now.month}.{now.day}"
elif args.no_turn_vis:
ex_appended = f"no_tvis_{now.month}.{now.day}"
else:
ex_appended = f"{now.month}.{now.day}"
if args.last_res:
ex_appended = f"lastres_{ex_appended}"
experiment = f"train_unilm_{args.dataset}_{args.img_add_pos}_e{args.num_train_epochs}" \
f"_res-mlm-{args.add_response_mlm}_Ebiad-{args.add_ent_bias}_Ematch-{args.add_ent_matching}" \
f"_lr{args.learning_rate}_seqlen{args.max_seq_length}_bs{args.per_gpu_train_batch_size}_" \
f"load{loading}_cap{int(args.add_textual_cap)}_ent{int(args.add_textual_ent)}" \
f"_ep{args.max_episode_length}_{ex_appended}"
save_folder = os.path.join(args.output_dir, experiment)
os.makedirs(os.path.join(save_folder, 'ckpt/model'), exist_ok=True)
t_writer = SummaryWriter(os.path.join(save_folder, 'train'), flush_secs=5)
# v_writer = SummaryWriter(os.path.join(save_folder, 'val'), flush_secs=5)
logging_file = f"train_unilm_{now.month}.{now.day}.log"
logging = Logger(os.path.join(save_folder, logging_file))
logging.info('\n*******************************************************************************\n')
logging.info("the configuration:")
logging.info(str(args).replace(',', '\n'))
logging.info('Loading models...')
## prepare model & resize tokenizer
config = BertConfig.from_pretrained(args.unilm_cache)
config = model_utils.sort_config(config, args)
tokenizer = BertTokenizer.from_pretrained(args.unilm_cache)
state = torch.load(os.path.join(args.unilm_cache, "pytorch_model.bin"))
state = model_utils.rearrange_unilm_weights(state)
model = BertForMMDialog(config, args.img_add_pos)
model.load_state_dict(state, strict=False)
print("Finish Loading Pre-trained UniLM Model Weight")
del state
tokenizer.bos_token = tokenizer.cls_token
tokenizer.eos_token = tokenizer.sep_token
# special_tokens_dict = {'bos_token': '[CLS]', 'eos_token': '[SEP]',
# 'additional_special_tokens': ['[CONSEP]', '[CAPSEP]']}
# num_added_toks = tokenizer.add_special_tokens(special_tokens_dict)
# logging.info(f'We have added {num_added_toks} tokens to the model..')
# model._resize_token_embeddings(len(tokenizer))
# model.bert._resize_token_embeddings(len(tokenizer))
# model.resize_token_embeddings(len(tokenizer))
bert_embedding_weight = model.bert.embeddings.word_embeddings.weight
# model.decoder = nn.Linear(bert_embedding_weight.size(1),
# bert_embedding_weight.size(0), bias=False)
model.vocab_size = bert_embedding_weight.size(0)
model.build_smoothing_loss()
if args.load is not None:
print("Loading Trained Model Weight...")
state = torch.load(os.path.join("output", args.load, "model_best_val.pt"))
model.load_state_dict(state)
del state
model.to(device)
model.train()
# model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
# if len(args.gpu) > 1:
# model = nn.DataParallel(model)
p_train_data = model_utils.jsonload(args.trainloader_file)[0]
p_val_data = model_utils.jsonload(args.valloader_file)[0]
if args.dataset == "WOW":
dialog_episodes = json.load(open(args.episodes_train_file))
val_episodes = json.load(open(args.episodes_val_file))
elif args.dataset == "DD":
dialog_episodes = model_utils.jsonload(args.episodes_train_file)
val_episodes = model_utils.jsonload(args.episodes_val_file)
else:
raise NotImplementedError
fea_prefix_path = os.path.join(data_dir, f"processed_resee_data/shared/processed_img_features_clip_{args.vit_version}/")
fea_paths = ["openimagev6_train_clip_vis_fea.pt",
"openimagev6_test_clip_vis_fea.pt",
"openimagev6_val_clip_vis_fea.pt",
"coco_train_clip_vis_fea.pt",
"coco_val_clip_vis_fea.pt",
"flickr30_clip_vis_fea.pt",
"nocaps_clip_vis_fea.pt"]
fea_paths = [os.path.join(fea_prefix_path, item) for item in fea_paths]
turn_img_feats = model_utils.load_pt_list(fea_paths)
if args.dataset == "WOW":
ent_clip_fea_file = "wow/processed_img_features/img_clip_features.pt"
elif args.dataset == "DD":
ent_clip_fea_file = "dd/processed_img_features/img_clip_features.pt"
else:
ent_clip_fea_file = None
ent_img_feats = model_utils.load_pt_list([os.path.join(data_dir, "processed_resee_data", ent_clip_fea_file)])
random_token_list = model_utils.gen_random_word_list(tokenizer, config.vocab_size)
train_data = MMDialogDataset(dialog_episodes, device,
p_train_data["entity_turn"],
p_train_data["caption_turn"],
p_train_data["entities_img_feats_ids"][0],
p_train_data["turn_img_feats_ids"],
turn_img_feats=turn_img_feats, ent_img_feats=ent_img_feats,
img_add_pos=args.img_add_pos, img_dim=args.img_feature_dim,
history_in_context=args.history_in_context,
max_episode_length=args.max_episode_length, tokenizer=tokenizer,
max_turn_img_seq_length=args.max_turn_img_length,
per_max_turn_img_seq_length=args.per_max_turn_img_seq_length,
max_ent_img_seq_length=args.max_ent_img_seq_length,
max_seq_length=args.max_seq_length,
max_seq_a_length=args.max_seq_a_length, is_train=args.do_train,
mask_prob=args.mask_prob, mask_response_prob=args.mask_response_prob,
response_sep="[SEP_0]", context_sep="[SEP_1]", caption_sep="[SEP_2]",
random_token_list=random_token_list, test=args.last_res,
max_masked_response_tokens=args.max_masked_response_tokens,
max_masked_context_tokens=args.max_masked_context_tokens,
dataset=args.dataset, add_cap=args.add_textual_cap,
add_ent=args.add_textual_ent, no_ent_vis=args.no_ent_vis,
no_turn_vis=args.no_turn_vis)
val_gen_data = MMDialogDataset(val_episodes, device,
p_val_data["entity_turn"],
p_val_data["caption_turn"],
p_val_data["entities_img_feats_ids"][0],
p_val_data["turn_img_feats_ids"],
turn_img_feats=turn_img_feats, ent_img_feats=ent_img_feats,
img_add_pos=args.img_add_pos, img_dim=args.img_feature_dim,
history_in_context=args.history_in_context,
max_episode_length=args.max_episode_length, tokenizer=tokenizer,
max_turn_img_seq_length=args.max_turn_img_length,
per_max_turn_img_seq_length=args.per_max_turn_img_seq_length,
max_ent_img_seq_length=args.max_ent_img_seq_length,
max_seq_length=args.max_seq_length,
max_seq_a_length=args.max_seq_a_length, is_train=False,
response_sep="[SEP_0]", context_sep="[SEP_1]", caption_sep="[SEP_2]",
test=True, dataset=args.dataset,
add_cap=False if not args.add_cap_labels else args.add_textual_cap,
add_ent=False if not args.add_ent_labels else args.add_textual_ent,
no_ent_vis=args.no_ent_vis,
no_turn_vis=args.no_turn_vis)
val_data = MMDialogDataset(val_episodes, device,
p_val_data["entity_turn"],
p_val_data["caption_turn"],
p_val_data["entities_img_feats_ids"][0],
p_val_data["turn_img_feats_ids"],
turn_img_feats=turn_img_feats, ent_img_feats=ent_img_feats,
img_add_pos=args.img_add_pos, img_dim=args.img_feature_dim,
history_in_context=args.history_in_context,
max_episode_length=args.max_episode_length, tokenizer=tokenizer,
max_turn_img_seq_length=args.max_turn_img_length,
per_max_turn_img_seq_length=args.per_max_turn_img_seq_length,
max_ent_img_seq_length=args.max_ent_img_seq_length,
max_seq_length=args.max_seq_length,
max_seq_a_length=args.max_seq_a_length, is_train=True,
response_sep="[SEP_0]", context_sep="[SEP_1]", caption_sep="[SEP_2]",
random_token_list=random_token_list,
test=True, dataset=args.dataset,
add_cap=args.add_textual_cap,
add_ent=args.add_textual_ent,
no_ent_vis=args.no_ent_vis,
no_turn_vis=args.no_turn_vis)
train_loader = DataLoader(train_data,
batch_size=args.train_batch_size,
pin_memory=False,
drop_last=False,
num_workers=args.workers,
shuffle=True)
val_sampler = SequentialSampler(val_data)
val_gen_loader = DataLoader(val_gen_data,
sampler=val_sampler,
batch_size=args.eval_batch_size,
pin_memory=False,
drop_last=False,
num_workers=args.workers)
val_loader = DataLoader(val_data,
sampler=val_sampler,
batch_size=args.eval_batch_size,
pin_memory=False,
drop_last=False,
num_workers=args.workers)
optimizer = AdamW(model.parameters(), lr=args.learning_rate, correct_bias=True, weight_decay=args.weight_decay)
if args.num_train_epochs is None:
total_step = args.iterations
else:
total_step = len(train_loader) * args.num_train_epochs
logging.info(f"Ignoring assigned iterations, total step is {total_step}..")
if args.scheduler == "linear":
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warm_up_ratio * total_step,
num_training_steps=total_step)
elif args.scheduler == "cosine":
scheduler = get_cosine_schedule_with_warmup(optimizer, num_warmup_steps=args.warm_up_ratio * total_step,
num_training_steps=total_step)
else:
raise NotImplementedError
logging.info("Begin training steps")
max_val_batches = args.max_val_batches # max num. of val batches
logging.info("Total step: %d" % total_step)
e = 0 # number of epoch
num_iters = 0
best_loss = 9e9
et = 0
eval_step = 0
optimizer.zero_grad()
scaler = torch.cuda.amp.GradScaler()
while (num_iters < total_step) and (et < args.early_stop):
# Run epoch
st = time.time()
# Training
print('Training loop. Batches:', len(train_loader))
logging.info('\n===========================================================================')
logging.info("Training loop. Batches: %d" % len(train_loader))
## record for every epoch losses
losses, masked_token_losses, masked_res_losses, ent_matching_losses = [], [], [], []
batch_masked_scores, batch_ent_scores, batch_masked_res_scores = [], [], []
for i, data_dict in enumerate(tqdm(train_loader)):
optimizer.zero_grad()
img_attention_mask = data_dict["attention_mask"][1] if args.img_add_pos == "cross_attn" else None
## remove padding tokens
masked_ids = data_dict["masked_ids"]
masked_ids = masked_ids[masked_ids != tokenizer.pad_token_id]
if args.add_response_mlm:
masked_res = data_dict["masked_res"][:, 1:] # remove BOS id at the beginning
masked_res = masked_res[masked_res != tokenizer.pad_token_id]
input_ids_rmask = data_dict["input_ids_rmask"]
else:
masked_res = None
input_ids_rmask = None
inputs = {"input_ids": data_dict["input_ids"],
"input_ids_rmask": input_ids_rmask,
"turn_img_feats": data_dict["turn_img_feats"],
"ent_img_feats": data_dict["ent_img_feats"],
"attention_mask": data_dict["attention_mask"][0],
"img_attention_mask": img_attention_mask,
"token_type_ids": data_dict["segment_ids"],
"masked_ids": masked_ids,
"masked_pos": data_dict["masked_pos"],
"masked_res": masked_res,
"masked_res_pos": data_dict["masked_res_pos"],
"add_ent_bias": args.add_ent_bias,
"add_cap_bias": args.add_cap_bias,
"cap_pos": data_dict["cap_pos"],
"ent_pos": data_dict["ent_pos"],
"add_ent_matching": args.add_ent_matching,
"ent_img_pos": data_dict["ent_img_pos"],
"ent_ids": data_dict["ent_ids"] if not args.no_ent_vis else None,
}
with torch.cuda.amp.autocast():
model_output = model(**inputs)
masked_token_loss, masked_res_loss, ent_matching_loss, class_logits, batch_masked_score, \
batch_masked_res_score, batch_ent_score = model_output[:7]
loss = args.alpha * masked_token_loss + (1 - args.alpha) * masked_res_loss + args.beta * ent_matching_loss
scaler.scale(loss).backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
scaler.step(optimizer)
scaler.update()
scheduler.step()
# return loss, masked_token_loss, masked_res_token_loss, ent_matching_loss, \
# batch_masked_score, batch_masked_res_score, batch_ent_score
# loss, masked_token_loss, masked_res_loss, ent_matching_loss, batch_masked_score, \
# batch_masked_res_score, batch_ent_score = unilm_train_step(model, optimizer, scaler,
# data_dict, tokenizer.pad_token_id, args)
if args.gpu and args.n_gpu > 1:
loss = loss.mean()
masked_token_loss = masked_token_loss.mean()
masked_res_loss = masked_res_loss.mean()
ent_matching_loss = ent_matching_loss.mean()
losses.append(loss.item())
masked_res_losses.append(masked_res_loss.item())
masked_token_losses.append(masked_token_loss.item())
ent_matching_losses.append(ent_matching_loss.item())
batch_masked_scores.append(batch_masked_score.item())
batch_ent_scores.append(batch_ent_score.item())
batch_masked_res_scores.append(batch_masked_res_score.item())
lr = scheduler.get_last_lr()[0]
# Log to Tensorboard
t_writer.add_scalar('loss', loss, num_iters)
t_writer.add_scalar('masked_token_loss', masked_token_loss, num_iters)
t_writer.add_scalar('masked_response_loss', masked_res_loss, num_iters)
t_writer.add_scalar('ent_matching_loss', ent_matching_loss, num_iters)
t_writer.add_scalar('masked_score', batch_masked_score, num_iters)
t_writer.add_scalar('ent_score', batch_ent_score, num_iters)
t_writer.add_scalar('masked_res_score', batch_masked_res_score, num_iters)
t_writer.add_scalar('lr', lr, num_iters)
num_iters += 1
if num_iters % args.test_iter == 0:
eval_step += 1
generate = eval_step%50==0
predict_file = os.path.join(save_folder, f"eval_out_{num_iters}.tsv")
curr_val_loss, _, _, _ = test(args, val_loader, val_gen_loader, model, tokenizer, predict_file,
logging, max_val_batches, generate, device)
if curr_val_loss <= best_loss:
best_loss = curr_val_loss
torch.save(model.state_dict(), os.path.join(save_folder, "model_best_val.pt"))
et = 0
else:
et += 1
model.train()
e += 1
logging.info(f"Finish Training for {e} Epochs..")
if e % args.log_epoch == 0:
masked_ppl = math.exp(min(np.mean(masked_token_losses), MIN_PPL_THR))
logging.info("========================= Training Status =========================")
logging.info("Epoch : {}".format(e))
logging.info("Avg. Loss : {:.4f}".format(np.mean(losses)))
logging.info("Avg. Masked Token Loss : {:.4f}".format(np.mean(masked_token_losses)))
logging.info("Avg. Masked PPL : {:.4f}".format(np.mean(masked_ppl)))
logging.info("Avg. Masked Response Loss : {:.4f}".format(np.mean(masked_res_losses)))
logging.info("Avg. Ent Matching Loss : {:.4f}".format(np.mean(ent_matching_losses)))
logging.info("Avg. Masked Acc. : {:.4f}".format(np.mean(batch_masked_scores)))
logging.info("Avg. Ent Acc. : {:.4f}".format(np.mean(batch_ent_scores)))
logging.info("Avg. Masked Response Acc. : {:.4f}".format(np.mean(batch_masked_res_scores)))
if et >= args.early_stop:
logging.info("Early Stopping..")
break
if (args.num_train_epochs is not None) and (e >= args.num_train_epochs):
break
## evaluate at the end of training
# predict_file = os.path.join(save_folder, f"eval_out_{num_iters}.tsv")
# _ = test(args, val_loader, model, tokenizer, predict_file,
# logging, max_val_batches, device)
def test_UniLM(args):
now = datetime.datetime.now()
# GPU
# if not torch.cuda.is_available(): args.no_gpu = True
gpu = not args.no_gpu
if gpu:
print("There are ", torch.cuda.device_count(), " available GPUs!")
print('Current single GPU: {}'.format(torch.cuda.current_device()))
args.n_gpu = len(args.gpu)
args.test_batch_size = args.per_gpu_eval_batch_size
device = torch.device(args.gpu[0] if gpu else "cpu")
# randomness
np.random.seed(args.seed)
prng = np.random.RandomState()
torch.random.manual_seed(args.seed)
if gpu: torch.cuda.manual_seed(args.seed); torch.cuda.manual_seed_all(args.seed)
experiment = "test"
# save_folder = os.path.join(args.eval_model_dir, experiment)
save_folder = os.path.join(args.output_dir, args.evaluate_cache)
# os.makedirs(save_folder, exist_ok=True)
logging_file = f"testing_{now.month}.{now.day}.log"
predict_file = os.path.join(save_folder, f"test_out-best_eval-topk{args.top_k}-topp{args.top_p}.tsv")
logging = Logger(os.path.join(save_folder, logging_file))
logging.info('\n*******************************************************************************\n')
logging.info("the configuration:")
logging.info(str(args).replace(',', '\n'))
logging.info('Loading models...')
## prepare model & resize tokenizer
config = BertConfig.from_pretrained(args.unilm_cache)
config = model_utils.sort_config(config, args)
tokenizer = BertTokenizer.from_pretrained(args.unilm_cache)
model = BertForMMDialog(config, args.img_add_pos)
tokenizer.bos_token = tokenizer.cls_token
tokenizer.eos_token = tokenizer.sep_token
bert_embedding_weight = model.bert.embeddings.word_embeddings.weight
model.vocab_size = bert_embedding_weight.size(0)
state = torch.load(os.path.join(args.output_dir, args.evaluate_cache, "model_best_val.pt"))
model.load_state_dict(state)
print("Finish Loading Model Weight")
del state
model.to(device)
p_test_data = model_utils.jsonload(args.testloader_file)[0]
if args.dataset == "WOW":
test_episodes = json.load(open(args.episodes_test_file))
elif args.dataset == "DD":
test_episodes = model_utils.jsonload(args.episodes_test_file)
else:
raise NotImplementedError
fea_prefix_path = os.path.join(data_dir, f"processed_resee_data/shared/processed_img_features_clip_{args.vit_version}/")
fea_paths = ["openimagev6_train_clip_vis_fea.pt",
"openimagev6_test_clip_vis_fea.pt",
"openimagev6_val_clip_vis_fea.pt",
"coco_train_clip_vis_fea.pt",
"coco_val_clip_vis_fea.pt",
"flickr30_clip_vis_fea.pt",
"nocaps_clip_vis_fea.pt"]
fea_paths = [os.path.join(fea_prefix_path, item) for item in fea_paths]
turn_img_feats = model_utils.load_pt_list(fea_paths)
if args.dataset == "WOW":
ent_clip_fea_file = "wow/processed_img_features/img_clip_features.pt"
elif args.dataset == "DD":
ent_clip_fea_file = "dd/processed_img_features/img_clip_features.pt"
else:
ent_clip_fea_file = None
ent_img_feats = model_utils.load_pt_list([os.path.join(data_dir, "processed_resee_data", ent_clip_fea_file)])
random_token_list = model_utils.gen_random_word_list(tokenizer, config.vocab_size)
test_data = MMDialogDataset(test_episodes, device,
p_test_data["entity_turn"],
p_test_data["caption_turn"],
p_test_data["entities_img_feats_ids"][0],
p_test_data["turn_img_feats_ids"],
turn_img_feats=turn_img_feats, ent_img_feats=ent_img_feats,
img_add_pos=args.img_add_pos, img_dim=args.img_feature_dim,
history_in_context=args.history_in_context,
max_episode_length=args.max_episode_length, tokenizer=tokenizer,
max_turn_img_seq_length=args.max_turn_img_length,
per_max_turn_img_seq_length=args.per_max_turn_img_seq_length,
max_ent_img_seq_length=args.max_ent_img_seq_length,
max_seq_length=args.max_seq_length,
max_seq_a_length=args.max_seq_a_length, is_train=True,
response_sep="[SEP_0]", context_sep="[SEP_1]", caption_sep="[SEP_2]",
random_token_list=random_token_list,
test=True, dataset=args.dataset, add_ent=args.add_textual_ent,
add_cap=args.add_textual_cap,
no_ent_vis=args.no_ent_vis,
no_turn_vis=args.no_turn_vis)
test_gen_data = MMDialogDataset(test_episodes, device,
p_test_data["entity_turn"],
p_test_data["caption_turn"],
p_test_data["entities_img_feats_ids"][0],
p_test_data["turn_img_feats_ids"],
turn_img_feats=turn_img_feats, ent_img_feats=ent_img_feats,
img_add_pos=args.img_add_pos, img_dim=args.img_feature_dim,
history_in_context=args.history_in_context,
max_episode_length=args.max_episode_length, tokenizer=tokenizer,
max_turn_img_seq_length=args.max_turn_img_length,
per_max_turn_img_seq_length=args.per_max_turn_img_seq_length,
max_ent_img_seq_length=args.max_ent_img_seq_length,
max_seq_length=args.max_seq_length,
max_seq_a_length=args.max_seq_a_length, is_train=False,
response_sep="[SEP_0]", context_sep="[SEP_1]", caption_sep="[SEP_2]",
test=True, dataset=args.dataset, add_ent=args.add_textual_ent,
add_cap=args.add_textual_cap,
no_ent_vis=args.no_ent_vis,
no_turn_vis=args.no_turn_vis)
test_sampler = SequentialSampler(test_data)
test_loader = DataLoader(test_data,
sampler=test_sampler,
batch_size=args.test_batch_size,
pin_memory=False,
drop_last=False,
num_workers=args.workers)
gen_loader = DataLoader(test_gen_data,
sampler=test_sampler,
batch_size=args.test_batch_size,
pin_memory=False,
drop_last=False,
num_workers=args.workers)
logging.info("Begin testing..")
test(args, test_loader, gen_loader, model, tokenizer, predict_file, logging, args.max_val_batches, True, device)
def train_T5(args):
now = datetime.datetime.now()
# GPU
# if not torch.cuda.is_available(): args.no_gpu = True
gpu = not args.no_gpu
if gpu:
print("There are ", torch.cuda.device_count(), " available GPUs!")
print('Current single GPU: {}'.format(torch.cuda.current_device()))
args.n_gpu = len(args.gpu)
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
else:
args.train_batch_size = args.per_gpu_train_batch_size
args.eval_batch_size = args.per_gpu_eval_batch_size
device = torch.device(args.gpu[0] if gpu else "cpu")
# randomness
np.random.seed(args.seed)
prng = np.random.RandomState()
torch.random.manual_seed(args.seed)
if gpu: torch.cuda.manual_seed(args.seed); torch.cuda.manual_seed_all(args.seed)
loading = True if args.load is not None else False
t5_version = args.t5_model_name
ex_t5 = t5_version.replace("/", "-")
if args.no_ent_vis and args.no_turn_vis:
ex_appended = f"novis_{now.month}.{now.day}"
elif args.no_ent_vis:
ex_appended = f"no_evis_{now.month}.{now.day}"
elif args.no_turn_vis:
ex_appended = f"no_tvis_{now.month}.{now.day}"
else:
ex_appended = f"{now.month}.{now.day}"
if args.last_res:
ex_appended = f"lastres_{ex_appended}"
eda_append = int(args.eda) if not args.eda else f"-{args.eda_pos}"
experiment = f"train_{ex_t5}_{args.dataset}_{args.img_add_pos}_e{args.num_train_epochs}" \
f"_res-mlm-{args.t5_mlm}_Ebias-{args.add_ent_bias}_lr{args.learning_rate}" \
f"_seqlen{args.max_seq_length}_res{args.max_seq_a_length}_bs{args.per_gpu_train_batch_size}" \
f"_load{loading}_cap{int(args.add_textual_cap)}_ent{int(args.add_textual_ent)}_" \
f"eda{eda_append}_kl{args.knowledge_len}_ev{args.ent_img_num}_{ex_appended}"
save_folder = os.path.join(args.output_dir, experiment)
t_writer = SummaryWriter(os.path.join(save_folder, 'train'), flush_secs=5)
v_writer = SummaryWriter(os.path.join(save_folder, 'val'), flush_secs=5)
logging_file = f"train_t5_{now.month}.{now.day}.log"
logging = Logger(os.path.join(save_folder, logging_file))
logging.info('\n*******************************************************************************\n')
logging.info("the configuration:")
logging.info(str(args).replace(',', '\n'))
logging.info('Loading models...')
## prepare model & resize tokenizer
config = T5Config.from_pretrained(t5_version, cache_dir=cache_dir)
config = model_utils.sort_config(config, args)
tokenizer = T5Tokenizer.from_pretrained(t5_version, cache_dir=cache_dir)
special_tokens_dict = {'mask_token': '<mask>', 'sep_token': '<sep>', 'bos_token': '<s>'}
num_added_toks = tokenizer.add_special_tokens(special_tokens_dict)
logging.info(f'We have added {num_added_toks} tokens to the model..')
model = T5ForMMDialogGeneration.from_pretrained(t5_version, cache_dir=cache_dir, config=config)
model._resize_token_embeddings(len(tokenizer))
if args.load is not None:
print("Loading Trained Model Weight...")
state = torch.load(os.path.join("output", args.load, "model_best_val.pt"))
model.load_state_dict(state)
del state
model.to(device)
model.train()
# model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
# if len(args.gpu) > 1:
# model = nn.DataParallel(model)
p_train_data = model_utils.jsonload(args.trainloader_file)[0]
p_val_data = model_utils.jsonload(args.valloader_file)[0]
if args.dataset == "WOW":
dialog_episodes = json.load(open(args.episodes_train_file))
val_episodes = json.load(open(args.episodes_val_file))
elif args.dataset == "DD":
dialog_episodes = model_utils.jsonload(args.episodes_train_file)
val_episodes = model_utils.jsonload(args.episodes_val_file)
else:
raise NotImplementedError
fea_prefix_path = os.path.join(data_dir, f"processed_resee_data/shared/processed_img_features_clip_{args.vit_version}/")
fea_paths = ["openimagev6_train_clip_vis_fea.pt",
"openimagev6_test_clip_vis_fea.pt",
"openimagev6_val_clip_vis_fea.pt",
"coco_train_clip_vis_fea.pt",
"coco_val_clip_vis_fea.pt",
"flickr30_clip_vis_fea.pt",
"nocaps_clip_vis_fea.pt"]
fea_paths = [os.path.join(fea_prefix_path, item) for item in fea_paths]
turn_img_feats = model_utils.load_pt_list(fea_paths)
if args.dataset == "WOW":
ent_clip_fea_file = "wow/processed_img_features/img_clip_features.pt"
num_aug = 9
elif args.dataset == "DD":
ent_clip_fea_file = "dd/processed_img_features/img_clip_features.pt"
num_aug = 5
else:
ent_clip_fea_file = None
num_aug = 0
ent_img_feats = model_utils.load_pt_list([os.path.join(data_dir, "processed_resee_data", ent_clip_fea_file)])
ent_img_feats = [ent_img_feats]
if args.ent_img_num > 1:
for count in range(1, args.ent_img_num):
## lets only consider DD for now
ent_img_feats.append(model_utils.load_pt_list([os.path.join(data_dir, "processed_resee_data/dd/processed_img_features",
f"img_clip_features-{count+1}.pt")]))
random_token_list = model_utils.gen_random_word_list_t5(tokenizer, len(tokenizer))
train_data = MMDialogDataset(dialog_episodes, device,
p_train_data["entity_turn"],
p_train_data["caption_turn"],
p_train_data["entities_img_feats_ids"][0],
p_train_data["turn_img_feats_ids"],
turn_img_feats=turn_img_feats, ent_img_feats=ent_img_feats,
img_add_pos=args.img_add_pos, img_dim=args.img_feature_dim,
history_in_context=args.history_in_context,
max_episode_length=args.max_episode_length, tokenizer=tokenizer,
max_turn_img_seq_length=args.max_turn_img_length,
per_max_turn_img_seq_length=args.per_max_turn_img_seq_length,
max_ent_img_seq_length=args.max_ent_img_seq_length,
max_seq_length=args.max_seq_length,
max_seq_a_length=args.max_seq_a_length, is_train=True,
mask_response_prob=args.mask_response_prob, mask_prob=0,
random_token_list=random_token_list, num_aug=num_aug,
max_masked_response_tokens=args.max_masked_response_tokens,
max_masked_context_tokens=0, test=args.last_res,
dataset=args.dataset, add_cap=args.add_textual_cap,
add_ent=args.add_textual_ent, seq2seq=True, EDA=args.eda,
eda_pos=args.eda_pos, no_ent_vis=args.no_ent_vis,
no_turn_vis=args.no_turn_vis, knowledge_len=args.knowledge_len)
val_data = MMDialogDataset(val_episodes, device,
p_val_data["entity_turn"],
p_val_data["caption_turn"],
p_val_data["entities_img_feats_ids"][0],
p_val_data["turn_img_feats_ids"],
turn_img_feats=turn_img_feats, ent_img_feats=ent_img_feats,
img_add_pos=args.img_add_pos, img_dim=args.img_feature_dim,
history_in_context=args.history_in_context,
max_episode_length=args.max_episode_length, tokenizer=tokenizer,
max_turn_img_seq_length=args.max_turn_img_length,
per_max_turn_img_seq_length=args.per_max_turn_img_seq_length,
max_ent_img_seq_length=args.max_ent_img_seq_length,
max_seq_length=args.max_seq_length,