forked from tianyizhou/DoCL
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdocl.py
1804 lines (1544 loc) · 78.2 KB
/
docl.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
# Copyright 2021-present, Tianyi Zhou ([email protected])
# All rights reserved.
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
from __future__ import print_function
import argparse
import os
import sys
import time
import shutil
import random
import copy
import math
import numpy as np
from numpy.testing import assert_array_almost_equal
from PIL import Image, ImageEnhance, ImageOps
from progress.bar import Bar as Bar
from sklearn.metrics import pairwise_distances
# from utils import CIFAR10PolicyAll, RandAugment
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.backends.cudnn as cudnn
import torch.optim as optim
import torch.utils.data as data
import torch.utils.data.sampler as sampler
import torchvision.transforms as transforms
import torchvision.datasets as datasets
import models.cifar as models
model_names = sorted(name for name in models.__dict__
if not name.startswith("__")
and callable(models.__dict__[name]))
parser = argparse.ArgumentParser(description='Curriculum Learning by Optimizing Learning Dynamics (DoCL)')
parser.add_argument('-d', '--dataset', type=str, default='CIFAR10')
# choices=['CIFAR10', 'CIFAR100', 'MNIST', 'FashionMNIST', 'SVHN', 'STL10'])
parser.add_argument('-spath', '--save_path', default='result', type=str, metavar='PATH',
help='path to save results')
parser.add_argument('-dpath', '--data_path', default='../data', type=str, metavar='PATH',
help='path to dataset directory')
parser.add_argument('--trialID', default='00/', type=str, metavar='PATH',
help='path to specific trial')
# Optimization parameters
parser.add_argument('--epochs', default=300, type=int, metavar='N',
help='total epochs to run')
parser.add_argument('--schedule', type=int, nargs='+', default=[0, 5, 10, 15, 20, 30, 40, 60, 90, 140, 210, 300],
help='splitting points (epoch number) for multiple episodes of training')
parser.add_argument('--selfsupervise_cut_epoch', default=200, type=int,
help='epoch to stop self-supervision (if consistency or contrastive > 0)')
parser.add_argument('--explore_cut_episode', default=5, type=int,
help='episode to stop update features, centrality, and increasing initial learning rate')
parser.add_argument('--train-batch', default=128, type=int, metavar='N',
help='train batchsize')
parser.add_argument('--test-batch', default=1024, type=int, metavar='N',
help='test batchsize')
parser.add_argument('--lr', '--learning-rate', default=2.0e-1, type=float,
metavar='LR', help='initial learning rate')
parser.add_argument('--lr_min', default=5.0e-4, type=float,
metavar='LR', help='ending learning rate of each episode')
parser.add_argument('--lr_min_decay', default=0.8, type=float,
help='decay factor for ending learning rate after each episode')
parser.add_argument('--lr_max', default=1.0e-1, type=float,
metavar='LR', help='starting learning rate')
parser.add_argument('--lr_max_decay', default=0.9, type=float,
help='decay factor applied to starting learning rate after each episode')
parser.add_argument('--momentum', default=0.9, type=float, metavar='M', help='momentum')
parser.add_argument('--weight-decay', '--wd', default=1e-4, type=float,
metavar='W', help='weight decay (default: 1e-4)')
parser.add_argument('--alpha', default=0.5, type=float,
help='beta distribution parameter for Mix-Up')
parser.add_argument('--alpha_rate', default=1.1, type=float,
help='rate for mixup alpha')
parser.add_argument('--batch_mul', default=4, type=int,
help='multiples of training batches during large-batch training')
# Neural Nets Architecture
parser.add_argument('--arch', '-a', metavar='ARCH', default='wrn')
parser.add_argument('--depth', type=int, default=28, help='Model depth.')
parser.add_argument('--repeat', type=int, default=3, help='number of blocks (WideResNet)')
parser.add_argument('--widen-factor', type=int, default=10, help='Widen factor. 4 -> 64, 8 -> 128, ...')
parser.add_argument('--drop', '--dropout', default=0.3, type=float, metavar='Dropout', help='Dropout rate')
# GPU and CPU
parser.add_argument('--gpu-id', default='0', type=str, help='id(s) for CUDA_VISIBLE_DEVICES')
parser.add_argument('-j', '--workers', default=4, type=int, metavar='N', help='number of data loading workers')
parser.add_argument('--manualSeed', type=int, help='manual seed')
parser.add_argument('--random_state', default=32, type=int, help='random state/seed')
# DoCL parameters
parser.add_argument('--k', default=1.0, type=float,
help='(initial) ratio for subset selected')
parser.add_argument('--dk', default=0.1, type=float,
help='increase rate of k')
parser.add_argument('--mk', default=0.3, type=float,
help='maximum/minimum of k')
parser.add_argument('--select_ratio', default=0.5, type=float,
help='(initial) ratio of further selection by centrality (when use_centrality=True)')
parser.add_argument('--select_ratio_rate', default=1.1, type=float,
help='multiplication factor applied to select_ratio for each episode')
parser.add_argument('--tmpt', default=0.1, type=float,
help='1/temperature used in computing softmax probability for subset sampling')
parser.add_argument('--tmpt_rate', default=0.9, type=float,
help='rate of temperature changing')
parser.add_argument('--mod', default=0.1, type=float,
help='(initial) weight of DoCL score (when use_centrality=True, since the weight for centrality is 1 - mod)')
parser.add_argument('--mod_rate', default=1.5, type=float,
help='multiplcation factor applied to mod every episode')
parser.add_argument('--ema_decay', default=0.5, type=float,
help='decay factor of exponential moving average')
parser.add_argument('--consistency', default=0., type=float,
help='weight of consistency regularization (provided by mean teacher) in the objective')
parser.add_argument('--consistency_rate', default=0.9, type=float,
help='multiplcation factor applied to the consistency weight every episode')
parser.add_argument('--contrastive', default=0., type=float,
help='weight of contrastive regularization (dictionary provided by mean teacher) in the objective')
parser.add_argument('--contrastive_rate', default=0.9, type=float,
help='multiplcation factor applied to the contrastive weight every episode')
# choices of different settings
parser.add_argument('--use_curriculum', action='store_true',
help='if using curriculum learning or not')
parser.add_argument('--use_mean_teacher', action='store_true',
help='maintain a time moving average of the model (ensemble over time) as a teacher for consistency regularization')
parser.add_argument('--use_random_subsample', action='store_true',
help='random subsample a subset before running centrality max to further select samples')
parser.add_argument('--use_centrality', action='store_true',
help='use centrality score in rating samples for more diversity')
parser.add_argument('--use_kernel_centrality', action='store_true',
help='compute centrality defined on a kernel matrix, if False, compute centrality based on the penultimate-layer features')
parser.add_argument('--use_noisylabel', action='store_true',
help='use training samples with noisy labels')
parser.add_argument('--label_noise_type', default='symmetric', type=str,
help='which type of label noise when --use_noisylabel: symmetric, pairflip')
parser.add_argument('--label_noise_rate', default=0.6, type=float,
help='noise rate on labels when --use_noisylabel')
parser.add_argument('--num_aug', default=0, type=int,
help='number of extra augmentations used for test/inference')
parser.add_argument('--save_dynamics', action='store_true',
help='save training dynamics (require large memory)')
args = parser.parse_args()
state = {k: v for k, v in args._get_kwargs()}
# Use CUDA
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_id
device_ids = list(map(int, args.gpu_id.split(',')))
device_ids = range(len(device_ids))
use_cuda = torch.cuda.is_available()
# Random seed
if args.manualSeed is None:
args.manualSeed = random.randint(1, 10000)
random.seed(args.manualSeed)
torch.manual_seed(args.manualSeed)
if use_cuda:
torch.cuda.manual_seed_all(args.manualSeed)
# name a folder by the trialID to store all the results
folder = args.trialID
def main():
global best_acc
if not os.path.isdir(args.save_path):
mkdir_p(args.save_path)
if not os.path.isdir(os.path.join(args.save_path, folder)):
mkdir_p(os.path.join(args.save_path, folder))
if not os.path.isdir(args.data_path):
mkdir_p(args.data_path)
# prepare datasets to train
print('==> Preparing dataset %s' % args.dataset)
if args.dataset == 'cifar10':
dataloader = datasets.CIFAR10
num_classes = 10
trans_mean = (0.4914, 0.4822, 0.4465)
trans_std = (0.2470, 0.2435, 0.2616)
cutout_size = 16
input_shape = (1, 3, 32, 32)
transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(5),
CIFAR10Policy(),
# CIFAR10PolicyAll(),
# RandAugment(),
cutout(cutout_size, 1.0, False),
transforms.ToTensor(),
transforms.Normalize(trans_mean, trans_std),
])
elif args.dataset == 'cifar100':
dataloader = datasets.CIFAR100
num_classes = 100
trans_mean = (0.5071, 0.4865, 0.4409)
trans_std = (0.2673, 0.2564, 0.2762)
cutout_size = 16
input_shape = (1, 3, 32, 32)
transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(5),
CIFAR10Policy(),
cutout(cutout_size, 1.0, False),
transforms.ToTensor(),
transforms.Normalize(trans_mean, trans_std),
])
elif args.dataset == 'svhn':
dataloader = datasets.SVHN
num_classes = 10
trans_mean = (0.5, 0.5, 0.5)
trans_std = (0.5, 0.5, 0.5)
cutout_size = 20
input_shape = (1, 3, 32, 32)
transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
cutout(cutout_size, 1.0, False),
transforms.ToTensor(),
transforms.Normalize(trans_mean, trans_std),
])
args.lr_max *= 0.1
args.lr_min *= 0.2
elif args.dataset == 'stl10':
dataloader = datasets.STL10
num_classes = 10
trans_mean = (0.5, 0.5, 0.5)
trans_std = (0.5, 0.5, 0.5)
cutout_size = 32
input_shape = (1, 3, 96, 96)
jitter_params = 0.4
light_params = 0.1
transform_train = transforms.Compose([
transforms.Pad(12),
transforms.RandomCrop(96),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(brightness=jitter_params, contrast=jitter_params, saturation=jitter_params, hue=0),
cutout(cutout_size, 1.0, False),
transforms.ToTensor(),
Lighting(light_params, _IMAGENET_PCA['eigval'], _IMAGENET_PCA['eigvec']),
transforms.Normalize(trans_mean, trans_std),
])
args.epochs *= 4
args.selfsupervise_cut_epoch *= 4
args.schedule = [sch*4 for sch in args.schedule]
print(args.epochs, args.schedule)
elif args.dataset == 'fmnist' or args.dataset == 'kmnist' or args.dataset == 'mnist':
if args.dataset == 'mnist':
dataloader = datasets.MNIST
trans_mean = np.array([0.1307])
trans_std = np.array([0.3081])
cutout_size = 8
elif args.dataset == 'fmnist':
dataloader = datasets.FashionMNIST
trans_mean = np.array([0.2860])
trans_std = np.array([0.3530])
cutout_size = 8
elif args.dataset == 'kmnist':
dataloader = datasets.KMNIST
trans_mean = np.array([0.1904])
trans_std = np.array([0.3475])
cutout_size = 14
num_classes = 10
input_shape = (1, 1, 28, 28)
args.lr_max *= 0.2
args.lr_min *= 0.2
transform_train = transforms.Compose([
# transforms.Pad(4),
# transforms.RandomCrop(28),
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
# transforms.RandomRotation(10),
cutout(cutout_size, 1.0, False),
transforms.ToTensor(),
transforms.Normalize(trans_mean, trans_std),
])
transform_train2 = TransformTwice(transform_train, args.use_mean_teacher)
transform_test = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(trans_mean, trans_std),
])
if args.dataset == 'stl10' or args.dataset == 'svhn':
trainset = dataloader(root=args.data_path, split='train', download=True, transform=transform_train2)
testset = dataloader(root=args.data_path, split='test', download=False, transform=transform_test)
trainset0 = dataloader(root=args.data_path, split='train', download=True, transform=transform_test)
if args.num_aug > 0:
testset_aug = dataloader(root=args.data_path, split='test', download=False, transform=transform_train)
trainset_aug = dataloader(root=args.data_path, split='train', download=False, transform=transform_train)
n_train = len(trainset.labels)
else:
trainset = dataloader(root=args.data_path+'/'+args.dataset, train=True, download=True, transform=transform_train2)
testset = dataloader(root=args.data_path+'/'+args.dataset, train=False, download=False, transform=transform_test)
trainset0 = dataloader(root=args.data_path+'/'+args.dataset, train=True, download=True, transform=transform_test)
if args.num_aug > 0:
testset_aug = dataloader(root=args.data_path+'/'+args.dataset, train=False, download=False, transform=transform_train)
trainset_aug = dataloader(root=args.data_path+'/'+args.dataset, train=True, download=False, transform=transform_train)
n_train = len(trainset.targets)
target_copy = trainset.targets
if args.use_noisylabel:
# noisy label
true_target_copy = target_copy.copy()
target_copy, actual_noise_rate = noisify(train_labels=np.asarray(target_copy), noise_type=args.label_noise_type, noise_rate=args.label_noise_rate, random_state=args.random_state, nb_classes=num_classes)
target_copy = target_copy.tolist()
trainset.targets = target_copy
noise_or_not_label = np.transpose(target_copy)==np.transpose(true_target_copy)
if use_cuda:
target_copy = torch.cuda.LongTensor(target_copy)
else:
target_copy = torch.LongTensor(target_copy)
trainloader_val = data.DataLoader(trainset0, batch_size=args.test_batch, shuffle=False, num_workers=args.workers)
testloader = data.DataLoader(testset, batch_size=args.test_batch, shuffle=False, num_workers=args.workers)
if args.num_aug > 0:
testloader_aug = data.DataLoader(testset_aug, batch_size=args.test_batch, shuffle=False, num_workers=args.workers)
trainloader_aug = data.DataLoader(trainset_aug, batch_size=args.test_batch, shuffle=False, num_workers=args.workers)
else:
testloader_aug = None
trainloader_aug = None
# Initialize Neural Nets Model
print("==> creating model '{}'".format(args.arch))
if args.dataset == 'fmnist' or args.dataset == 'kmnist' or args.dataset == 'mnist':
model = torch.nn.DataParallel(models.PreActResNet34_MNIST().cuda(), device_ids=device_ids)
if args.use_mean_teacher:
ema_model = torch.nn.DataParallel(models.PreActResNet34_MNIST().cuda(), device_ids=device_ids)
for param in ema_model.parameters():
param.detach_()
init_as_ema(ema_model, model)
else:
ema_model = None
else:
model = torch.nn.DataParallel(wrn(input_shape = input_shape, num_classes=num_classes, depth=args.depth,
widen_factor=args.widen_factor, repeat = args.repeat, dropRate=args.drop, bias=True).cuda(), device_ids=device_ids)
if args.use_mean_teacher:
ema_model = torch.nn.DataParallel(wrn(input_shape = input_shape, num_classes=num_classes, depth=args.depth,
widen_factor=args.widen_factor, repeat = args.repeat, dropRate=args.drop, bias=True).cuda(), device_ids=device_ids)
for param in ema_model.parameters():
param.detach_()
init_as_ema(ema_model, model)
else:
ema_model = None
print(' Total params: %.2fM' % (sum(p.numel() for p in model.parameters())/1000000.0))
optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay, nesterov=True)
cudnn.benchmark = True
criterion = nn.CrossEntropyLoss()
# initialize log files
title = args.dataset + args.arch
logger = Logger(os.path.join(args.save_path, folder + 'log.txt'), title=title)
logger.set_names(['Learning Rate', 'Train Loss', 'Valid Loss', 'Train Acc.', 'Valid Acc.'])
log_backup = open(os.path.join(args.save_path, folder + 'log_backup.txt'), 'w')
log_result = ()
# initialization of variables storing training results
best_acc = 0.0
schedule_idx = 0
train_sample_num = 0
batches_counter = 0
rand_train_set_old = None
batches_this_episode = np.ceil(float(n_train) / float(args.train_batch)) * (args.schedule[schedule_idx+1] - args.schedule[schedule_idx])
epoch_time = np.zeros(args.epochs)
decrease_budget = True if args.k > args.mk else False
all_pace = np.zeros(n_train)
selects = np.zeros(n_train)
if args.save_dynamics:
select_dy_train = ()
correct_dy_train = ()
loss_dy_train = ()
pred_dy_train = ()
correct_dy_test = ()
loss_dy_test = ()
pred_dy_test = ()
# training epochs
for epoch in range(args.epochs):
epoch_start_time = time.time()
# At the begining of each episode, before training begins, we need to update some training parameters
if epoch == args.schedule[min(schedule_idx + 1, len(args.schedule) - 1)]:
batches_counter = 0
schedule_idx += 1
if schedule_idx < args.explore_cut_episode:
args.lr_max /= args.lr_max_decay
else:
args.lr_max *= args.lr_max_decay
args.lr_min *= args.lr_min_decay
args.consistency *= args.consistency_rate
if args.consistency < 1.e-2 or epoch >= args.selfsupervise_cut_epoch:
args.consistency = 0.
args.contrastive *= args.contrastive_rate
if args.contrastive < 1.e-3 or epoch >= args.selfsupervise_cut_epoch:
args.contrastive = 0.
if args.use_curriculum:
# update k0 (random subsampling budget)
if decrease_budget:
args.k = max([args.k * (1.0-args.dk), args.mk])
else:
args.k = min([args.k * (1.0+args.dk), args.mk])
k0 = int(np.floor(args.k * n_train))
args.select_ratio = min(args.select_ratio_rate * args.select_ratio, 1.0)
# update subset size and trade-off weight between centrality and DoCL score
if args.use_centrality:
subset_size = int(args.select_ratio * k0)
batches_per_epoch = np.ceil(float(subset_size) / float(args.train_batch))
else:
batches_per_epoch = np.ceil(float(k0) / float(args.train_batch))
# update DoCL score's weight, softmax's temperature, and mixup's alpha
if decrease_budget:
args.mod *= min(0.8, args.mod_rate)
args.alpha = min(1.2, args.alpha * max(1.1, args.alpha_rate))
else:
args.mod *= max(1.2, args.mod_rate)
args.alpha = max(0.2, args.alpha * min(0.9, args.alpha_rate))
# update the size of selected subset and the number of training batches in the current episode
batches_this_episode = batches_per_epoch * (args.schedule[schedule_idx+1] - args.schedule[schedule_idx]) + np.ceil(float(n_train) / float(args.train_batch))
# one-pass inference of all training samples
train_loss, train_acc, all_loss, all_correct, _, train_fea, train_pred, _ = test(trainloader_val, trainloader_aug, model if ema_model is None else ema_model, criterion, epoch, args.num_aug, use_cuda)
# save dynamics
if args.save_dynamics:
all_loss_epoch = all_loss.cpu().numpy().astype(float)
diff_loss = np.minimum(np.abs(all_loss_epoch - all_loss_old), 10.0)
all_correct_epoch = all_correct.cpu().numpy().astype(float)
diff_flips = np.abs(all_correct_epoch - all_correct_old)
all_pred_epoch = train_pred.cpu().numpy()
loss_dy_train = loss_dy_train + (all_loss_epoch, )
correct_dy_train = correct_dy_train + (all_correct_epoch, )
pred_dy_train = pred_dy_train + (all_pred_epoch, )
# update the centrality of every training sample using pairwise similarity in the penultimate-layer feature space
# it measures how representative of each sample for the data distribution and encourages diversity in data selection
# it can be replaced by facility location function or other submodular functions if submodulax maximization is available
if args.use_centrality and epoch <= args.schedule[args.explore_cut_episode]:
modular_estimate = 1.0
centrality_estimate = 1.0
if args.use_kernel_centrality:
if epoch > args.schedule[1]:
del train_sims, centrality
train_sims, centrality = compute_sims(np.clip(train_fea, 1.0e-10, 1.0e+10), sigma = 20., metric = 'cos')
else:
train_fea, centrality = compute_feamat(np.clip(train_fea, 1.0e-10, 1.0e+10))
else:
batches_per_epoch = np.ceil(float(n_train) / float(args.train_batch))
batches_this_episode = batches_per_epoch * (args.schedule[schedule_idx+1] - args.schedule[schedule_idx])
# initialize mean teacher model by the current model before training begins
if args.use_mean_teacher:
if epoch == args.schedule[1]:
print('Initialize mean teacher!')
init_as_ema(ema_model, model)
else:
print('Re-initialize model by Ema!')
init_as_ema(model, ema_model, 1.e-2)
# warm starting episode (epochs from args.schedule[0] to args.schedule[1]) and large-batch training in DoCL: one full pass of all samples
if (not args.use_curriculum) or (epoch < args.schedule[1] or schedule_idx > 0 and epoch == args.schedule[min(schedule_idx, len(args.schedule) - 1)]):
print('\nEpoch: [%d | %d] LR: %f' % (epoch + 1, args.epochs, state['lr']))
# one-pass training on the whole training set
old_lr = args.lr
indices = np.random.permutation(n_train)
train_sample_num += n_train
trainloader = data.DataLoader(Subset(trainset, indices), batch_size=args.train_batch, shuffle=False, num_workers=args.workers)
if epoch < args.schedule[1]:
train_loss, train_acc, all_loss_per_sample, all_correct_per_sample, all_pred_per_sample, all_prob_per_sample, all_logit_per_sample, batches_counter, pace_delta = train(trainloader, model, ema_model, criterion, optimizer, batches_counter, batches_this_episode, use_cuda, args.ema_decay, 0., 0., 0.2, lr_schedule = 'linear')
elif args.use_curriculum and schedule_idx > 0 and epoch == args.schedule[min(schedule_idx, len(args.schedule) - 1)]:
train_loss, train_acc, all_loss_per_sample, all_correct_per_sample, all_pred_per_sample, all_prob_per_sample, all_logit_per_sample, batches_counter, pace_delta = train(trainloader, model, ema_model, criterion, optimizer, batches_counter, batches_this_episode, use_cuda, args.ema_decay, 0., 0., 0.2, lr_schedule = 'cosine', mul = args.batch_mul, mse_loss = True)
else:
train_loss, train_acc, all_loss_per_sample, all_correct_per_sample, all_pred_per_sample, all_prob_per_sample, all_logit_per_sample, batches_counter, pace_delta = train(trainloader, model, ema_model, criterion, optimizer, batches_counter, batches_this_episode, use_cuda, args.ema_decay, 0., 0., args.alpha, lr_schedule = 'cosine')
all_pace += pace_delta
all_prob_per_sample[indices] = all_prob_per_sample.clone()
all_logit_per_sample[indices] = all_logit_per_sample.clone()
if epoch > 0:
# time difference of logits
diff_logits = all_logit_per_sample - all_logit_old
# compute residual
res = -all_prob_per_sample.clone()
res[range(n_train), target_copy] += 1
# update docl score: acc_score
if epoch == 1:
acc_score = (res * diff_logits).sum(dim=1)#.abs()
else:
rc = torch.from_numpy(np.power(all_pace, 1.e-1)).float().to(diff_logits.get_device())
acc_score = args.ema_decay * acc_score + (1. - args.ema_decay) * (res * diff_logits).sum(dim=1) / ((n_train / rc.sum()) * rc)
# store the current feedback for next iteration comparison
all_logit_old = all_logit_per_sample.clone()
all_prob_old = all_prob_per_sample.clone()
all_pace[indices] = 0.0
# save dynamics
if args.save_dynamics:
all_loss_epoch = np.zeros(n_train)
all_loss_epoch[indices] = all_loss_per_sample.cpu().numpy().astype(float)
all_correct_epoch = np.zeros(n_train)
all_correct_epoch[indices] = all_correct_per_sample.cpu().numpy().astype(float)
all_pred_epoch = np.zeros(n_train)
all_pred_epoch[indices] = all_pred_per_sample.cpu().numpy()
loss_dy_train = loss_dy_train + (all_loss_epoch, )
correct_dy_train = correct_dy_train + (all_correct_epoch, )
pred_dy_train = pred_dy_train + (all_pred_epoch, )
# every episode except the warm starting episode (after epoch args.schedule[1]) selects a subset of data for training
if args.use_curriculum and epoch >= args.schedule[1]:
print('\nEpoch: [%d | %d] LR: %f, sampled, subset: %d, %d' % (epoch + 1, args.epochs, state['lr'], k0, subset_size))
# compute sampling probability
rand_prob = acc_score - acc_score.min()
rand_prob = rand_prob.cpu().numpy().astype(float)
rand_prob = np.exp(args.tmpt * (rand_prob - np.max(rand_prob)))
rand_prob /= np.sum(rand_prob)
print('rand_prob max, min: ', rand_prob.max(), rand_prob.min())
# sampling a subset according to rand_prob
if args.use_centrality:
mod = args.mod * (centrality_estimate / modular_estimate)
if args.use_random_subsample:
rand_train_set = np.random.choice(n_train, k0, p = rand_prob, replace=False)
train_subset = np.argpartition(mod * rand_prob[rand_train_set] + centrality[rand_train_set], -subset_size)[-subset_size:]
rand_train_set = rand_train_set[train_subset]
else:
p = mod * rand_prob + centrality
rand_train_set = np.random.choice(n_train, subset_size, p = p / p.sum(), replace=False)
centrality_estimate, modular_estimate = centrality[rand_train_set].sum(), rand_prob[rand_train_set].sum()
else:
rand_train_set = np.random.choice(n_train, k0, p = rand_prob, replace=False)
# train model on the selected subset rand_train_set
train_sample_num += len(rand_train_set)
trainselectloader = data.DataLoader(trainset, batch_size=args.train_batch, sampler=SubsetSampler(rand_train_set), num_workers=args.workers)
if rand_train_set_old is not None:
print('new samples selected comparing to previous epoch:', len(np.setdiff1d(rand_train_set, rand_train_set_old, assume_unique=True)))
rand_train_set_old = rand_train_set.copy()
old_lr = args.lr
train_loss1, train_acc, loss_rest, correct_rest, pred_rest, prob_rest, logit_rest, batches_counter, pace_delta2 = train(trainselectloader, model, ema_model, criterion, optimizer, batches_counter, batches_this_episode, use_cuda, args.ema_decay, args.consistency, args.contrastive, args.alpha, lr_schedule = 'cosine')
if epoch > args.schedule[min(schedule_idx, len(args.schedule)-1)]:
train_loss = train_loss1
# update after the training
selects[rand_train_set] += 1.0
all_pace += pace_delta2
diff_logits = logit_rest - all_logit_old[rand_train_set]
res = -prob_rest.clone()
res[range(len(rand_train_set)), target_copy[rand_train_set]] += 1
rc = torch.from_numpy(np.power(all_pace[rand_train_set], 1.e-1)).float().to(diff_logits.get_device())
acc_score[rand_train_set] = args.ema_decay * acc_score[rand_train_set] + (1. - args.ema_decay) * (res * diff_logits).sum(dim=1) / ((len(rand_train_set) / rc.sum()) * rc)
# store the current feedback for next iteration comparison
all_prob_old[rand_train_set] = prob_rest.clone()
all_logit_old[rand_train_set] = logit_rest.clone()
all_pace[rand_train_set] = 0.0
# save dynamics
if args.save_dynamics:
all_loss_epoch = all_loss_old.copy()
loss_rest = loss_rest.cpu().numpy().astype(float)
all_loss_epoch[rand_train_set] = loss_rest
all_correct_epoch = all_correct_old.copy()
all_correct_epoch[rand_train_set] = correct_rest.cpu().numpy().astype(float)
all_pred_epoch = all_pred_old.copy()
all_pred_epoch[rand_train_set] = pred_rest.cpu().numpy()
loss_dy_train = loss_dy_train + (all_loss_epoch, )
correct_dy_train = correct_dy_train + (all_correct_epoch, )
pred_dy_train = pred_dy_train + (all_pred_epoch, )
# inference on test set
epoch_time[epoch] = time.time() - epoch_start_time
test_loss, test_acc, all_loss_test, all_correct_test, _, test_fea, all_pred_test, _ = test(testloader, None, model, criterion, epoch, 0, use_cuda)
if args.use_mean_teacher:
ema_test_loss, ema_test_acc, _, _, _, _, _, _ = test(testloader, None, ema_model, criterion, epoch, 0, use_cuda)
if args.save_dynamics:
loss_dy_test = loss_dy_test + (all_loss_test.cpu().numpy().astype(float), )
correct_dy_test = correct_dy_test + (all_correct_test.cpu().numpy().astype(float), )
pred_dy_test = pred_dy_test + (all_pred_test.cpu().numpy(), )
select_dy_train = select_dy_train + (selects.copy(), )
torch.cuda.empty_cache()
# save results to log files
logger.append([state['lr'], train_loss, test_loss, train_acc, test_acc])
if args.use_mean_teacher:
result_iter = np.asarray([epoch, train_loss, test_loss, train_acc, test_acc, ema_test_loss, ema_test_acc, epoch_time[epoch], train_sample_num])
else:
result_iter = np.asarray([epoch, train_loss, test_loss, train_acc, test_acc, epoch_time[epoch], train_sample_num])
log_result = log_result + (result_iter, )
log_backup.write("\t".join(map(str, result_iter))+'\n')
log_backup.flush()
best_acc = np.max([best_acc, test_acc, ema_test_acc])
# save result and log files
log_backup.close()
logger.close()
np.savetxt(os.path.join(args.save_path, folder + 'log_result.txt'), np.vstack(log_result))
if args.save_dynamics:
np.savez(os.path.join(args.save_path, folder + args.dataset + '_dy.npz'),
loss_train = np.stack(loss_dy_train), loss_test = np.stack(loss_dy_test),
correct_train = np.stack(correct_dy_train), correct_test = np.stack(correct_dy_test),
pred_train = np.stack(pred_dy_train), pred_test = np.stack(pred_dy_test),
select_train = np.stack(select_dy_train), time_train = epoch_time)
print('Best acc:', best_acc)
#--------------------------------DATA AUTOAUGMENTATION-------------------------------
class CIFAR10Policy(object):
""" Randomly choose one of the best 25 Sub-policies on CIFAR10.
Example:
>>> policy = CIFAR10Policy()
>>> transformed = policy(image)
Example as a PyTorch Transform:
>>> transform=transforms.Compose([
>>> transforms.Resize(256),
>>> CIFAR10Policy(),
>>> transforms.ToTensor()])
"""
def __init__(self, fillcolor=(128, 128, 128)):
self.policies = [
SubPolicy(0.1, "invert", 7, 0.2, "contrast", 6, fillcolor),
SubPolicy(0.7, "rotate", 2, 0.3, "translateX", 9, fillcolor),
SubPolicy(0.8, "sharpness", 1, 0.9, "sharpness", 3, fillcolor),
SubPolicy(0.5, "shearY", 8, 0.7, "translateY", 9, fillcolor),
SubPolicy(0.5, "autocontrast", 8, 0.9, "equalize", 2, fillcolor),
SubPolicy(0.2, "shearY", 7, 0.3, "posterize", 7, fillcolor),
SubPolicy(0.4, "color", 3, 0.6, "brightness", 7, fillcolor),
SubPolicy(0.3, "sharpness", 9, 0.7, "brightness", 9, fillcolor),
SubPolicy(0.6, "equalize", 5, 0.5, "equalize", 1, fillcolor),
SubPolicy(0.6, "contrast", 7, 0.6, "sharpness", 5, fillcolor),
SubPolicy(0.7, "color", 7, 0.5, "translateX", 8, fillcolor),
SubPolicy(0.3, "equalize", 7, 0.4, "autocontrast", 8, fillcolor),
SubPolicy(0.4, "translateY", 3, 0.2, "sharpness", 6, fillcolor),
SubPolicy(0.9, "brightness", 6, 0.2, "color", 8, fillcolor),
SubPolicy(0.5, "solarize", 2, 0.0, "invert", 3, fillcolor),
SubPolicy(0.2, "equalize", 0, 0.6, "autocontrast", 0, fillcolor),
SubPolicy(0.2, "equalize", 8, 0.8, "equalize", 4, fillcolor),
SubPolicy(0.9, "color", 9, 0.6, "equalize", 6, fillcolor),
SubPolicy(0.8, "autocontrast", 4, 0.2, "solarize", 8, fillcolor),
SubPolicy(0.1, "brightness", 3, 0.7, "color", 0, fillcolor),
SubPolicy(0.4, "solarize", 5, 0.9, "autocontrast", 3, fillcolor),
SubPolicy(0.9, "translateY", 9, 0.7, "translateY", 9, fillcolor),
SubPolicy(0.9, "autocontrast", 2, 0.8, "solarize", 3, fillcolor),
SubPolicy(0.8, "equalize", 8, 0.1, "invert", 3, fillcolor),
SubPolicy(0.7, "translateY", 9, 0.9, "autocontrast", 1, fillcolor)
]
def __call__(self, img):
policy_idx = random.randint(0, len(self.policies) - 1)
return self.policies[policy_idx](img)
def __repr__(self):
return "AutoAugment CIFAR10 Policy"
class SVHNPolicy(object):
""" Randomly choose one of the best 25 Sub-policies on SVHN.
Example:
>>> policy = SVHNPolicy()
>>> transformed = policy(image)
Example as a PyTorch Transform:
>>> transform=transforms.Compose([
>>> transforms.Resize(256),
>>> SVHNPolicy(),
>>> transforms.ToTensor()])
"""
def __init__(self, fillcolor=(128, 128, 128)):
self.policies = [
SubPolicy(0.9, "shearX", 4, 0.2, "invert", 3, fillcolor),
SubPolicy(0.9, "shearY", 8, 0.7, "invert", 5, fillcolor),
SubPolicy(0.6, "equalize", 5, 0.6, "solarize", 6, fillcolor),
SubPolicy(0.9, "invert", 3, 0.6, "equalize", 3, fillcolor),
SubPolicy(0.6, "equalize", 1, 0.9, "rotate", 3, fillcolor),
SubPolicy(0.9, "shearX", 4, 0.8, "autocontrast", 3, fillcolor),
SubPolicy(0.9, "shearY", 8, 0.4, "invert", 5, fillcolor),
SubPolicy(0.9, "shearY", 5, 0.2, "solarize", 6, fillcolor),
SubPolicy(0.9, "invert", 6, 0.8, "autocontrast", 1, fillcolor),
SubPolicy(0.6, "equalize", 3, 0.9, "rotate", 3, fillcolor),
SubPolicy(0.9, "shearX", 4, 0.3, "solarize", 3, fillcolor),
SubPolicy(0.8, "shearY", 8, 0.7, "invert", 4, fillcolor),
SubPolicy(0.9, "equalize", 5, 0.6, "translateY", 6, fillcolor),
SubPolicy(0.9, "invert", 4, 0.6, "equalize", 7, fillcolor),
SubPolicy(0.3, "contrast", 3, 0.8, "rotate", 4, fillcolor),
SubPolicy(0.8, "invert", 5, 0.0, "translateY", 2, fillcolor),
SubPolicy(0.7, "shearY", 6, 0.4, "solarize", 8, fillcolor),
SubPolicy(0.6, "invert", 4, 0.8, "rotate", 4, fillcolor),
SubPolicy(0.3, "shearY", 7, 0.9, "translateX", 3, fillcolor),
SubPolicy(0.1, "shearX", 6, 0.6, "invert", 5, fillcolor),
SubPolicy(0.7, "solarize", 2, 0.6, "translateY", 7, fillcolor),
SubPolicy(0.8, "shearY", 4, 0.8, "invert", 8, fillcolor),
SubPolicy(0.7, "shearX", 9, 0.8, "translateY", 3, fillcolor),
SubPolicy(0.8, "shearY", 5, 0.7, "autocontrast", 3, fillcolor),
SubPolicy(0.7, "shearX", 2, 0.1, "invert", 5, fillcolor)
]
def __call__(self, img):
policy_idx = random.randint(0, len(self.policies) - 1)
return self.policies[policy_idx](img)
def __repr__(self):
return "AutoAugment SVHN Policy"
class SubPolicy(object):
def __init__(self, p1, operation1, magnitude_idx1, p2, operation2, magnitude_idx2, fillcolor=(128, 128, 128)):
Subranges = {
"shearX": np.linspace(0, 0.3, 10),
"shearY": np.linspace(0, 0.3, 10),
"translateX": np.linspace(0, 150 / 331, 10),
"translateY": np.linspace(0, 150 / 331, 10),
"rotate": np.linspace(0, 30, 10),
"color": np.linspace(0.0, 0.9, 10),
"posterize": np.round(np.linspace(8, 4, 10), 0).astype(np.int),
"solarize": np.linspace(256, 0, 10),
"contrast": np.linspace(0.0, 0.9, 10),
"sharpness": np.linspace(0.0, 0.9, 10),
"brightness": np.linspace(0.0, 0.9, 10),
"autocontrast": [0] * 10,
"equalize": [0] * 10,
"invert": [0] * 10
}
# from https://stackoverflow.com/questions/5252170/specify-image-filling-color-when-rotating-in-python-with-pil-and-setting-expand
def rotate_with_fill(img, magnitude):
rot = img.convert("RGBA").rotate(magnitude)
return Image.composite(rot, Image.new("RGBA", rot.size, (128,) * 4), rot).convert(img.mode)
Subfunc = {
"shearX": lambda img, magnitude: img.transform(
img.size, Image.AFFINE, (1, magnitude * random.choice([-1, 1]), 0, 0, 1, 0),
Image.BICUBIC, fillcolor=fillcolor),
"shearY": lambda img, magnitude: img.transform(
img.size, Image.AFFINE, (1, 0, 0, magnitude * random.choice([-1, 1]), 1, 0),
Image.BICUBIC, fillcolor=fillcolor),
"translateX": lambda img, magnitude: img.transform(
img.size, Image.AFFINE, (1, 0, magnitude * img.size[0] * random.choice([-1, 1]), 0, 1, 0),
fillcolor=fillcolor),
"translateY": lambda img, magnitude: img.transform(
img.size, Image.AFFINE, (1, 0, 0, 0, 1, magnitude * img.size[1] * random.choice([-1, 1])),
fillcolor=fillcolor),
"rotate": lambda img, magnitude: rotate_with_fill(img, magnitude),
# "rotate": lambda img, magnitude: img.rotate(magnitude * random.choice([-1, 1])),
"color": lambda img, magnitude: ImageEnhance.Color(img).enhance(1 + magnitude * random.choice([-1, 1])),
"posterize": lambda img, magnitude: ImageOps.posterize(img, magnitude),
"solarize": lambda img, magnitude: ImageOps.solarize(img, magnitude),
"contrast": lambda img, magnitude: ImageEnhance.Contrast(img).enhance(
1 + magnitude * random.choice([-1, 1])),
"sharpness": lambda img, magnitude: ImageEnhance.Sharpness(img).enhance(
1 + magnitude * random.choice([-1, 1])),
"brightness": lambda img, magnitude: ImageEnhance.Brightness(img).enhance(
1 + magnitude * random.choice([-1, 1])),
"autocontrast": lambda img, magnitude: ImageOps.autocontrast(img),
"equalize": lambda img, magnitude: ImageOps.equalize(img),
"invert": lambda img, magnitude: ImageOps.invert(img)
}
# self.name = "{}_{:.2f}_and_{}_{:.2f}".format(
# operation1, ranges[operation1][magnitude_idx1],
# operation2, ranges[operation2][magnitude_idx2])
self.p1 = p1
self.operation1 = Subfunc[operation1]
self.magnitude1 = Subranges[operation1][magnitude_idx1]
self.p2 = p2
self.operation2 = Subfunc[operation2]
self.magnitude2 = Subranges[operation2][magnitude_idx2]
def __call__(self, img):
if random.random() < self.p1: img = self.operation1(img, self.magnitude1)
if random.random() < self.p2: img = self.operation2(img, self.magnitude2)
return img
def mixup_data(x, y, alpha=1.0, use_cuda=True):
'''Returns mixed inputs, pairs of targets, and lambda'''
if alpha > 0:
lam = np.random.beta(alpha, alpha)
else:
lam = 1
batch_size = x.size()[0]
if use_cuda:
index = torch.randperm(batch_size).cuda()
else:
index = torch.randperm(batch_size)
mixed_x = lam * x + (1 - lam) * x[index, :]
y_a, y_b = y, y[index]
return mixed_x, y_a, y_b, lam, index
# -----------------------------DATASET and DATALOADER-----------------------------------
class Dataset(object):
"""An abstract class representing a Dataset.
All other datasets should subclass it. All subclasses should override
``__len__``, that provides the size of the dataset, and ``__getitem__``,
supporting integer indexing in range from 0 to len(self) exclusive.
"""
def __getitem__(self, index):
raise NotImplementedError
def __len__(self):
raise NotImplementedError
def __add__(self, other):
return ConcatDataset([self, other])
class SubsetSampler(sampler.Sampler):
"""Samples elements randomly from a given list of indices, without replacement.
Arguments:
indices (list): a list of indices
"""
def __init__(self, indices):
self.indices = indices
def __iter__(self):
return (self.indices[i] for i in range(len(self.indices)))
def __len__(self):
return len(self.indices)
class Subset(Dataset):
def __init__(self, dataset, indices):
self.dataset = dataset
self.indices = indices
def __getitem__(self, idx):
return self.dataset[self.indices[idx]]
def __len__(self):
return len(self.indices)
#------------------------------Neural Network Structures (WideResNet)--------------------------------
class BasicBlock(nn.Module):
def __init__(self, in_planes, out_planes, stride, dropRate=0.0):
super(BasicBlock, self).__init__()
self.bn1 = nn.BatchNorm2d(in_planes)#, momentum = 0.001
self.relu1 = nn.LeakyReLU(inplace=False)#, negative_slope = 0.02)
self.conv1 = nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(out_planes)#, momentum = 0.001
self.relu2 = nn.LeakyReLU(inplace=False)#, negative_slope = 0.02)
self.conv2 = nn.Conv2d(out_planes, out_planes, kernel_size=3, stride=1,
padding=1, bias=False)
self.droprate = dropRate
self.equalInOut = (in_planes == out_planes)
self.convShortcut = (not self.equalInOut) and nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride,
padding=0, bias=False) or None
def forward(self, x):
if not self.equalInOut:
x = self.relu1(self.bn1(x))
else:
out = self.relu1(self.bn1(x))
out = self.relu2(self.bn2(self.conv1(out if self.equalInOut else x)))
if self.droprate > 0:
out = F.dropout(out, p=self.droprate, training=self.training)
out = self.conv2(out)
return torch.add(x if self.equalInOut else self.convShortcut(x), out)
class NetworkBlock(nn.Module):
def __init__(self, nb_layers, in_planes, out_planes, block, stride, dropRate=0.0):
super(NetworkBlock, self).__init__()
self.layer = self._make_layer(block, in_planes, out_planes, nb_layers, stride, dropRate)
def _make_layer(self, block, in_planes, out_planes, nb_layers, stride, dropRate):
layers = []
for i in range(nb_layers):
layers.append(block(i == 0 and in_planes or out_planes, out_planes, i == 0 and stride or 1, dropRate))
return nn.Sequential(*layers)
def forward(self, x):
return self.layer(x)
class Flatten(nn.Module):
def __init__(self, d):
super(Flatten, self).__init__()
self.d = d
def forward(self, x):
x = x.view(x.size(0), -1)
return x
class WideResNet(nn.Module):
def __init__(self, input_shape, depth, num_classes, widen_factor=1, dropRate=0.0, repeat=3, bias=True):
super(WideResNet, self).__init__()
nChannels = [16]
if widen_factor > 20:
for ii in range(repeat):
nChannels.append(2**ii * widen_factor)
else:
for ii in range(repeat):
nChannels.append(2**ii * 16 * widen_factor)
assert (depth - 4) % 6 == 0, 'depth should be 6n+4'
n = (depth - 4) // 6
block = BasicBlock
# 1st conv before any network block
self.conv1 = nn.Conv2d(input_shape[1], nChannels[0], kernel_size=3, stride=1,
padding=1, bias=False)
self.blocks = [NetworkBlock(n, nChannels[0], nChannels[1], block, 1, dropRate)]
for ii in range(repeat - 1):
self.blocks.append(NetworkBlock(n, nChannels[ii+1], nChannels[ii+2], block, 2, dropRate))
self.blocks = nn.ModuleList(self.blocks)
self.bn1 = nn.BatchNorm2d(nChannels[-1])
self.relu = nn.LeakyReLU(inplace=True)
self.flatten = Flatten(nChannels[-1])
# compute conv feature size
with torch.no_grad():
self.feature_size = self._forward_conv(
torch.zeros(*input_shape)).reshape(-1).shape[0]
self.fc = nn.Linear(self.feature_size, num_classes, bias = bias)
self.nChannels = nChannels[-1]
for m in self.modules():
if isinstance(m, nn.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
elif isinstance(m, nn.Linear):
if bias:
m.bias.data.zero_()
def _forward_conv(self, x):
out = self.conv1(x)
for i, blk in enumerate(self.blocks):
out = blk(out)
out = self.relu(self.bn1(out))
out = F.adaptive_avg_pool2d(out, output_size=1)
return out
def forward(self, x):
x = self._forward_conv(x)
outfea = self.flatten(x)
x = self.fc(outfea)
return outfea, x
def wrn(**kwargs):
"""
Constructs a Wide Residual Networks.