-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathpeft_trainer.py
1301 lines (1113 loc) · 66.2 KB
/
peft_trainer.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
from transformers import T5Tokenizer, T5ForConditionalGeneration,T5Model, T5Config, Seq2SeqTrainer, Seq2SeqTrainingArguments, Trainer, AutoModelForSequenceClassification, Seq2SeqAdapterTrainer
from datasets import load_dataset, load_metric, concatenate_datasets
import numpy as np
from torch.nn import CrossEntropyLoss
import torch
from transformers import (
GPT2TokenizerFast,
AdamW,
Adafactor,
get_scheduler,
EarlyStoppingCallback
)
from transformers import AutoModelForSeq2SeqLM, AutoAdapterModel
from torch.utils.data import DataLoader
import numpy as np
import os
import time
from functools import partial
from transformers import (
AutoConfig,
AutoTokenizer,
DataCollatorForSeq2Seq,
default_data_collator,
set_seed,
create_optimizer,
get_constant_schedule
)
from transformers import get_linear_schedule_with_warmup, get_constant_schedule_with_warmup
from transformers.optimization import Adafactor, AdafactorSchedule, AdamW
from utils import get_embedding_layer, get_soft_prompt_token_list, get_all_params, round_up
import transformers
from peft import get_peft_config, get_peft_model, LoraConfig, TaskType, PromptTuningConfig,PrefixTuningConfig
from util.ni_dataset_collator import DataCollatorForNI
from copy import deepcopy
import datetime
class PEFTTrainer:
def __init__(self, training_args, data_args, model_args, peft_args):
self.training_args = training_args
self.data_args = data_args
self.model_args = model_args
self.peft_args = peft_args
# self.arguments = arguments
self.model_name_or_path = self.model_args.model_name_or_path
self.verbalizers = self.get_dataset_verbalizers(self.data_args.dataset_name if self.data_args.dataset_name != "super_glue" else self.data_args.dataset_config_name)
self.load_model()
self.model_trainable_params = sum(p.numel() for p in self.model.parameters())
self.org_vocab_size = self.tokenizer.vocab_size
# assert len(self.models) == len(self.tokenizers) == len(self.configs)
self.new_vocab_sizes = [self.org_vocab_size]
self.model_cache = deepcopy(self.model)
self.default_optimizer_n_scheduler = self.training_args.default_optimizer_n_scheduler
"""
Model is loaded, now we need to set up the trainer
1. prepare peft model
2. set up trainer
"""
task_type = TaskType.SEQ_2_SEQ_LM
init_from_text = False
if self.data_args.dataset_name == "sst2" and self.model_args.model_arch == "encoder":
task_type = TaskType.SEQ_CLS
prompt_tuning_init = None # NOTE: it decides whether prompt init is used
prompt_tuning_init_text = None
init_text_tokenizer_name_or_path = None
prompt_tuning_init = "TEXT"
print("task type is seq_cls")
prompt_tuning_init_text=" ".join(self.verbalizers)
init_text_tokenizer_name_or_path = self.model_name_or_path
# given that model is already loaded, we can check if
# model path exsits locally
if not os.path.exists(self.model_args.model_name_or_path):
self.configure_n_convert_peft()
else:
# TODO: maybe read available peft module rather than
# read by module name
adapter_name = self.model_args.tuning_mode
self.model.set_active_adapters(adapter_name)
if self.peft_args.trainable_params_percentage and not self.model_args.tuning_mode in ["compactor","fine_tuning"]:
# not check compactor
assert abs(self.peft_args.trainable_params_percentage - cur_trainable_params_percentage) < 0.002, f"trainable_params_percentage {self.peft_args.trainable_params_percentage} is not matched with cur_trainable_params_percentage {cur_trainable_params_percentage}"
# deactivate
# NOTE: set lm head trainable again
# if hasattr(self.model, "lm_head"): # peft model
# self.model.lm_head.weight.requires_grad = True
# else: # adapter model
# self.model.heads["seq2seq-head-sst-2"][0].weight.requires_grad
# self.training_args.run_name += f"lm_head_trainable"
time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
# self.model = self.model_cache
if self.peft_args.trainable_params_percentage:
self.training_args.run_name += f"_trainable_params_percentage_{self.peft_args.trainable_params_percentage}"
# prepare runname before passing to trainer
if self.data_args.num_training_tasks:
self.training_args.run_name += f"_num_training_tasks_{self.data_args.num_training_tasks}"
self.training_args.run_name += f"_{time}"
# import pdb; pdb.set_trace()
# print("check run_name", self.training_args.run_name)
self.set_up_hf_trainer()
self.tokenizer = self.tokenizer
def configure_n_convert_peft(self):
# model loading procedure:
# 1. load model from model_name_or_path (self.load_model())
# 2. not satisfied with peft, load model from self.model_cache and convert again. self.model = deepcopy(self.model_cache)
if self.model_args.tuning_mode == "prompt_tuning":
cur_prompt_len = 1
assert self.peft_args.trainable_params_percentage is not None or self.peft_args.num_soft_tokens > 0, "either prompt_len or trainable_params_percentage should be set"
config = PromptTuningConfig(
task_type=task_type,
num_virtual_tokens=cur_prompt_len,
inference_mode=False,
device= self.peft_args.module_device,
# prompt_tuning_init="TEXT",
# prompt_tuning_init_text=prompt_tuning_init_text,
# tokenizer_name_or_path=init_text_tokenizer_name_or_path,
)
cur_trainable_params_percentage = self.convert_to_peft(config)
while self.peft_args.trainable_params_percentage and cur_trainable_params_percentage < self.peft_args.trainable_params_percentage:
config = PromptTuningConfig(
task_type=task_type,
num_virtual_tokens=cur_prompt_len,
inference_mode=False,
device= self.peft_args.module_device,
# prompt_tuning_init="TEXT",
# prompt_tuning_init_text=prompt_tuning_init_text,
# tokenizer_name_or_path=init_text_tokenizer_name_or_path,
)
cur_trainable_params_percentage = self.convert_to_peft(config, reset_peft=True)
print("prompt length is {}".format(cur_prompt_len))
print("trainable params percentage is {}".format(cur_trainable_params_percentage))
cur_prompt_len += 1
self.training_args.run_name += "_prompt_len_{}".format(cur_prompt_len-1)
elif self.model_args.tuning_mode == "prefix_tuning":
from transformers.adapters import PrefixTuningConfig
# from peft import PrefixTuningConfig
cur_prefix_len = 1 if self.peft_args.prefix_len is None else self.peft_args.prefix_len
bottleneck_size = 576
assert self.peft_args.trainable_params_percentage is not None or self.peft_args.prefix_len > 0, "either prefix_len or trainable_params_percentage should be set"
config = PrefixTuningConfig(prefix_length=cur_prefix_len, bottleneck_size=bottleneck_size,
encoder_prefix=True,
cross_prefix=True)
cur_trainable_params_percentage = self.convert_to_peft(config)
while self.peft_args.trainable_params_percentage and cur_trainable_params_percentage < self.peft_args.trainable_params_percentage:
cur_prefix_len += 1
# config = PrefixTuningConfig(prefix_length=cur_prefix_len, flat=True)
config = PrefixTuningConfig(prefix_length=cur_prefix_len, bottleneck_size=bottleneck_size,
encoder_prefix=True,
cross_prefix=True)
cur_trainable_params_percentage = self.convert_to_peft(config, reset_peft=True)
print("prefix length is {}".format(cur_prefix_len))
self.training_args.run_name += "_prefix_len_{}".format(cur_prefix_len)
elif self.model_args.tuning_mode == "lora":
# peft package
cur_lora_r = 15 if self.peft_args.lora_r is None else self.peft_args.lora_r
assert self.peft_args.trainable_params_percentage is not None or self.peft_args.lora_r > 0, "either lora_r or trainable_params_percentage should be set"
# config = LoraConfig(
# task_type=task_type,
# inference_mode=False,
# r=cur_lora_r,
# lora_alpha=32,
# lora_dropout=0.1,
# # lora_modules
# target_modules= list(self.peft_args.lora_modules) if self.peft_args.lora_modules else ["q", "v"],
# )
# cur_trainable_params_percentage = self.convert_to_peft(config, reset_peft=True)
# print("cur_lora_r", cur_lora_r, "cur_trainable_params_percentage", cur_trainable_params_percentage)
# while self.peft_args.trainable_params_percentage and cur_trainable_params_percentage < self.peft_args.trainable_params_percentage:
# cur_lora_r += 1
# config = LoraConfig(
# task_type=task_type,
# inference_mode=False,
# r=cur_lora_r,
# lora_alpha=32,
# lora_dropout=0.1,
# target_modules=list(self.peft_args.lora_modules) if self.peft_args.lora_modules else ["q", "v"],
# )
# cur_trainable_params_percentage = self.convert_to_peft(config, reset_peft=True)
# print("cur_lora_r", cur_lora_r, "cur_trainable_params_percentage", cur_trainable_params_percentage)
from transformers.adapters import LoRAConfig
config = LoRAConfig(r=cur_lora_r,
alpha=16,
attn_matrices=list(self.peft_args.lora_modules) if self.peft_args.lora_modules else ["q", "v"]
)
# self.model.add_adapter("lora_adapter", config=config)
cur_trainable_params_percentage = self.convert_to_peft(config)
while self.peft_args.trainable_params_percentage and cur_trainable_params_percentage < self.peft_args.trainable_params_percentage:
cur_lora_r += 1
config = LoRAConfig(
r=cur_lora_r,
alpha=16,
attn_matrices=list(self.peft_args.lora_modules) if self.peft_args.lora_modules else ["q", "v"]
)
cur_trainable_params_percentage = self.convert_to_peft(config, reset_peft=True)
print("cur_lora_r", cur_lora_r, "cur_trainable_params_percentage", cur_trainable_params_percentage)
# cur_trainable_params_percentage = self.convert_to_peft(config, reset_peft=True)
self.peft_args.lora_r = cur_lora_r
self.training_args.run_name += "_lora_r_" + str(cur_lora_r)
if self.peft_args.lora_modules:
self.training_args.run_name += "_lora_modules_" + self.peft_args.lora_modules
elif self.model_args.tuning_mode == "ia3":
from transformers.adapters import IA3Config
cur_lora_r = 15 if self.peft_args.lora_r is None else self.peft_args.lora_r
assert self.peft_args.trainable_params_percentage is not None or self.peft_args.lora_r > 0, "either lora_r or trainable_params_percentage should be set"
config = IA3Config(
r = cur_lora_r,
)
cur_trainable_params_percentage = self.convert_to_peft(config)
print("cur_lora_r", cur_lora_r, "cur_trainable_params_percentage", cur_trainable_params_percentage)
while self.peft_args.trainable_params_percentage and cur_trainable_params_percentage < self.peft_args.trainable_params_percentage:
cur_lora_r += 1
config = IA3Config(
r = cur_lora_r,
)
cur_trainable_params_percentage = self.convert_to_peft(config, reset_peft=True)
print("cur_lora_r", cur_lora_r, "cur_trainable_params_percentage", cur_trainable_params_percentage)
self.training_args.run_name += "_lora_r_" + str(cur_lora_r)
elif self.model_args.tuning_mode in ["adapter", "compactor"]:
cur_reduction_factor = 64 if self.peft_args.reduction_factor is None else self.peft_args.reduction_factor
assert self.peft_args.trainable_params_percentage is not None or self.peft_args.reduction_factor > 0, "either reduction_factor or trainable_params_percentage should be set"
from transformers.adapters import AdapterConfig, HoulsbyConfig, CompacterConfig
# check existing adapter and remove them
# config = AdapterConfig()
if self.model_args.tuning_mode == "adapter":
config = HoulsbyConfig(reduction_factor=cur_reduction_factor)
else:
config = CompacterConfig(reduction_factor=cur_reduction_factor,
phm_dim=self.peft_args.phm_dimension )
cur_trainable_params_percentage = self.convert_to_peft(config)
while self.peft_args.trainable_params_percentage and cur_trainable_params_percentage < self.peft_args.trainable_params_percentage:
cur_reduction_factor /=1.01
if self.model_args.tuning_mode == "adapter":
config = HoulsbyConfig(reduction_factor=cur_reduction_factor)
else:
config = CompacterConfig(reduction_factor=cur_reduction_factor,
phm_dim=self.peft_args.phm_dimension)
cur_trainable_params_percentage = self.convert_to_peft(config, reset_peft=True)
print(f"cur_trainable_params_percentage: {cur_trainable_params_percentage}, cur_reduction_factor: {cur_reduction_factor}")
# only keep 4 digits for reduction factor
self.training_args.run_name += f"_reduction_factor_{cur_reduction_factor:.4f}"
if self.model_args.tuning_mode == "compactor":
self.training_args.run_name += f"_phm_dim_{self.peft_args.phm_dimension}"
elif self.model_args.tuning_mode == "parallel_adapter":
from transformers.adapters import ParallelConfig
config = ParallelConfig(reduction_factor= self.peft_args.reduction_factor)
cur_trainable_params_percentage = self.convert_to_peft(config)
print(f"cur_trainable_params_percentage: {cur_trainable_params_percentage}")
self.training_args.run_name += f"_reduction_factor_{self.peft_args.reduction_factor:.4f}"
elif self.model_args.tuning_mode == "embedding_tuning":
self.convert_to_embedding_tuning()
if self.peft_args.num_soft_tokens > 0:
self.peft_args.num_soft_tokens = 0
print("num_soft_tokens is set to 0 for embedding tuning mode")
elif self.model_args.tuning_mode == "bitfit":
self.convert_to_peft()
elif self.model_args.tuning_mode == "fine_tuning":
# no converting needed
pass
elif self.model_args.tuning_mode == "lm_head_tuning":
for param in self.model.parameters():
param.requires_grad = False
# lm head takes almost 10% paramaters
for name, module in self.model.named_modules():
if "lm_head" in name:
module.weight.requires_grad = True
elif self.model_args.tuning_mode == "layer_tuning":
for param in self.model.parameters():
param.requires_grad = False
layers = []
# NOTE: we only fine-tune attention weights for now
if self.peft_args.layer_name == "first_encoder_layer":
layers.append(self.model.encoder.block[0].layer[0])
elif self.peft_args.layer_name == "last_encoder_layer":
layers.append(self.model.encoder.block[-1].layer[0])
elif self.peft_args.layer_name == "first_decoder_layer":
layers.append(self.model.decoder.block[0].layer[0])
elif self.peft_args.layer_name == "last_decoder_layer":
layers.append(self.model.decoder.block[-1].layer[0])
elif self.peft_args.layer_name == "custom":
# all decoder layer
modules = self.model.decoder.block
for m in modules:
layers.append(m.layer[0])
elif self.peft_args.layer_name == "custom2":
# all decoder layer
modules = self.model.decoder.block
for m in modules:
layers.append(m.layer[0])
else:
raise NotImplementedError(f"layer_name {self.peft_args.layer_name} is not implemented")
for l in layers:
for name, module in l.named_modules():
# if "selfattention" in name.lower():
if hasattr(module, "weight"):
module.weight.requires_grad = True
print("activate gradient for ", name)
elif self.model_args.tuning_mode == "lora+adapter":
from transformers.adapters import AdapterConfig, HoulsbyConfig, CompacterConfig
# model.base_model.model_frozen = True
# print('cur_trainable_params_percentage', self.check_trainable_parameters())
# # peft package
# cur_lora_r = 15 if self.peft_args.lora_r is None else self.peft_args.lora_r
# assert self.peft_args.trainable_params_percentage is not None or self.peft_args.lora_r > 0, "either lora_r or trainable_params_percentage should be set"
# task_type = TaskType.SEQ_2_SEQ_LM
# config = LoraConfig(
# task_type=task_type,
# inference_mode=False,
# r=cur_lora_r,
# lora_alpha=32,
# lora_dropout=0.1
# )
# cur_trainable_params_percentage = self.convert_to_peft(config, reset_peft=True)
# lora 2
cur_lora_r = 40 if self.peft_args.lora_r is None else self.peft_args.lora_r
freeze_model = True
from transformers.adapters import LoRAConfig
config = LoRAConfig(r=cur_lora_r, alpha=16)
self.model.add_adapter("lora_adapter", config=config)
# cur_trainable_params_percentage = self.convert_to_peft(config, reset_peft=True)
self.check_trainable_parameters()
print("lora_adapter is added, trainable parameters: ", self.check_trainable_parameters())
# adapter
cur_reduction_factor=18
peft_config = HoulsbyConfig(reduction_factor=cur_reduction_factor)
reset_peft = False
# cur_trainable_params_percentage = self.convert_to_peft(config, reset_peft=True)
self.model.add_adapter(self.model_args.tuning_mode,
config = peft_config, overwrite_ok=reset_peft)
self.model.train_adapter(self.model_args.tuning_mode)
print('cur_reduction_factor', cur_reduction_factor, 'cur_trainable_params_percentage', self.check_trainable_parameters())
# do not freeze model as model itself is already frozen
# we don't want to freeze the lora module
# self.model.train_adapter(["sst-2","lora_adapter"] , freeze_model=freeze_model)
self.model.train_adapter(["sst-2", "lora_adapter"], freeze_model=freeze_model)
if self.model_args.model_arch == "encoder":
self.model.add_classification_head("classification-head-sst-2", num_labels=2, overwrite_ok=reset_peft)
elif self.model_args.model_arch == "encoder-decoder":
self.model.add_seq2seq_lm_head("seq2seq-head-sst-2", overwrite_ok=reset_peft)
# reset weight self.model.heads["seq2seq-head-sst-2"][0].weight
self.model.heads["seq2seq-head-sst-2"][0].weight = self.model_lm_head_weight
self.model.heads["seq2seq-head-sst-2"][0].weight.requires_grad = False
else:
raise NotImplementedError(
f"Not implemented for model arch: {self.model_args.model_arch}"
)
self.model.set_active_adapters(["sst-2", "lora_adapter"])
print('cur_reduction_factor', cur_reduction_factor, 'cur_trainable_params_percentage', self.check_trainable_parameters())
self.peft_args.lora_r = cur_lora_r
self.training_args.run_name += "_lora_r_" + str(cur_lora_r)
# invalid
elif self.model_args.tuning_mode == "unipelt":
from transformers.adapters import UniPELTConfig, PrefixTuningConfig, PfeifferConfig, LoRAConfig, HoulsbyConfig
gating = False
reset_peft=False
# peft_config = UniPELTConfig(
# PrefixTuningConfig(prefix_length=1, use_gating=self.peft_args.use_pelt_gate),
# PfeifferConfig(reduction_factor=500, use_gating=self.peft_args.use_pelt_gate),
# LoRAConfig(r=self.peft_args.lora_r, use_gating=self.peft_args.use_pelt_gate),
# )
peft_config = UniPELTConfig(
PrefixTuningConfig(prefix_length=1, use_gating=self.peft_args.use_pelt_gate),
HoulsbyConfig(reduction_factor=500, use_gating=self.peft_args.use_pelt_gate),
LoRAConfig(r=self.peft_args.lora_r, use_gating=self.peft_args.use_pelt_gate),
)
self.model.add_adapter(adapter_name, config = peft_config)
self.model.train_adapter(adapter_name)
self.model.add_seq2seq_lm_head(f"seq2seq-head-{adapter_name}", overwrite_ok=reset_peft)
self.model.set_active_adapters(adapter_name)
# reset weight self.model.heads["seq2seq-head-sst-2"][0].weight
self.model.heads[f"seq2seq-head-{adapter_name}"][0].weight = self.model_lm_head_weight
self.model.heads[f"seq2seq-head-{adapter_name}"][0].weight.requires_grad = False
print('cur_trainable_params_percentage', self.check_trainable_parameters())
self.training_args.run_name += "lora_r_" + str(self.peft_args.lora_r)
self.training_args.run_name += "_use_peft_gate_" + str(self.peft_args.use_pelt_gate)
else:
raise NotImplementedError(f"mode {self.model_args.tuning_mode} is not implemented")
def set_up_hf_trainer(self):
del self.model_cache
if self.training_args.model_parallel_gpus > 1 and torch.cuda.device_count() > 1:
if torch.cuda.device_count() != self.training_args.model_parallel_gpus:
print(f"WARNING: model parallel is enabled but the number of GPUs does not match the number of GPUs specified in the model_parallel_gpus argument. Using all available GPUs. ({torch.cuda.device_count()} GPUs found)")
if hasattr(self.model, "parallelize"):
self.model.parallelize()
else:
print(f"Model {self.model_name_or_path} cannot be parallelized")
optimizer = Adafactor(
self.model.parameters(),
# filter(lambda p: p.requires_grad, self.model.parameters()),
lr= self.training_args.learning_rate,
eps=(1e-30, 1e-3),
clip_threshold=1.0,
decay_rate=-0.8,
beta1=None,
weight_decay=1e-5,
# fixed learning rate
scale_parameter=False,
relative_step=False,
warmup_init=False,
)
lr_scheduler = get_constant_schedule(optimizer)
if self.data_args.dataset_name == "ni":
dataset_dependent_data_collator = DataCollatorForNI(
self.tokenizer,
model=self.model,
padding="max_length" if self.data_args.pad_to_max_length else "longest",
max_source_length=self.data_args.max_source_length,
max_target_length=self.data_args.max_target_length,
label_pad_token_id=self.tokenizer.pad_token_id,
pad_to_multiple_of=8 if self.training_args.bf16 else None,
add_task_name=self.data_args.add_task_name,
add_task_definition=self.data_args.add_task_definition,
num_pos_examples=self.data_args.num_pos_examples,
num_neg_examples=self.data_args.num_neg_examples,
add_explanation=self.data_args.add_explanation,
tk_instruct=self.data_args.tk_instruct
)
self.training_args.remove_unused_columns = False
else:
dataset_dependent_data_collator = default_data_collator
if self.model_args.tuning_mode in ["adapter", "compactor", "prefix_tuning", "ia3", "parallel_adapter"]: # "prefix_tuning",
self.trainer = Seq2SeqAdapterTrainer(
model = self.model,
tokenizer = self.tokenizer,
train_dataset = None,
eval_dataset = None,
args = self.training_args,
optimizers=[optimizer, lr_scheduler] if not self.default_optimizer_n_scheduler else [None, None],
compute_metrics=partial(self.compute_metrics, is_pred_logits = not self.training_args.predict_with_generate),
data_collator=dataset_dependent_data_collator,
)
else:
# self.trainer = Seq2SeqAdapterTrainer(
# model = self.model,
# tokenizer = self.tokenizer,
# train_dataset = None,
# eval_dataset = None,
# args = self.arguments,
# optimizers=[optimizer, lr_scheduler] if not self.default_optimizer_n_scheduler else [None, None],
# compute_metrics=partial(self.compute_metrics, is_pred_logits = not self.training_args.predict_with_generate),
# data_collator=dataset_dependent_data_collator,
# )
self.trainer = Seq2SeqTrainer(
model = self.model,
tokenizer = self.tokenizer,
train_dataset = None,
eval_dataset = None,
args = self.training_args,
optimizers=[optimizer, lr_scheduler] if not self.default_optimizer_n_scheduler else [None, None],
compute_metrics=partial(self.compute_metrics, is_pred_logits = not self.training_args.predict_with_generate),
data_collator=dataset_dependent_data_collator,
)
self.trainer.model = self.model
def _format_prompts(self, prefix, source_strs, verbal_targets, include_labels_in_input = False):
prompts = [""] * len(source_strs)
prompts = [""]* len(source_strs)
formatted_inputs = [f"{prefix} {s_1} {prompt} " for s_1, prompt in zip(source_strs, prompts)]
formatted_inputs = [f"{input} " for input in formatted_inputs]
if include_labels_in_input:
labels = ",".join(verbal_targets)
formatted_inputs = [f"{input} Decide the label in {self.verbalizers}." for input in formatted_inputs]
return formatted_inputs, verbal_targets
def get_dataset_verbalizers(self, dataset):
if dataset in ['sst-2', 'sst2',
'yelp-2', 'mr', 'cr']:
# verbalizers = ['\u0120terrible', '\u0120great'] # num_classes
verbalizers = ['terrible', 'great']
elif dataset == 'agnews':
verbalizers = ['World', 'Sports', 'Business', 'Tech'] # num_classes
elif dataset in ['sst-5', 'yelp-5']:
# verbalizers = ['\u0120terrible', '\u0120bad', '\u0120okay',
# '\u0120good', '\u0120great'] # num_classes
verbalizers = ['terrible', 'bad', 'okay', 'good', 'great']
elif dataset == 'subj':
# verbalizers = ['\u0120subjective', '\u0120objective']
verbalizers = ['subjective', 'objective']
elif dataset == 'trec':
verbalizers = ['Description', 'Entity',
'Expression', 'Human',
'Location', 'Number']
verbalizers = ['Description', 'Entity', 'Expression', 'Human', 'Location', 'Number']
elif dataset == 'yahoo':
verbalizers = ['culture', 'science',
'health', 'education',
'computer', 'sports',
'business', 'music',
'family', 'politics']
elif dataset == 'dbpedia':
verbalizers = ['\u0120Company', '\u0120Education',
'\u0120Artist', '\u0120Sports',
'\u0120Office', '\u0120Transportation',
'\u0120Building', '\u0120Natural',
'\u0120Village', '\u0120Animal',
'\u0120Plant', '\u0120Album',
'\u0120Film', '\u0120Written']
verbalizers = ['Company', 'Education', 'Artist', 'Sports', 'Office', 'Transportation', 'Building', 'Natural', 'Village', 'Animal', 'Plant', 'Album', 'Film', 'Written']
elif dataset in ['axb', 'axg']:
verbalizers = ["No", "Yes"]
elif dataset in ['cb', "rtx"]:
verbalizers = ["Yes", "No"]
elif dataset in ['copa', ]:
verbalizers = ["choice1", "choice2"]
elif dataset in ['boolq','multirc', 'wic', 'wsc', 'wsc_fixed']:
if dataset == 'boolq':
# only boolq
verbalizers = ["True", "False"]
else:
verbalizers = ["False", "True"]
elif dataset == 'record':
verbalizers = [None] # answer is text
elif dataset == 'yelp_review_full':
verbalizers = ['1', '2', '3', '4', '5']
elif dataset == 'ag_news':
verbalizers = ['World', 'Sports', 'Business', 'Sci/Tech']
elif dataset == 'ni':
verbalizers = []
else:
raise NotImplementedError("Dataset not supported: " + dataset)
return verbalizers
def load_model(self):
print(
"Loading",
self.model_args.model_name_or_path,
"(for large models, this might take a while)",
)
print("Files will be cached at:", self.training_args.cache_dir)
print(
"Ensure this directory is persistent if you do not want to download model files again!"
)
# for self.model_name_or_path in self.model_name_or_path:
# self.config = AutoConfig.from_pretrained(
# self.model_name_or_path,
# cache_dir=self.training_args.cache_dir,
# gradient_checkpointing=self.arguments.gradient_checkpointing,
# use_cache=not self.arguments.gradient_checkpointing,
# )
self.tokenizer = AutoTokenizer.from_pretrained(
self.model_name_or_path,
cache_dir=self.training_args.cache_dir,
# use_cache = self.arguments.use_cache,
use_fast=True,
return_tensors="pt"
)
# m = T5ForConditionalGeneration.from_pretrained(
# self.model_name_or_path,
# )
if "t5" in self.model_name_or_path or "bart" in self.model_name_or_path:
if self.model_args.tuning_mode in ["fine_tuning"]:
# trainer not compactiable with AdapterTrainer yet due to forward function not returning loss
self.model = AutoModelForSeq2SeqLM.from_pretrained(self.model_name_or_path, cache_dir=self.training_args.cache_dir,)
elif self.model_args.tuning_mode in ["adapter", "compactor", "prefix_tuning", "ia3", "lora", "parallel_adapter"]:
self.model = AutoAdapterModel.from_pretrained(self.model_name_or_path, cache_dir=self.training_args.cache_dir)
elif self.model_args.tuning_mode in ["prompt_tuning"]:
self.model = PeftModelForSeq2Seq.from_pretrained(self.model_name_or_path, cache_dir=self.training_args.cache_dir)
else:
raise NotImplementedError("Tuning mode not supported: " + self.model_args.tuning_mode)
self.model_lm_head_weight = AutoModelForSeq2SeqLM.from_pretrained(self.model_name_or_path, cache_dir=self.training_args.cache_dir).lm_head.weight
elif "roberta" in self.model_name_or_path:
self.model = AutoModelForSequenceClassification.from_pretrained(self.model_name_or_path)
elif "gpt2" in self.model_name_or_path or "bloom" in self.model_name_or_path or "opt" in self.model_name_or_path:
from transformers import AutoModelForCausalLM
self.model = AutoModelForCausalLM.from_pretrained(
self.model_name_or_path,
# from_tf=bool(".ckpt" in self.model_name_or_path),
# config=m_config,
cache_dir=self.training_args.cache_dir,
)
else:
raise NotImplementedError("Model not supported: " + self.model_name_or_path)
# Wrap model in adapter package
# NOTE: temp implementation
# import AutoModelWithHeads
from transformers import AutoModelWithHeads
if self.model_args.tuning_mode in ["adapter", "compactor", "prefix_tuning", "ia3"] : # "prefix_tuning",
self.model = AutoAdapterModel.from_pretrained(self.model_name_or_path, cache_dir=self.training_args.cache_dir,)
if self.tokenizer.pad_token is None:
assert self.model_name_or_path == "gpt2", "Only gpt2 is expected not having pad tokens for now"
# gpt2 model
self.tokenizer.add_special_tokens({'pad_token': '[PAD]'})
m.resize_token_embeddings(len(self.tokenizer))
self.padding = "max_length" if self.data_args.pad_to_max_length else False
print('gpt2 requires padding to max length')
if "gpt2" in self.model_name_or_path:
self.padding = "max_length"
self.training_args.run_name += self.model_name_or_path
def preprocess(self, examples, class_ids = [0,1], evaluation=False):
# disable prefix prompt
# prefix_prompt = "".join(get_soft_prompt_token_list(self.peft_args.num_soft_tokens))
prefix_prompt = ""
if self.peft_args.num_soft_tokens ==0:
assert prefix_prompt == ""
if self.model_args.model_arch == "encoder":
add_prompt_for_gen = False
else:
add_prompt_for_gen = True
if self.data_args.dataset_name == "sst2":
prompt_for_gen = "Sentiment:"
inputs =["Sentence: " + sent for sent, label_id in zip(examples["sentence"], examples["label"]) if label_id in class_ids]
# verbalize the sentiment id to tokens
# it's not used for t5 evaluation though (label id is used instead)
verbal_targets = [self.verbalizers[l]
for l in examples["label"] if l in class_ids]
elif self.data_args.dataset_name == "yelp_review_full":
prompt_for_gen = "give the review score from 1-5 stars:"
inputs =["Sentence: " + sent for sent, label_id in zip(examples["text"], examples["label"]) if label_id in class_ids]
verbal_targets = [self.verbalizers[l]
for l in examples["label"] if l in class_ids]
elif self.data_args.dataset_name == "super_glue":
prompt_for_gen = "Answer:"
if self.data_args.dataset_config_name in['axb', 'axg'] :
raise NotImplementedError("axb and axg are not implemented yet")
inputs = ["Premise: " + premise + " Hypothesis: " + hypothesis + "Given the premise, is the hypothesis correct? Answer: " for premise, hypothesis in zip(examples["premise"], examples["hypothesis"])]
verbal_targets = [self.verbalizers[l]
for l in examples["label"] if l in class_ids]
elif self.data_args.dataset_config_name in ["boolq", "multirc"]:
if self.data_args.dataset_config_name == "boolq":
inputs = ["Question: " + question + "Passage: " + passage for question, passage, label_id in zip(examples["question"], examples["passage"], examples["label"]) if label_id in class_ids]
elif self.data_args.dataset_config_name == "multirc":
inputs = ["Question: " + question + "Paragraph: " + paragraph for question, paragraph, label_id in zip(examples["question"], examples["paragraph"], examples["label"]) if label_id in class_ids]
# inputs = ["" for question, paragraph in zip(examples["question"], examples["paragraph"])]
# inputs = ["Passage: " + passage + " Question: " + question for question, passage in zip(examples["question"], examples["passage"])]
verbal_targets = [self.verbalizers[l]
for l in examples["label"] if l in class_ids]
elif self.data_args.dataset_config_name in ["wic"]:
prompt_for_gen = "" # NOTE: wic has a special format
inputs = ["Sentence1: " + sentence1 + " Sentence2: " + sentence2 + f" Question: does the word '{word}' have the same meaning in the two sentences? Answer: " for sentence1, sentence2, word, label_id in zip(examples["sentence1"], examples["sentence2"], examples["word"], examples["label"]) if label_id in class_ids]
verbal_targets = [self.verbalizers[l] for l, label_id in zip(examples["label"], examples["label"]) if label_id in class_ids]
elif self.data_args.dataset_name in ["trec"]:
prompt_for_gen = " What's the type of the question? "
inputs = ["Question: " + t for t, label_id in zip(examples["text"],examples["coarse_label"]) if label_id in class_ids]
verbal_targets = [self.verbalizers[l] for l, label_id in zip(examples["coarse_label"], examples["coarse_label"]) if label_id in class_ids]
else:
raise NotImplementedError("Dataset not supported: " + self.data_args.dataset_name)
if add_prompt_for_gen:
inputs = [inp + " " + prompt_for_gen for inp in inputs]
formatted_inputs, verbal_targets =\
self._format_prompts(prefix_prompt,
inputs,
# [self.prefix_prompt]*len(inputs),
verbal_targets,
)
print("Sample input: ", formatted_inputs[0])
model_inputs = {}
# multi_model_inputs_l = []# initalize as list of dict, finally merge dict into one dict
# multi_model_inputs = {}
tokenizer = self.tokenizer
model_inputs = tokenizer(
formatted_inputs,
max_length=self.data_args.max_source_length,
padding=self.padding,
truncation=True,
return_tensors='pt'
)
# build label input ids
with tokenizer.as_target_tokenizer():
labels = tokenizer(
verbal_targets,
return_tensors='pt', padding="max_length",
max_length=self.data_args.max_target_length,
# padding=self.padding,
truncation=True,
)
if self.padding == "max_length" and self.data_args.ignore_pad_token_for_loss:
labels["input_ids"][labels["input_ids"]==0] = -100
# labels["input_ids"] = [
# [(l if l != self.tokenizer.pad_token_id else -100)
# for l in label]
# for label in labels["input_ids"]
# ]
# if encoder only model
if self.model_args.model_arch == "encoder":
# for SequenceClassificationModel
# model_inputs["label"] = [l for l in examples["label"] if l in class_ids]
model_inputs["labels"] = [l for l in examples["label"] if l in class_ids]
else:
# model_inputs["label"] = labels["input_ids"]
model_inputs["labels"] = labels["input_ids"]
if evaluation:
label_name = "label" if self.data_args.dataset_name not in ["trec"] else "coarse_label"
label_class_ids = [l for l in examples[label_name] if l in class_ids]
model_inputs["class_ids"] = torch.tensor(label_class_ids)
# model_inputs = add_idx_to_inputs_keys(model_inputs, model_idx)
return model_inputs
def load_dataset(self, train, valid, test=False):
"""
dataset loading pipeline:
1. load dataset from huggingface datasets
2. preprocess dataset
3. tokenize dataset and have multi-model inputs like (input_ids_0, input_ids_1, input_ids_2, labels), also padding and convert to tensor
4. dataloader with tokenizer inside, it requires tokenizer to provide padding token id
5.
"""
if self.data_args.dataset_name == "ni":
assert self.data_args.task_dir is not None, "task_dir is required for NaturalInstructions dataset"
assert self.data_args.data_dir is not None, "data_dir is required for NaturalInstructions dataset"
# Get the NaturalInstructions dataset
raw_datasets = load_dataset(
"util/ni_dataset.py",
data_dir=self.data_args.data_dir,
task_dir=self.data_args.task_dir,
cache_dir=self.training_args.cache_dir,
max_num_instances_per_task=self.data_args.max_num_instances_per_task,
max_num_instances_per_eval_task=self.data_args.max_num_instances_per_eval_task,
download_mode = "reuse_dataset_if_exists" if not self.data_args.overwrite_cache else "force_redownload",
)
flat_data_dir = self.data_args.data_dir.replace("/", "_")
self.training_args.run_name += flat_data_dir
self.training_args.run_name += f"_max_num_instances_per_task_{self.data_args.max_num_instances_per_task}"
if self.training_args.dev:
raw_datasets["validation"] = raw_datasets["validation"].select(range(20))
self.trainer.train_dataset = raw_datasets["train"]
self.trainer.eval_dataset = raw_datasets["validation"]
self.eval_dataset = raw_datasets["validation"]
self.test_dataset = raw_datasets["test"]
else:
raw_datasets = load_dataset(self.data_args.dataset_name, self.data_args.dataset_config_name)
column_names = raw_datasets["train"].column_names
if train:
self.train_dataset = raw_datasets["train"].map(
self.preprocess,
batched=True,
remove_columns= column_names,
num_proc=1,
# load_from_cache_file=self.arguments.dataset_cache,
fn_kwargs = {"evaluation": False},
# fn_kwargs = {"evaluation": True},
desc="Running tokenizer on train dataset",
)
# sample_input = np.array(self.train_dataset[0]["input_ids_0"])
sample_input = np.array(self.train_dataset[0]["input_ids"])
# sample_label = np.array(self.train_dataset[0]["labels_0"])
sample_label = np.array(self.train_dataset[0]["labels"])
sample_input[sample_input==-100] = 0
sample_label[sample_label==-100] = 0
print("train dataset input sample", sample_input, "\n", self.tokenizer.decode(sample_input))
print("train dataset label sample", sample_label,"\n", self.tokenizer.decode(sample_label))
self.train_dataset.set_format(type="torch")
self.trainer.train_dataset = self.train_dataset
if valid:
if self.data_args.dataset_name in ["yelp_review_full", "ag_news", "trec"]:
valid_split_name = "test"
else:
valid_split_name = "validation"
if self.training_args.dev:
true_val_dataset = raw_datasets[valid_split_name].map(
self.preprocess,
batched=True,
remove_columns=column_names,
num_proc=1,
# load_from_cache_file=self.arguments.dataset_cache,
# fn_kwargs = {"evaluation": False}, # hf internal validation
fn_kwargs = {"evaluation": True, "class_ids":[0]},
desc="Running tokenizer on validation dataset",
)
false_val_dataset = raw_datasets[valid_split_name].map(
self.preprocess,
batched=True,
remove_columns=column_names,
num_proc=1,
# load_from_cache_file=self.arguments.dataset_cache,
# fn_kwargs = {"evaluation": False}, # hf internal validation
fn_kwargs = {"evaluation": True, "class_ids":[1]},
desc="Running tokenizer on validation dataset",
)
# select 100 data from each class and merge them
self.eval_dataset = concatenate_datasets([true_val_dataset.select(range(100)),false_val_dataset.select(range(100))])
else:
self.eval_dataset = raw_datasets[valid_split_name].map(
self.preprocess,
batched=True,
remove_columns=column_names,
num_proc=1,
# load_from_cache_file=self.arguments.dataset_cache,
# fn_kwargs = {"evaluation": False}, # hf internal validation
fn_kwargs = {"evaluation": True},
desc="Running tokenizer on validation dataset",
)
self.eval_dataset.set_format(type="torch")
self.trainer.eval_dataset = self.eval_dataset
def _deactivate_relevant_gradients(self, trainable_components):
"""
https://github.com/benzakenelad/BitFit/blob/7ead19a8350a01d5701f9e2df896a1c5b42c3723/glue_evaluator.py#L612
"""
for param in self.model.parameters():
param.requires_grad = False
if trainable_components:
trainable_components = trainable_components + ['pooler.dense.bias']
trainable_components = trainable_components + ['classifier']
# it iterates exsiting parameters only
# bias init must be done before this
for name, param in self.model.named_parameters():
for component in trainable_components:
# print(f"check component {name}")
# if component in name:
# print(f"activate {name}")
# param.requires_grad = True
# break
# brute force bias activation
if "bias" in name:
print(f"activate {name}")
param.requires_grad = True
break
# import pdb; pdb.set_trace()
# print('break point for bias activiation')
def convert_to_peft(self, peft_config=None, reset_peft=False):
"""
1. prepare peft model
2. set up trainer
Args:
peft_config (_type_): _description_
"""
adapter_name = self.model_args.tuning_mode
if self.model_args.tuning_mode in ["adapter", "compactor", "prefix_tuning", "ia3", "lora", "parallel_adapter"]: # prefix_tuning
# add and activate adapter
self.model.add_adapter(adapter_name, config = peft_config, overwrite_ok=reset_peft)
self.model.train_adapter(adapter_name)
if self.model_args.model_arch == "encoder":
self.model.add_classification_head(f"classification-head-{adapter_name}", num_labels=2, overwrite_ok=reset_peft)
elif self.model_args.model_arch == "encoder-decoder":
self.model.add_seq2seq_lm_head(f"seq2seq-head-{adapter_name}", overwrite_ok=reset_peft)
# reset weight self.model.heads["seq2seq-head-sst-2"][0].weight
self.model.heads[f"seq2seq-head-{adapter_name}"][0].weight = self.model_lm_head_weight
self.model.heads[f"seq2seq-head-{adapter_name}"][0].weight.requires_grad = False
else:
raise NotImplementedError(
f"Not implemented for model arch: {self.model_args.model_arch}"
)
self.model.set_active_adapters(self.model_args.tuning_mode)
# self.model.freeze_model(True)
elif self.model_args.tuning_mode == "bitfit":
# if self.model_args.model_arch == "encoder":
# # deactivate gradients except for bias terms
# BIAS_TERMS_DICT = {
# 'intermediate': 'intermediate.dense.bias',
# 'key': 'attention.self.key.bias',
# 'query': 'attention.self.query.bias',
# 'value': 'attention.self.value.bias',
# 'output': 'output.dense.bias',
# 'output_layernorm': 'output.LayerNorm.bias',
# 'attention_layernorm': 'attention.output.LayerNorm.bias',
# 'all': 'bias',
# }
# elif self.model_args.model_arch == "encoder-decoder":
# BIAS_TERMS_DICT = {
# 'intermediate': 'intermediate.dense.bias',
# 'key': 'attention.self.key.bias',
# 'query': 'attention.self.query.bias',
# 'value': 'attention.self.value.bias',
# 'output': 'output.dense.bias',
# 'output_layernorm': 'output.LayerNorm.bias',
# 'attention_layernorm': 'attention.output.LayerNorm.bias',
# 'all': 'bias',
# }
# def convert_to_actual_components(components):
# return [BIAS_TERMS_DICT[component] for component in components]
# is_bias_init = False
# for name, module in self.model.named_modules():
# if hasattr(module, "bias") and "lm_head" not in name:
# if module.bias is None:
# print("found none bias, init bias for ", name)
# module.bias = torch.nn.Parameter(torch.randn(module.out_features))
# is_bias_init = True
# if not module.bias.requires_grad: