-
Notifications
You must be signed in to change notification settings - Fork 1
/
utils.py
1148 lines (966 loc) · 54.3 KB
/
utils.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
from copy import deepcopy
import os
import random
import math
import pickle
import json
import avalanche as avl
from PIL import Image
from avalanche.benchmarks.utils.dataset_definitions import IDatasetWithTargets
from avalanche.benchmarks.datasets import default_dataset_location
from avalanche.benchmarks.generators import dataset_benchmark
from avalanche.benchmarks.utils import AvalancheDataset, AvalancheSubset, AvalancheDatasetType, AvalancheConcatDataset
import numpy as np
import torch
from avalanche.benchmarks.utils.data_loader import TaskBalancedDataLoader
from avalanche.benchmarks.utils.dataset_utils import TupleTLabel
from torchvision.transforms import InterpolationMode, autoaugment
from transformers.trainer import Trainer
from tqdm import tqdm
from pandas import read_csv, concat
from sklearn.model_selection import train_test_split
from datasets import Features, Value, ClassLabel, Dataset, Split, load_metric, concatenate_datasets
from transformers import TrainingArguments, DataCollatorForLanguageModeling, EarlyStoppingCallback, AutoTokenizer, \
BeitFeatureExtractor, BeitForImageClassification, BeitForMaskedImageModeling
from torchvision import transforms
from dall_e import load_model
from dall_e.utils import map_pixels
base_data = '/disk3/cossu'
base_save = '/disk2/cossu'
root_path = os.path.join(base_data, 'arxiv_archive-master/processed_data/20200101/per_year')
save_path = os.path.join(base_data, 'arxiv_archive-master/saved')
save_path_small = os.path.join(base_data, 'arxiv_small')
save_path_generic = os.path.join(base_data, 'data', 'emotion')
save_path_qnli = os.path.join(base_data, 'data', 'glue', 'qnli', '1.0.0')
cache_dir = os.path.join(base_save, 'huggingface')
wb_save = os.path.join(base_data, 'wandb')
log_dir = os.path.join(base_data, 'logs')
f1 = load_metric("f1", cache_dir=cache_dir)
accuracy = load_metric('accuracy', cache_dir=cache_dir)
classes = ['acc-phys', 'cmp-lg', 'cs.CE', 'cs.GT', 'cs.PL', 'funct-an', 'math.DS', 'math.OC', 'nucl-ex',
'physics.gen-ph', 'q-bio.MN', 'q-fin.ST', 'adap-org', 'comp-gas', 'cs.CG', 'cs.HC', 'cs.RO', 'gr-qc',
'math.FA', 'math-ph', 'nucl-th', 'physics.geo-ph', 'q-bio.NC', 'q-fin.TR', 'alg-geom', 'cond-mat.dis-nn',
'cs.CL', 'cs.IR', 'cs.SC', 'hep-ex', 'math.GM', 'math.PR', 'patt-sol', 'physics.hist-ph', 'q-bio.OT',
'quant-ph', 'ao-sci', 'cond-mat.mes-hall', 'cs.CR', 'cs.IT', 'cs.SD', 'hep-lat', 'math.GN', 'math.QA',
'physics.acc-ph', 'physics.ins-det', 'q-bio.PE', 'solv-int', 'astro-ph.CO', 'cond-mat.mtrl-sci', 'cs.CV',
'cs.LG', 'cs.SE', 'hep-ph', 'math.GR', 'math.RA', 'physics.ao-ph', 'physics.med-ph', 'q-bio.QM', 'stat.AP',
'astro-ph.EP', 'cond-mat.other', 'cs.CY', 'cs.LO', 'cs.SI', 'hep-th', 'math.GT', 'math.RT', 'physics.app-ph',
'physics.optics', 'q-bio.SC', 'stat.CO', 'astro-ph.GA', 'cond-mat.quant-gas', 'cs.DB', 'cs.MA', 'cs.SY',
'math.AC', 'math.HO', 'math.SG', 'physics.atm-clus', 'physics.plasm-ph', 'q-bio.TO', 'stat.ME',
'astro-ph.HE', 'cond-mat.soft', 'cs.DC', 'cs.MM', 'dg-ga', 'math.AG', 'math.IT', 'math.SP',
'physics.atom-ph', 'physics.pop-ph', 'q-bio', 'stat.ML', 'astro-ph.IM', 'cond-mat.stat-mech', 'cs.DL',
'cs.MS', 'econ.EM', 'math.AP', 'math.KT', 'math.ST', 'physics.bio-ph', 'physics.soc-ph', 'q-fin.CP',
'stat.OT', 'astro-ph.SR', 'cond-mat.str-el', 'cs.DM', 'cs.NA', 'econ.GN', 'math.AT', 'math.LO', 'mtrl-th',
'physics.chem-ph', 'physics.space-ph', 'q-fin.EC', 'stat.TH', 'astro-ph', 'cond-mat.supr-con', 'cs.DS',
'cs.NE', 'econ.TH', 'math.CA', 'math.MG', 'nlin.AO', 'physics.class-ph', 'plasm-ph', 'q-fin.GN', 'supr-con',
'atom-ph', 'cond-mat', 'cs.ET', 'cs.NI', 'eess.AS', 'math.CO', 'math.MP', 'nlin.CD', 'physics.comp-ph',
'q-alg', 'q-fin.MF', 'bayes-an', 'cs.AI', 'cs.FL', 'cs.OH', 'eess.IV', 'math.CT', 'math.NA', 'nlin.CG',
'physics.data-an', 'q-bio.BM', 'q-fin.PM', 'chao-dyn', 'cs.AR', 'cs.GL', 'cs.OS', 'eess.SP', 'math.CV',
'math.NT', 'nlin.PS', 'physics.ed-ph', 'q-bio.CB', 'q-fin.PR', 'chem-ph', 'cs.CC', 'cs.GR', 'cs.PF',
'eess.SY', 'math.DG', 'math.OA', 'nlin.SI', 'physics.flu-dyn', 'q-bio.GN', 'q-fin.RM']
filtered_classes = [53, 120, 65, 35, 37, 17, 49, 109, 97, 108]
# ['hep-ph', 'astro-ph', 'hep-th', 'quant-ph', 'cond-mat.mes-hall', 'gr-qc', 'cond-mat.mtrl-sci',
# 'cond-mat.str-el', 'cond-mat.stat-mech', 'astro-ph.SR']
additional_tokens_word_cl = [
['observations', 'distribution', 'temperature', 'luminosity', 'parameters', 'evolution', 'structure', 'galaxies',
'emission', 'neutrino', 'observed', 'magnetic', 'clusters', 'spectrum', 'velocity', 'galactic', 'redshift',
'obtained', 'symmetry', 'spectral', 'cluster', 'sources', 'optical', 'spectra', 'discuss', 'galaxy', 'masses',
'theory', 'higgs', 'gamma', 'quark', 'decay', 'gauge', 'solar', 'flux', 'qcd', 'gev', '$\\', '}$'],
['supersymmetric', 'entanglement', 'hamiltonian', 'measurement', 'interaction', 'dimensions', 'particular',
'conditions', 'operators', 'equations', 'functions', 'conformal', 'entangled', 'potential', 'solutions',
'classical', 'boundary', 'equation', 'possible', 'theories', 'approach', 'particle', 'solution', 'coupling',
'dynamics', 'algebra', 'systems', 'quantum', 'entropy', 'scalar', 'finite', 'photon', 'cavity', 'qubits', 'matrix',
'branes', 'scheme', 'qubit', 'brane', 'dual', '$-', ')$'],
['gravitational', 'cosmological', 'topological', 'conductance', 'scattering', 'transition', 'relativity',
'electronic', 'spacetime', 'numerical', 'transport', 'tunneling', 'parameter', 'electrons', 'einstein', 'constant',
'electron', 'universe', 'graphene', 'momentum', 'horizon', 'coupled', 'regime', 'tensor', 'metric', 'modes',
'dirac', 'noise'],
['antiferromagnetic', 'magnetization', 'ferromagnetic', 'measurements', 'experimental', 'calculations',
'temperatures', 'conductivity', 'interactions', 'correlation', 'structures', 'dependence', 'insulator',
'materials', 'exchange', 'orbital', 'studied', 'crystal', 'thermal', 'lattice', 'doping', 'phases', 'phonon',
'strain', 'kondo', 'films', 'fermi', 'bulk', '_2', '_3'],
['simulations', 'equilibrium', 'accretion', 'particles', 'diffusion', 'rotation', 'scaling', 'coronal', 'dwarf',
'flare']]
additional_tokens_word = ['antiferromagnetic', 'renormalization', 'superconducting', 'electromagnetic', 'supersymmetric',
'experimentally', 'characteristic', 'representation', 'configurations', 'gravitational',
'corresponding', 'approximation', 'ferromagnetic', 'magnetization', 'distributions',
'significantly', 'thermodynamic', 'understanding', 'contributions', 'characterized',
'perturbations', 'configuration', 'spectroscopic', 'observational', 'distribution', 'observations',
'measurements', 'interactions', 'experimental', 'entanglement', 'calculations', 'fluctuations',
'cosmological', 'polarization', 'temperatures', 'correlations', 'investigated', 'relativistic',
'applications', 'contribution', 'oscillations', 'conductivity', 'respectively', 'spectroscopy',
'perturbation', 'differential', 'intermediate', 'perturbative', 'temperature', 'interaction',
'topological', 'simulations', 'investigate', 'correlation', 'theoretical', 'equilibrium',
'demonstrate', 'experiments', 'transitions', 'hamiltonian', 'measurement', 'corrections',
'constraints', 'probability', 'interacting', 'generalized', 'predictions', 'fundamental',
'statistical', 'excitations', 'calculation', 'possibility', 'furthermore', 'conductance',
'instability', 'frequencies', 'metallicity', 'observation', 'anisotropic', 'numerically',
'observables', 'variability', 'transition', 'parameters', 'scattering', 'particular', 'electronic',
'conditions', 'consistent', 'dependence', 'structures', 'dimensions', 'considered', 'calculated',
'determined', 'absorption', 'luminosity', 'relaxation', 'components', 'previously', 'anisotropy',
'analytical', 'additional', 'excitation', 'transverse', 'correlated', 'comparison', 'dispersion',
'simulation', 'continuous', 'relativity', 'mechanical', 'techniques', 'amplitudes', 'difference',
'variations', 'experiment', 'asymptotic', 'structural', 'abundances', 'statistics', 'introduced',
'stochastic', 'symmetries', 'approaches', 'structure', 'potential', 'evolution', 'equations',
'classical', 'solutions', 'parameter', 'particles', 'transport', 'functions', 'numerical',
'discussed', 'dynamical', 'materials', 'processes', 'electrons', 'molecular', 'mechanism',
'agreement', 'presented', 'radiation', 'expansion', 'amplitude', 'spacetime', 'accretion',
'determine', 'calculate', 'diffusion', 'symmetric', 'magnitude', 'invariant', 'formalism',
'resulting', 'detection', 'arbitrary', 'operators', 'resonance', 'nonlinear', 'technique',
'insulator', 'conformal', 'tunneling', 'predicted', 'continuum', 'performed', 'algorithm',
'therefore', 'stability', 'abundance', 'variables', 'couplings', 'curvature', 'anomalous',
'mechanics', 'principle', 'increases', 'phenomena', 'behaviour', 'telescope', 'polarized',
'radiative', 'localized', 'influence', 'introduce', 'entangled', 'inflation', 'intrinsic',
'densities', 'magnetic', 'electron', 'observed', 'emission', 'coupling', 'symmetry', 'obtained',
'dynamics', 'particle', 'equation', 'approach', 'spectrum', 'graphene', 'galaxies', 'possible',
'velocity', 'spectral', 'presence', 'constant', 'theories', 'momentum', 'solution', 'boundary',
'clusters', 'neutrino', 'proposed', 'universe', 'recently', 'rotation', 'compared', 'strongly',
'previous', 'infrared', 'exchange', 'measured', 'describe', 'einstein', 'provides', 'geometry',
'extended', 'galactic', 'energies', 'addition', 'detected', 'increase', 'disorder', 'elements',
'scenario', 'fermions', 'coherent', 'estimate', 'detailed', 'contrast', 'rotating', 'metallic',
'redshift', 'identify', 'fraction', 'periodic', 'binaries', 'enhanced', 'suggests', 'gaussian',
'moreover', 'indicate', 'analyzed', 'exhibits', 'directly', 'hydrogen', 'infinite', 'profiles',
'surfaces', 'resolved', 'collapse', 'combined', 'impurity', 'examples', 'explicit', 'quantum',
'systems', 'optical', 'lattice', 'discuss', 'thermal', 'spectra', 'cluster', 'however', 'studied',
'entropy', 'provide', 'several', 'orbital', 'physics', 'sources', 'coupled', 'regions', 'scaling',
'various', 'applied', 'methods', 'further', 'spatial', 'propose', 'neutron', 'finally', 'studies',
'crystal', 'measure', 'horizon', 'diagram', 'analyze', 'compare', 'exhibit', 'compute', 'compact',
'fermion', 'degrees', 'explain', 'depends', 'coulomb', 'reduced', 'smaller', 'becomes', 'coronal',
'samples', 'observe', 'minimal', 'photons', 'kinetic', 'formula', 'appears', 'typical', 'towards',
'theory', 'finite', 'scalar', 'photon', 'regime', 'phases', 'matrix', 'obtain', 'masses', 'galaxy',
'derive', 'tensor', 'larger', 'vacuum', 'phonon', 'chiral', 'scheme', 'metric', 'scales', 'mixing',
'radial', 'signal', 'survey', 'plasma', 'cosmic', 'strain', 'cavity', 'dwarfs', 'layers', 'cannot',
'latter', 'doping', 'beyond', 'curves', 'reveal', 'appear', 'decays', 'moment', 'landau', 'occurs',
'solar', 'gauge', 'gamma', 'modes', 'decay', 'noise', 'higgs', 'fermi', 'ratio', 'exact', 'delta',
'quark', 'leads', 'dirac', 'atoms', 'qubit', 'sigma', 'quasi', 'dwarf', 'brane', 'films', 'monte',
'carlo', 'means', 'boson', 'omega', 'novel', 'basis', 'spins', 'fluid', 'gives', 'probe', 'kondo',
'giant', 'flare', 'doped', 'argue', 'disks', 'pairs', 'flux', 'bulk', 'dual', '^{-', 'qcd', 'tau',
'}$,', '+/-', '}$.', 'gev', 'ngc', '$\\', '}$', ')$', '$-', '_2', '2d', '_3', 'm_', '3d', '$)']
def remap_classes(example):
example['primary_cat'] = filtered_classes.index(example['primary_cat'])
return example
class CustomTrainer(Trainer):
def __init__(self, *args, eval_device='cpu', **kwargs):
super().__init__(*args, **kwargs)
self.eval_device = torch.device(eval_device)
def prediction_step(self, model, inputs, prediction_loss_only, ignore_keys=None):
"""https://github.com/huggingface/transformers/issues/7232"""
loss, logits, labels = super().prediction_step(model, inputs, prediction_loss_only, ignore_keys)
if prediction_loss_only:
return loss, None, None
else:
try:
return loss, logits.to(self.eval_device), labels.to(self.eval_device)
except RuntimeError:
print(loss, logits, labels)
def create_tokenizer(tokenizer_name, add_tokens=False):
tokenizer = AutoTokenizer.from_pretrained(tokenizer_name, cache_dir=cache_dir)
if add_tokens:
new_tokens_count = tokenizer.add_tokens(additional_tokens_word)
print(f"Added {new_tokens_count} tokens to the tokenizer. Total tokens to add were {len(additional_tokens_word)}.")
return tokenizer
def create_tokenizer_cl(tokenizer_name, exp_id, add_tokens=False):
tokenizer = AutoTokenizer.from_pretrained(tokenizer_name, cache_dir=cache_dir)
if add_tokens:
toadd = []
for i in range(exp_id+1):
toadd += additional_tokens_word_cl[i]
new_tokens_count = tokenizer.add_tokens(toadd)
print(f"Added {new_tokens_count} tokens to the tokenizer. Total tokens to add were {len(toadd)}.")
return tokenizer
def filter_by_classes(dataset, classes_filter):
return dataset.filter(lambda el: el['primary_cat'] in classes_filter)
def create_small_dataset(dataset, patterns_per_class=None, num_patterns=None,
patterns_per_class_per_test=0):
assert patterns_per_class or num_patterns
if patterns_per_class is None:
return dataset.select(range(num_patterns))
else:
classes_list = list(set(dataset['primary_cat']))
dts = []
dts_test = []
for class_id in classes_list:
dts_cls = filter_by_classes(dataset, [class_id])
assert len(dts_cls) > patterns_per_class_per_test + patterns_per_class
dts.append(dts_cls.select(range(patterns_per_class))) # take from beginning
if patterns_per_class_per_test > 0: # take from end
dts_test.append(dts_cls.select(range(len(dts_cls) - patterns_per_class_per_test, len(dts_cls))))
if patterns_per_class_per_test > 0:
return concatenate_datasets(dts), concatenate_datasets(dts_test)
else:
return concatenate_datasets(dts)
def load_generic_finetuning(use_test=False, add_tokens=False, exp_id=None, use_bert=False):
append_to_save_dir = f'{exp_id}_new_tokens' if add_tokens else ''
if use_bert:
append_to_save_dir += '_bert'
tr_d = Dataset.load_from_disk(os.path.join(save_path_generic, 'train_tokenized'+append_to_save_dir))
if use_test:
ts_d = Dataset.load_from_disk(os.path.join(save_path_generic, 'test_tokenized'+append_to_save_dir))
else:
ts_d = Dataset.load_from_disk(os.path.join(save_path_generic, 'valid_tokenized'+append_to_save_dir))
return tr_d.remove_columns(['text']), ts_d.remove_columns(['text'])
def load_qnli(add_tokens=False, exp_id=None, use_bert=False):
append_to_save_dir = f'{exp_id}_new_tokens' if add_tokens else ''
if use_bert:
append_to_save_dir += '_bert'
tr_d = Dataset.load_from_disk(os.path.join(save_path_qnli, 'train_tokenized'+append_to_save_dir))
ts_d = Dataset.load_from_disk(os.path.join(save_path_qnli, 'valid_tokenized'+append_to_save_dir))
return tr_d.remove_columns(['sentence', 'question', 'idx']).rename_column('label', 'labels'), \
ts_d.remove_columns(['sentence', 'question', 'idx']).rename_column('label', 'labels')
def compute_metrics(predictions):
logits, targets = predictions
predicted_classes = np.argmax(logits, axis=-1)
out = {}
out['accuracy'] = accuracy.compute(predictions=predicted_classes, references=targets)['accuracy']
out['f1'] = f1.compute(predictions=predicted_classes, references=targets, average='macro')['f1'] # per class average without weight
return out
def pretrain_model(args, tr_d, ts_d, model, tokenizer, eval_strategy, log_strategy, eval_only=False,
run_name=None):
"""Do not compute metrics other than loss to save memory space.
https://github.com/huggingface/transformers/issues/8143"""
model.train()
ckp = os.path.join(log_dir, "masked_checkpoint_dir")
training_args = TrainingArguments(ckp,
evaluation_strategy=eval_strategy, eval_steps=args.eval_every,
logging_strategy=log_strategy, logging_steps=args.log_every,
save_strategy=eval_strategy, save_steps=args.eval_every,
num_train_epochs=args.epochs,
per_device_train_batch_size=args.train_batch_size,
per_device_eval_batch_size=args.eval_batch_size,
learning_rate=args.lr,
weight_decay=args.weight_decay,
no_cuda=args.no_cuda,
report_to=["wandb"],
run_name=run_name,
load_best_model_at_end=True,
metric_for_best_model='eval_loss',
greater_is_better=False)
callbacks = []
callbacks.append(EarlyStoppingCallback(early_stopping_patience=2, early_stopping_threshold=0))
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=0.15)
trainer = CustomTrainer(model=model, args=training_args, train_dataset=tr_d,
eval_dataset=ts_d, data_collator=data_collator,
eval_device='cpu',
callbacks=callbacks)
if not eval_only:
trainer.train(resume_from_checkpoint=False)
trainer.evaluate()
def finetune_model(args, tr_d, ts_d, model, eval_strategy, log_strategy, eval_only=False,
run_name=None):
model.train()
ckp = os.path.join(log_dir, "checkpoint_dir")
training_args = TrainingArguments(ckp,
evaluation_strategy=eval_strategy, eval_steps=args.eval_every,
logging_strategy=log_strategy, logging_steps=args.log_every,
save_strategy=eval_strategy, save_steps=args.eval_every,
num_train_epochs=args.epochs,
per_device_train_batch_size=args.train_batch_size,
per_device_eval_batch_size=args.eval_batch_size,
learning_rate=args.lr,
weight_decay=args.weight_decay,
no_cuda=args.no_cuda,
report_to=["wandb"],
run_name=run_name,
load_best_model_at_end=True,
metric_for_best_model='eval_loss',
greater_is_better=False)
callbacks = []
callbacks.append(EarlyStoppingCallback(early_stopping_patience=2, early_stopping_threshold=0))
trainer = CustomTrainer(model=model, args=training_args, train_dataset=tr_d,
eval_dataset=ts_d, compute_metrics=compute_metrics,
eval_device='cpu',
callbacks=callbacks)
if not eval_only:
trainer.train(resume_from_checkpoint=False)
trainer.evaluate()
def preprocess_and_tokenize_by_task(preprocess=False, tokenize=False, tokenizer=None,
tokenize_field="abstract", append_to_save_dir=''):
assert preprocess or tokenize
if preprocess:
features = Features({tokenize_field: Value('string'),
'primary_cat': ClassLabel(num_classes=len(classes), names=classes),
'created': Value('int64')})
datas = []
class_to_idx = dict(zip(classes, list(range(len(classes)))))
for year in list(range(1993, 2020)):
data = read_csv(os.path.join(root_path, f'{year}.tsv'), sep='\t',
usecols=[tokenize_field, 'primary_cat', 'created'],
converters={'created': lambda el: int(el[:4]),
'primary_cat': lambda el: int(class_to_idx[el])})
datas.append(data)
data_all = concat(datas)
print("Read all data")
data_all = data_all.groupby('primary_cat').filter(lambda el: len(el) > 1)
train_data, test_data = train_test_split(data_all, test_size=0.15, shuffle=True,
stratify=data_all['primary_cat'])
pretrain_train, finetune_train = train_test_split(train_data, test_size=0.5, shuffle=True,
stratify=train_data['primary_cat'])
pretrain_test, finetune_test = train_test_split(test_data, test_size=0.5, shuffle=True,
stratify=test_data['primary_cat'])
print("Split train test completed")
pr_tr = Dataset.from_pandas(pretrain_train, features=features, split=Split.TRAIN)
pr_ts = Dataset.from_pandas(pretrain_test, features=features, split=Split.TEST)
f_tr = Dataset.from_pandas(finetune_train, features=features, split=Split.TRAIN)
f_ts = Dataset.from_pandas(finetune_test, features=features, split=Split.TEST)
pr_tr.save_to_disk(os.path.join(save_path, 'train', 'pretrain_task_filtered'))
pr_ts.save_to_disk(os.path.join(save_path, 'test', 'pretrain_task_filtered'))
f_tr.save_to_disk(os.path.join(save_path, 'train', 'finetuning_task_filtered'))
f_ts.save_to_disk(os.path.join(save_path, 'test', 'finetuning_task_filtered'))
print("Saved split data")
if tokenize:
def tokenize(examples):
return tokenizer(examples[tokenize_field], padding="max_length", truncation=True)
pr_tr = Dataset.load_from_disk(os.path.join(save_path, 'train', 'pretrain_task_filtered'))
pr_ts = Dataset.load_from_disk(os.path.join(save_path, 'test', 'pretrain_task_filtered'))
f_tr = Dataset.load_from_disk(os.path.join(save_path, 'train', 'finetuning_task_filtered'))
f_ts = Dataset.load_from_disk(os.path.join(save_path, 'test', 'finetuning_task_filtered'))
print("Loaded data")
pr_tr_tk = pr_tr.map(tokenize, batched=True, batch_size=1000)
pr_tr_tk.save_to_disk(os.path.join(save_path, 'train', 'tokenized', 'pretrain_task_filtered'+append_to_save_dir))
pr_ts_tk = pr_ts.map(tokenize, batched=True, batch_size=1000)
pr_ts_tk.save_to_disk(os.path.join(save_path, 'test', 'tokenized', 'pretrain_task_filtered'+append_to_save_dir))
f_tr_tk = f_tr.map(tokenize, batched=True, batch_size=1000)
f_tr_tk.save_to_disk(os.path.join(save_path, 'train', 'tokenized', 'finetuning_task_filtered'+append_to_save_dir))
f_ts_tk = f_ts.map(tokenize, batched=True, batch_size=1000)
f_ts_tk.save_to_disk(os.path.join(save_path, 'test', 'tokenized', 'finetuning_task_filtered'+append_to_save_dir))
print("Tokenization and saving completed")
class HGAvalancheDataset(AvalancheDataset):
def _process_pattern(self, element, idx: int):
pattern = element['input_ids']
label = element['labels']
pattern, label = self._apply_transforms(pattern, label)
return TupleTLabel((pattern, label, element['attention_mask'], element['token_type_ids'],
self.targets_task_labels[idx]))
def create_abstracts_avalanche_benchmark(experiences, add_tokens=False, test_on_test=False):
train_exps, test_exps = [], []
for exp_id, classes in enumerate(experiences):
append_to_save_dir = f'{exp_id}_new_tokens' if add_tokens and exp_id == 0 else f'{exp_id}'
tr_d = Dataset.load_from_disk(os.path.join(save_path_small, 'train', 'tokenized', 'finetuning_task_filtered'+append_to_save_dir))
if test_on_test:
ts_d = Dataset.load_from_disk(os.path.join(save_path_small, 'test', 'tokenized', 'finetuning_task_filtered'+append_to_save_dir))
else:
ts_d = Dataset.load_from_disk(os.path.join(save_path_small, 'valid', 'tokenized', 'finetuning_task_filtered'+append_to_save_dir))
tr_d = tr_d.map(remap_classes)
ts_d = ts_d.map(remap_classes)
tr_d = tr_d.remove_columns(['abstract', 'created']).rename_column('primary_cat', 'labels')
ts_d = ts_d.remove_columns(['abstract', 'created']).rename_column('primary_cat', 'labels')
tr_d.set_format(type="torch")
ts_d.set_format(type="torch")
tr_dataset = HGAvalancheDataset(tr_d, task_labels=0, dataset_type=avl.benchmarks.utils.AvalancheDatasetType.CLASSIFICATION,
targets=tr_d['labels'])
ts_dataset = HGAvalancheDataset(ts_d, task_labels=0, dataset_type=avl.benchmarks.utils.AvalancheDatasetType.CLASSIFICATION,
targets=ts_d['labels'])
train_exps.append(tr_dataset)
test_exps.append(ts_dataset)
benchmark = avl.benchmarks.dataset_benchmark(train_datasets=train_exps, test_datasets=test_exps,
dataset_type=avl.benchmarks.utils.AvalancheDatasetType.CLASSIFICATION)
return benchmark
class HGFeatureExtractorBackbone(avl.models.FeatureExtractorBackbone):
def forward(self, x):
self.model(input_ids=x[0],
attention_mask=x[1],
token_type_ids=x[2])
return self.output
def resize_token_embeddings(self, x):
self.model.resize_token_embeddings(x)
class HGStreamingLDA(avl.training.StreamingLDA):
@property
def mb_attention_mask(self):
return self.mbatch[2]
@property
def mb_token_type_ids(self):
return self.mbatch[3]
def forward(self, return_features=False):
self.model.eval()
feat = self.model([self.mb_x, self.mb_attention_mask, self.mb_token_type_ids])
out = self.predict(feat)
if return_features:
return out, feat
else:
return out
class HGBaseStrategy(avl.training.BaseStrategy):
@property
def mb_attention_mask(self):
return self.mbatch[2]
@property
def mb_token_type_ids(self):
return self.mbatch[3]
def forward(self):
out = self.model(input_ids=self.mb_x,
attention_mask=self.mb_attention_mask,
token_type_ids=self.mb_token_type_ids)
return out.logits
class HGJointTraining(avl.training.JointTraining):
@property
def mb_attention_mask(self):
return self.mbatch[2]
@property
def mb_token_type_ids(self):
return self.mbatch[3]
def forward(self):
out = self.model(input_ids=self.mb_x,
attention_mask=self.mb_attention_mask,
token_type_ids=self.mb_token_type_ids)
return out.logits
class ClassificationPresetTrain:
"""https://github.com/pytorch/vision/blob/main/references/classification/presets.py"""
def __init__(
self,
crop_size,
mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225),
interpolation=InterpolationMode.BILINEAR,
hflip_prob=0.5,
auto_augment_policy=None,
random_erase_prob=0.0,
vit=False
):
trans = [transforms.RandomResizedCrop(crop_size, interpolation=interpolation)]
if hflip_prob > 0:
trans.append(transforms.RandomHorizontalFlip(hflip_prob))
if auto_augment_policy is not None:
if auto_augment_policy == "ra":
trans.append(autoaugment.RandAugment(interpolation=interpolation))
elif auto_augment_policy == "ta_wide":
trans.append(autoaugment.TrivialAugmentWide(interpolation=interpolation))
else:
aa_policy = autoaugment.AutoAugmentPolicy(auto_augment_policy)
trans.append(autoaugment.AutoAugment(policy=aa_policy, interpolation=interpolation))
trans.extend(
[
transforms.PILToTensor(),
transforms.ConvertImageDtype(torch.float),
]
)
if not vit:
trans.extend([transforms.Normalize(mean=mean, std=std), ])
if random_erase_prob > 0:
trans.append(transforms.RandomErasing(p=random_erase_prob))
self.transforms = transforms.Compose(trans)
def __call__(self, img):
return self.transforms(img)
class ClassificationPresetEval:
"""https://github.com/pytorch/vision/blob/main/references/classification/presets.py"""
def __init__(
self,
crop_size,
resize_size=256,
mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225),
interpolation=InterpolationMode.BILINEAR,
vit=False
):
trans = [
transforms.Resize(resize_size, interpolation=interpolation),
transforms.CenterCrop(crop_size),
transforms.PILToTensor(),
transforms.ConvertImageDtype(torch.float),
]
if not vit:
trans.extend([transforms.Normalize(mean=mean, std=std), ])
self.transforms = transforms.Compose(trans)
def __call__(self, img):
return self.transforms(img)
def get_remove_idxs(d):
t = transforms.ToTensor()
remove_idxs = []
for i, el in enumerate(d):
tel = t(el[0])
if len(tel.size()) != 3 or tel.size(0) != 3:
remove_idxs.append(i)
return remove_idxs
class CustomExpDataTargets(IDatasetWithTargets):
def __init__(self, images, targets):
self.images = images
self.targets = targets
def __getitem__(self, index):
return self.images[index], self.targets[index]
def __len__(self):
return len(self.images)
def split_inaturalist(dataset, transforms_dict, test_size=0.25, target_type="super", in_memory=False):
remove_idxs = [23513, 68871, 68888, 72691, 102203, 108354, 108358, 108385, 108440, 111947, 111948, 111958, 111967,
115891, 117965, 125014, 144830, 144834, 144839, 144846, 145495, 145496, 146083, 155378, 155379,
155381, 155382, 155383, 155384, 155386, 155387, 155388, 155389, 155390, 155392, 155393, 200064,
224063, 238914, 247984, 264555, 264557, 264561, 264562, 264565, 264566, 264567, 264569, 264570,
264571, 265398, 267506, 270337, 270342, 270345, 270350, 270351, 270352, 270354, 270363, 270364,
270365, 270368, 270377, 271500, 271951, 272514, 273215, 273290, 273294, 273295, 273296, 273297,
273302, 273320, 273392, 273554, 273558, 273566, 273707, 273737, 273805, 273865, 274083, 274173,
274532, 275093, 275095, 275098, 275100, 275103, 275105, 275106, 275107, 275108, 275110, 275111,
275112, 275116, 275171, 275175, 275177, 275178, 275179, 275180, 275182, 275183, 275184, 275185,
275187, 275190, 275191, 275192, 275193, 275197, 275206, 275211, 275214, 275220, 275221, 275223,
275224, 275227, 275228, 275229, 275230, 275233, 275236, 275237, 275238, 275301, 275397, 275419,
275434, 275498, 275574, 275580, 275581, 275605, 275668, 275799, 275824, 275928, 275937, 275939,
275940, 275942, 275944, 275945, 275946, 275948, 275949, 275950, 275951, 275952, 275954, 275955,
275957, 275958, 275959, 275960, 275961, 275962, 275963, 276329, 276556, 276742, 276743, 276745,
276746, 276747, 276749, 276750, 276754, 276756, 276768, 276779, 276781, 276831, 276832, 276834,
276835, 276837, 276838, 276840, 276842, 276844, 276846, 276847, 276849, 276850, 276851, 276852,
276853, 278278, 278315, 278441, 278444, 278451, 278581, 278627, 278866, 278930, 278957, 278963,
278966, 279007, 279034, 279050, 279060, 279073, 279094, 279128, 279139, 279181, 279297, 279573,
279685, 279686, 279690, 279696, 279706, 279708, 279709, 279710, 279712, 279714, 279715, 279716,
279717, 279718, 279719, 279722, 279724, 280828, 280878, 281020, 281274, 281338, 281344, 281346,
281558, 281563, 282062, 282526, 283876, 284824, 284852, 284853, 284857, 284860, 284861, 284863,
285438, 285441, 285442, 285444, 290245, 290246, 290247, 290249, 290250, 290252, 290256, 290257,
290259, 290261, 290263, 290264, 290266, 290267, 290268, 290269, 290271, 290274, 291206, 291209,
291211, 300037, 300040, 300053, 300054, 300056, 308274, 309049, 309507, 327167, 338681, 451448,
453391, 453397]
if in_memory:
images = []
targets = []
print("Start reading Inaturalist images.")
for cat_id, fname in dataset.index:
with Image.open(os.path.join(dataset.root, dataset.all_categories[cat_id], fname)) as m:
images.append(m.copy())
targets.append(dataset.categories_map[cat_id][target_type])
print("Inaturalist images loaded in memory.")
dts = AvalancheDataset(dataset=CustomExpDataTargets(images, targets), task_labels=0,
dataset_type=AvalancheDatasetType.CLASSIFICATION,
transform=None)
else:
try:
with open(os.path.join(base_data, 'data', 'inaturalist_targets.pickle'), 'rb') as f:
targets = pickle.load(f)
except FileNotFoundError:
print("Inaturalist targets not found. Creating them now.")
targets = []
for cat_id, fname in dataset.index:
targets.append(dataset.categories_map[cat_id][target_type])
with open(os.path.join(base_data, 'data', 'inaturalist_targets.pickle'), 'wb') as f:
pickle.dump(targets, f)
dts = AvalancheDataset(dataset, task_labels=0, targets=targets,
dataset_type=AvalancheDatasetType.CLASSIFICATION)
train_idxs, test_idxs = train_test_split(range(len(targets)), stratify=targets, test_size=test_size, shuffle=True)
for el in remove_idxs:
if el in train_idxs:
train_idxs.remove(el)
if el in test_idxs:
test_idxs.remove(el)
if transforms_dict is not None:
ttr = transforms_dict['train']
tts = transforms_dict['val']
else:
ttr, tts = None, None
dtrain = AvalancheSubset(dts, train_idxs, dataset_type=AvalancheDatasetType.CLASSIFICATION,
transform=ttr)
dtest = AvalancheSubset(dts, test_idxs, dataset_type=AvalancheDatasetType.CLASSIFICATION,
transform=tts)
return dtrain, dtest
def create_downstream_model(model, hidden_size, n_classes, device, vit=False, path=None):
model_down = deepcopy(model)
if path is not None:
model_down.load_state_dict(torch.load(path))
if vit:
model_down.head = torch.nn.Linear(hidden_size, n_classes)
else:
model_down.fc = torch.nn.Linear(hidden_size, n_classes)
return model_down.to(device)
class UnsupBEiTSeqCLF(torch.nn.Module):
def __init__(self, num_classes, pretrained_path, pretrained_path_extractor, device='cpu'):
super().__init__()
self.feature_extractor = BeitFeatureExtractor.from_pretrained(
pretrained_path_extractor, cache_dir=os.path.join(base_data, 'data', 'beit'))
self.vit = BeitForImageClassification.from_pretrained(
pretrained_path, cache_dir=os.path.join(base_data, 'data', 'beit'),
num_labels=num_classes)
self.device = device
def forward(self, x):
if not isinstance(x, list):
x = list(x.cpu())
inputs = self.feature_extractor(images=x, return_tensors="pt") # 1024 features
outputs = self.vit(inputs['pixel_values'].to(self.device))
return outputs.logits
def save_pretrained(self, save_path):
self.vit.save_pretrained(save_path)
class UnsupBEiTLM(torch.nn.Module):
def __init__(self, pretrained_path, pretrained_path_extractor, device='cpu'):
super().__init__()
self.feature_extractor = BeitFeatureExtractor.from_pretrained(
pretrained_path_extractor, cache_dir=os.path.join(base_data, 'data', 'beit'))
self.vit = BeitForMaskedImageModeling.from_pretrained(
pretrained_path, cache_dir=os.path.join(base_data, 'data', 'beit'))
self.device = device
window_size = self.vit.beit.embeddings.patch_embeddings.patch_shape
num_masking_patches = 75
max_mask_patches_per_block = None
min_mask_patches_per_block = 16
# generating mask for the corresponding image
self.mask_generator = MaskingGenerator(
window_size, num_masking_patches=num_masking_patches,
max_num_patches=max_mask_patches_per_block,
min_num_patches=min_mask_patches_per_block)
self.encoder = load_model("https://cdn.openai.com/dall-e/encoder.pkl", device)
self.transforms_beit = transforms.Compose([
transforms.Resize((112, 112),
transforms.InterpolationMode.LANCZOS),
transforms.ToTensor(),
])
def forward(self, x):
if not isinstance(x, list):
x = list(x.cpu())
with torch.no_grad():
"""https://github.com/NielsRogge/Transformers-Tutorials/blob/master/BEiT/Understanding_BeitForMaskedImageModeling.ipynb"""
inputs = self.feature_extractor(images=x, return_tensors="pt")['pixel_values'].to(self.device)
xtrans = torch.stack([self.transforms_beit(el) for el in x], dim=0).to(self.device)
z_logits = self.encoder(map_pixels(xtrans))
input_ids = torch.argmax(z_logits, dim=1).flatten(1)
# bool_masked_pos = torch.from_numpy(self.mask_generator()).flatten().to(torch.bool)
# bool_masked_pos = bool_masked_pos.unsqueeze(0).repeat(input_ids.size(0), 1)
bool_masked_pos = torch.stack([torch.from_numpy(self.mask_generator()).flatten().to(torch.bool)
for _ in range(input_ids.size(0))], dim=0)
labels = input_ids[bool_masked_pos]
outputs = self.vit(inputs, bool_masked_pos, labels=labels) # batch_size, 1024, 512
logits = outputs.logits[bool_masked_pos]
return logits, outputs.loss
def save_pretrained(self, save_path):
self.vit.save_pretrained(save_path)
class UnsupBEiTNaive(avl.training.BaseStrategy):
def make_train_dataloader(self, num_workers=0, shuffle=True,
pin_memory=True, **kwargs):
""" Data loader initialization.
Called at the start of each learning experience after the dataset
adaptation.
:param num_workers: number of thread workers for the data loading.
:param shuffle: True if the data should be shuffled, False otherwise.
:param pin_memory: If True, the data loader will copy Tensors into CUDA
pinned memory before returning them. Defaults to True.
"""
self.dataloader = TaskBalancedDataLoader(
self.adapted_dataset,
oversample_small_groups=True,
num_workers=num_workers,
batch_size=self.train_mb_size,
shuffle=shuffle,
pin_memory=pin_memory,
**kwargs)
def make_eval_dataloader(self, num_workers=0, pin_memory=True,
**kwargs):
"""
Initializes the eval data loader.
:param num_workers: How many subprocesses to use for data loading.
0 means that the data will be loaded in the main process.
(default: 0).
:param pin_memory: If True, the data loader will copy Tensors into CUDA
pinned memory before returning them. Defaults to True.
:param kwargs:
:return:
"""
self.dataloader = torch.utils.data.DataLoader(
self.adapted_dataset,
num_workers=num_workers,
batch_size=self.eval_mb_size,
pin_memory=pin_memory,
**kwargs)
def _unpack_minibatch(self):
assert len(self.mbatch) >= 3
for i in range(1, len(self.mbatch)):
self.mbatch[i] = self.mbatch[i].to(self.device)
def training_epoch(self, **kwargs):
for self.mbatch in self.dataloader:
if self._stop_training:
break
self._unpack_minibatch()
self._before_training_iteration(**kwargs)
self.optimizer.zero_grad()
self.loss = 0
# Forward
self._before_forward(**kwargs)
self.mb_output, self.loss = self.forward()
self._after_forward(**kwargs)
self._before_backward(**kwargs)
self.loss.backward()
self._after_backward(**kwargs)
# Optimization step
self._before_update(**kwargs)
self.optimizer.step()
self._after_update(**kwargs)
self._after_training_iteration(**kwargs)
def eval_epoch(self, **kwargs):
for self.mbatch in self.dataloader:
self._unpack_minibatch()
self._before_eval_iteration(**kwargs)
self._before_eval_forward(**kwargs)
self.mb_output, self.loss = self.forward()
self._after_eval_forward(**kwargs)
self._after_eval_iteration(**kwargs)
class MaskingGenerator:
def __init__(
self, input_size, num_masking_patches, min_num_patches=4, max_num_patches=None,
min_aspect=0.3, max_aspect=None):
if not isinstance(input_size, tuple):
input_size = (input_size, ) * 2
self.height, self.width = input_size
self.num_patches = self.height * self.width
self.num_masking_patches = num_masking_patches
self.min_num_patches = min_num_patches
self.max_num_patches = num_masking_patches if max_num_patches is None else max_num_patches
max_aspect = max_aspect or 1 / min_aspect
self.log_aspect_ratio = (math.log(min_aspect), math.log(max_aspect))
def __repr__(self):
repr_str = "Generator(%d, %d -> [%d ~ %d], max = %d, %.3f ~ %.3f)" % (
self.height, self.width, self.min_num_patches, self.max_num_patches,
self.num_masking_patches, self.log_aspect_ratio[0], self.log_aspect_ratio[1])
return repr_str
def get_shape(self):
return self.height, self.width
def _mask(self, mask, max_mask_patches):
delta = 0
for attempt in range(10):
target_area = random.uniform(self.min_num_patches, max_mask_patches)
aspect_ratio = math.exp(random.uniform(*self.log_aspect_ratio))
h = int(round(math.sqrt(target_area * aspect_ratio)))
w = int(round(math.sqrt(target_area / aspect_ratio)))
if w < self.width and h < self.height:
top = random.randint(0, self.height - h)
left = random.randint(0, self.width - w)
num_masked = mask[top: top + h, left: left + w].sum()
# Overlap
if 0 < h * w - num_masked <= max_mask_patches:
for i in range(top, top + h):
for j in range(left, left + w):
if mask[i, j] == 0:
mask[i, j] = 1
delta += 1
if delta > 0:
break
return delta
def __call__(self):
mask = np.zeros(shape=self.get_shape(), dtype=np.int)
mask_count = 0
while mask_count < self.num_masking_patches:
max_mask_patches = self.num_masking_patches - mask_count
max_mask_patches = min(max_mask_patches, self.max_num_patches)
delta = self._mask(mask, max_mask_patches)
if delta == 0:
break
else:
mask_count += delta
return mask
def freeze_half_model(model, freeze, bert=False):
if not freeze:
return
modelname = 'bert' if bert else 'roberta'
base_model = getattr(model, modelname)
total_layers = len(base_model.encoder.layer)
layers_to_freeze = total_layers // 2
for p in base_model.embeddings.parameters():
p.requires_grad_(False)
for p in base_model.encoder.layer[:layers_to_freeze].parameters():
p.requires_grad_(False)
def freeze_model_but_classifier(model, linear_eval, head_name):
if not isinstance(head_name, (list, tuple)):
head_name = [head_name]
if not linear_eval:
return
for p in model.parameters():
p.requires_grad_(False)
for n, p in model.named_parameters():
if any([el in n for el in head_name]):
p.requires_grad_(True)
def json_save_dict(d, filepath):
with open(filepath, 'w') as f:
json.dump(d, f, indent=2)
def pickle_save_dict(d, filepath):
with open(filepath, 'wb') as f:
pickle.dump(d, f, protocol=pickle.HIGHEST_PROTOCOL)
class FixCORe50Dataset(avl.benchmarks.datasets.CORe50Dataset):
def _load_metadata(self) -> bool:
if self.mini:
bp = "core50_32x32"
else:
bp = "core50_128x128"
if not (self.root / bp).exists():
return False
if not (self.root / "batches_filelists").exists():
return False
with open(self.root / "paths.pkl", "rb") as f:
self.train_test_paths = pickle.load(f)
if self.verbose:
print("Loading labels...")
with open(self.root / "labels.pkl", "rb") as f:
self.all_targets = pickle.load(f)
self.train_test_targets = []
for i in range(self._nbatch + 1):
self.train_test_targets += self.all_targets[self._scen][
self._run
][i]
if self.verbose:
print("Loading LUP...")
with open(self.root / "LUP.pkl", "rb") as f:
self.LUP = pickle.load(f)
if self.verbose:
print("Loading labels names...")
with open(self.root / "labels2names.pkl", "rb") as f:
self.labels2names = pickle.load(f)
self.idx_list = []
if self.train:
for i in range(self._nbatch): # fix bug by removing + 1
self.idx_list += self.LUP[self._scen][self._run][i]
else:
self.idx_list = self.LUP[self._scen][self._run][-1]
self.paths = []
self.targets = []
for idx in self.idx_list:
self.paths.append(self.train_test_paths[idx])
div = 1
if not self.object_level:
div = 5
self.targets.append(self.train_test_targets[idx] // div)
with open(self.root / "labels2names.pkl", "rb") as f:
self.labels2names = pickle.load(f)
if not (self.root / "NIC_v2_79_cat").exists():
self._create_cat_filelists()
return True
nbatch = {
"ni": 8,
"nc": 9,
"nic": 79,
"nicv2_79": 79,
"nicv2_196": 196,
"nicv2_391": 391,
}
scen2dirs = {
"ni": "batches_filelists/NI_inc/",
"nc": "batches_filelists/NC_inc/",
"nic": "batches_filelists/NIC_inc/",
"nicv2_79": "NIC_v2_79/",
"nicv2_196": "NIC_v2_196/",
"nicv2_391": "NIC_v2_391/",
}
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
_default_train_transform = transforms.Compose(
[transforms.ToTensor(), transforms.RandomHorizontalFlip(), normalize]
)
_default_eval_transform = transforms.Compose([transforms.ToTensor(), normalize])
def InMemoryCORe50(*, scenario: str = "nicv2_391", val_perc=0, run: int = 0, object_lvl: bool = True, mini: bool = False,
train_transform=_default_train_transform, eval_transform=_default_eval_transform, dataset_root=None,
only_test=False):
assert 0 <= run <= 9, (
"Pre-defined run of CORe50 are only 10. Indicate "
"a number between 0 and 9."
)