-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataset.py
1516 lines (1311 loc) · 76 KB
/
dataset.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
# %%
"""The `denoise_model` will be our U-Net defined above. We'll employ the Huber loss between the true and the predicted noise.
## Define a PyTorch Dataset + DataLoader
Here we define a regular [PyTorch Dataset](https://pytorch.org/tutorials/beginner/basics/data_tutorial.html). The dataset simply consists of images from a real dataset, like Fashion-MNIST, CIFAR-10 or ImageNet, scaled linearly to \\([−1, 1]\\).
Each image is resized to the same size. Interesting to note is that images are also randomly horizontally flipped. From the paper:
> We used random horizontal flips during training for CIFAR10; we tried training both with and without flips, and found flips to improve sample quality slightly.
Here we use the 🤗 [Datasets library](https://huggingface.co/docs/datasets/index) to easily load the Fashion MNIST dataset from the [hub](https://huggingface.co/datasets/fashion_mnist). This dataset consists of images which already have the same resolution, namely 28x28.
"""
import glob
import os
import pathlib
from random import sample
from typing import Callable, List, Tuple, Union
from functools import lru_cache
import warnings
from datasets import load_dataset, concatenate_datasets
import datasets
from datasets.dataset_dict import DatasetDict
from matplotlib import pyplot as plt
import numpy as np
import torch
from torchvision import transforms
from torchvision.transforms import Compose, ToTensor, Lambda, ToPILImage, CenterCrop, Resize
from torchvision.utils import make_grid, save_image
from torch.utils.data import DataLoader, ConcatDataset, Subset, Dataset, IterableDataset
from torchvision.datasets import MNIST, CIFAR10, SVHN, FashionMNIST
from PIL import Image
from joblib import Parallel, delayed
from util import Log, normalize
DEFAULT_VMIN = float(-1.0)
DEFAULT_VMAX = float(1.0)
class DatasetLoader(object):
# Dataset generation mode
MODE_FIXED = "FIXED"
MODE_FLEX = "FLEX"
MODE_NONE = "NONE"
MODE_EXTEND = "EXTEND"
# Dataset names
MNIST = "MNIST"
CIFAR10 = "CIFAR10"
CELEBA = "CELEBA"
LSUN_CHURCH = "LSUN-CHURCH"
LSUN_BEDROOM = "LSUN-BEDROOM"
CELEBA_HQ = "CELEBA-HQ"
CELEBA_HQ_LATENT_PR05 = "CELEBA-HQ-LATENT_PR05"
CELEBA_HQ_LATENT = "CELEBA-HQ-LATENT"
# Inpaint Type
INPAINT_BOX: str = "INPAINT_BOX"
INPAINT_LINE: str = "INPAINT_LINE"
TRAIN = "train"
TEST = "test"
PIXEL_VALUES = "pixel_values"
PIXEL_VALUES_TRIGGER = "pixel_values_trigger"
TRIGGER = "trigger"
TARGET = "target"
IS_CLEAN = "is_clean"
R_trigger_only = "R_trigger_only"
IMAGE = "image"
LABEL = "label"
def __init__(self, name: str, label: int=None, root: str=None, channel: int=None, image_size: int=None, vmin: Union[int, float]=DEFAULT_VMIN, vmax: Union[int, float]=DEFAULT_VMAX, batch_size: int=512, shuffle: bool=True, seed: int=0):
self.__root = root
self.__name = name
if label != None and not isinstance(label, list)and not isinstance(label, tuple):
self.__label = [label]
else:
self.__label = label
self.__channel = channel
self.__vmin = vmin
self.__vmax = vmax
self.__batch_size = batch_size
self.__shuffle = shuffle
self.__dataset = self.__load_dataset(name=name)
self.__set_img_shape(image_size=image_size)
self.__trigger_type = self.__target_type = None
self.__trigger = self.__target = self.__poison_rate = self.__ext_poison_rate = None
self.__clean_rate = 1
self.__seed = seed
self.__rand_generator = torch.Generator()
self.__rand_generator.manual_seed(self.__seed)
if root != None:
self.__backdoor = Backdoor(root=root)
self.__R_trigger_only: bool = False
# self.__prep_dataset()
def set_poison(self, trigger_type: str, target_type: str, target_dx: int=-5, target_dy: int=-3, clean_rate: float=1.0, poison_rate: float=0.2, ext_poison_rate: float=0.0) -> 'DatasetLoader':
if self.__root == None:
raise ValueError("Attribute 'root' is None")
self.__clean_rate = clean_rate
self.__ext_poison_rate = ext_poison_rate
self.__poison_rate = poison_rate
self.__trigger_type = trigger_type
self.__target_type = target_type
self.__trigger = self.__backdoor.get_trigger(type=trigger_type, channel=self.__channel, image_size=self.__image_size, vmin=self.__vmin, vmax=self.__vmax)
self.__target = self.__backdoor.get_target(type=target_type, trigger=self.__trigger, dx=target_dx, dy=target_dy, vmin=self.__vmin, vmax=self.__vmax)
return self
def __load_dataset(self, name: str):
datasets.config.IN_MEMORY_MAX_SIZE = 50 * 2 ** 30
split_method = 'train+test'
if name == DatasetLoader.MNIST:
return load_dataset("mnist", split=split_method)
elif name == DatasetLoader.CIFAR10:
return load_dataset("cifar10", split=split_method)
elif name == DatasetLoader.CELEBA:
return load_dataset("student/celebA", split='train')
elif name == DatasetLoader.CELEBA_HQ:
# return load_dataset("huggan/CelebA-HQ", split=split_method)
return load_dataset("datasets/celeba_hq_256", split='train')
elif name == DatasetLoader.CELEBA_HQ_LATENT_PR05:
return load_from_disk("datasets/celeba_hq_256_pr05")
elif name == DatasetLoader.CELEBA_HQ_LATENT:
return LatentDataset(ds_root='datasets/celeba_hq_256_latents')
else:
raise NotImplementedError(f"Undefined dataset: {name}")
def __set_img_shape(self, image_size: int) -> None:
# Set channel
if self.__name == self.MNIST:
self.__channel = 1 if self.__channel == None else self.__channel
# self.__vmin = -1
# self.__vmax = 1
self.__cmap = "gray"
elif self.__name == self.CIFAR10 or self.__name == self.CELEBA or self.__name == self.CELEBA_HQ or self.__name == self.LSUN_CHURCH or self.__name == self.CELEBA_HQ_LATENT_PR05 or self.__name == self.CELEBA_HQ_LATENT:
self.__channel = 3 if self.__channel == None else self.__channel
# self.__vmin = -1
# self.__vmax = 1
self.__cmap = None
else:
raise NotImplementedError(f"No dataset named as {self.__name}")
# Set image size
if image_size == None:
if self.__name == self.MNIST:
self.__image_size = 32
elif self.__name == self.CIFAR10:
self.__image_size = 32
elif self.__name == self.CELEBA:
self.__image_size = 64
elif self.__name == self.CELEBA_HQ or self.__name == self.LSUN_CHURCH or self.__name == self.CELEBA_HQ_LATENT_PR05 or self.__name == self.CELEBA_HQ_LATENT:
self.__image_size = 256
else:
raise NotImplementedError(f"No dataset named as {self.__name}")
else:
self.__image_size = image_size
def __get_transform(self, prev_trans: List=[], next_trans: List=[]):
if self.__channel == 1:
channel_trans = transforms.Grayscale(num_output_channels=1)
elif self.__channel == 3:
channel_trans = transforms.Lambda(lambda x: x.convert("RGB"))
aug_trans = []
if self.__dataset != DatasetLoader.LSUN_CHURCH:
aug_trans = [transforms.RandomHorizontalFlip()]
trans = [channel_trans,
transforms.Resize([self.__image_size, self.__image_size]),
transforms.ToTensor(),
transforms.Lambda(lambda x: normalize(vmin_in=0, vmax_in=1, vmin_out=self.__vmin, vmax_out=self.__vmax, x=x)),
# transforms.Normalize([0.5], [0.5]),
] + aug_trans
return Compose(prev_trans + trans + next_trans)
# trans = [transforms.Resize(self.__image_size),
# transforms.ToTensor(),
# transforms.Lambda(lambda x: normalize(vmin=self.__vmin, vmax=self.__vmax, x=x))]
# return Compose(prev_trans + self.TRANSFORM_OPS + + next_trans)
def __fixed_sz_dataset_old(self):
gen = torch.Generator()
gen.manual_seed(self.__seed)
# Apply transformations
self.__full_dataset = self.__dataset.with_transform(self.__transform_generator(self.__name, True))
# Generate poisoned dataset
if self.__poison_rate > 0:
full_ds_len = len(self.__full_dataset[DatasetLoader.TRAIN])
perm_idx = torch.randperm(full_ds_len, generator=gen).long()
self.__poison_n = int(full_ds_len * float(self.__poison_rate))
self.__clean_n = full_ds_len - self.__poison_n
# print(f"perm_idx: {perm_idx}")
# print(f"len(perm_idx): {len(perm_idx)}, max: {torch.max(perm_idx)}, min: {torch.min(perm_idx)}")
# print(f"Clean n: {self.__clean_n}, Poison n: {self.__poison_n}")
self.__full_dataset[DatasetLoader.TRAIN] = Subset(self.__full_dataset[DatasetLoader.TRAIN], perm_idx[:self.__clean_n].tolist())
# print(f"Clean dataset len: {len(self.__full_dataset[DatasetLoader.TRAIN])}")
self.__backdoor_dataset = self.__dataset.with_transform(self.__transform_generator(self.__name, False))
self.__backdoor_dataset = Subset(self.__backdoor_dataset[DatasetLoader.TRAIN], perm_idx[self.__clean_n:].tolist())
# print(f"Backdoor dataset len: {len(self.__backdoor_dataset)}")
self.__full_dataset[DatasetLoader.TRAIN] = ConcatDataset([self.__full_dataset[DatasetLoader.TRAIN], self.__backdoor_dataset])
# print(f"self.__full_dataset[DatasetLoader.TRAIN] len: {len(self.__full_dataset[DatasetLoader.TRAIN])}")
self.__full_dataset = self.__full_dataset[DatasetLoader.TRAIN]
def manual_split():
pass
def __fixed_sz_dataset(self):
gen = torch.Generator()
gen.manual_seed(self.__seed)
if float(self.__poison_rate) < 0 or float(self.__poison_rate) > 1:
raise ValueError(f"In {DatasetLoader.MODE_FIXED}, poison rate should <= 1.0 and >= 0.0")
ds_n = len(self.__dataset)
backdoor_n = int(ds_n * float(self.__poison_rate))
ds_ls = []
# Apply transformations
if float(self.__poison_rate) == 0.0:
self.__clean_dataset = self.__dataset
self.__backdoor_dataset = None
elif float(self.__poison_rate) == 1.0:
self.__clean_dataset = None
self.__backdoor_dataset = self.__dataset
else:
full_dataset: datasets.DatasetDict = self.__dataset.train_test_split(test_size=backdoor_n)
self.__clean_dataset = full_dataset[DatasetLoader.TRAIN]
self.__backdoor_dataset = full_dataset[DatasetLoader.TEST]
if self.__clean_dataset != None:
clean_n = len(self.__clean_dataset)
self.__clean_dataset = self.__clean_dataset.add_column(DatasetLoader.IS_CLEAN, [True] * clean_n)
ds_ls.append(self.__clean_dataset)
# print(f"TRAIN IS_CLEAN N: {len(self.__full_dataset[DatasetLoader.TRAIN].filter(lambda x: x[DatasetLoader.IS_CLEAN]))}")
if self.__backdoor_dataset != None:
backdoor_n = len(self.__backdoor_dataset)
self.__backdoor_dataset = self.__backdoor_dataset.add_column(DatasetLoader.IS_CLEAN, [False] * backdoor_n)
ds_ls.append(self.__backdoor_dataset)
# print(f"TEST !IS_CLEAN N: {len(self.__full_dataset[DatasetLoader.TEST].filter(lambda x: not x[DatasetLoader.IS_CLEAN]))}")
def trans(x):
if x[DatasetLoader.IS_CLEAN][0]:
# print(f"IS_CLEAN: {x[DatasetLoader.IS_CLEAN]}")
return self.__transform_generator(self.__name, True, self.__R_trigger_only)(x)
return self.__transform_generator(self.__name, False, self.__R_trigger_only)(x)
self.__full_dataset = concatenate_datasets(ds_ls)
# print(f"IS_CLEAN N: {len(self.__full_dataset.filter(lambda x: x[DatasetLoader.IS_CLEAN]))}")
self.__full_dataset = self.__full_dataset.with_transform(trans)
# print(f"__full_dataset len: {len(self.__full_dataset)}, features: {self.__full_dataset.features}, keys: {self.__full_dataset[0].keys()}")
def __flex_sz_dataset_old(self):
# Apply transformations
self.__full_dataset = self.__dataset.with_transform(self.__transform_generator(self.__name, True))
full_ds_len = len(self.__full_dataset[DatasetLoader.TRAIN])
# Shrink the clean dataset
if self.__clean_rate != 1:
self.__clean_n = int(full_ds_len * float(self.__clean_rate))
self.__full_dataset[DatasetLoader.TRAIN] = Subset(self.__full_dataset[DatasetLoader.TRAIN], list(range(0, self.__clean_n, 1)))
# MODIFIED: Only 1 poisoned training sample
# self.__full_dataset[DatasetLoader.TRAIN] = Subset(self.__full_dataset[DatasetLoader.TRAIN], list(range(0, 1, 1)))
# Generate poisoned dataset
if self.__poison_rate > 0:
self.__backdoor_dataset = self.__dataset.with_transform(self.__transform_generator(self.__name, False))
self.__poison_n = int(full_ds_len * float(self.__poison_rate))
self.__backdoor_dataset = Subset(self.__backdoor_dataset[DatasetLoader.TRAIN], list(range(0, self.__poison_n, 1)))
self.__full_dataset[DatasetLoader.TRAIN] = ConcatDataset([self.__full_dataset[DatasetLoader.TRAIN], self.__backdoor_dataset])
# MODIFIED: Only 1 clean training sample
# self.__backdoor_dataset = Subset(self.__backdoor_dataset[DatasetLoader.TRAIN], list(range(0, 1, 1)))
# self.__full_dataset[DatasetLoader.TRAIN] = self.__backdoor_dataset
self.__full_dataset = self.__full_dataset[DatasetLoader.TRAIN]
def __flex_sz_dataset(self):
gen = torch.Generator()
gen.manual_seed(self.__seed)
def portion_sz(rate: float, n: int):
return int(n * float(rate))
def slice_ds(dataset, rate: float, ds_size: int):
if float(rate) == 0.0:
return None
elif float(rate) == 1.0:
return dataset
else:
return dataset.train_test_split(test_size=portion_sz(rate=rate, n=ds_size))[DatasetLoader.TEST]
ds_ls: List = []
ds_n = len(self.__dataset)
print(f"Total Dataset Size: {ds_n}")
# Apply transformations
self.__full_dataset: datasets.DatasetDict = self.__dataset.train_test_split()
clean_ds = slice_ds(dataset=self.__dataset, rate=float(self.__clean_rate), ds_size=ds_n)
if clean_ds is not None:
print(f"[Mode Flex] Clean Dataset Size: {len(clean_ds)}")
ds_ls.append(clean_ds.add_column(DatasetLoader.IS_CLEAN, [True] * portion_sz(rate=self.__clean_rate, n=ds_n)))
else:
print(f"[Mode Flex] Clean Dataset Size: 0")
backdoor_ds = slice_ds(dataset=self.__dataset, rate=float(self.__poison_rate), ds_size=ds_n)
if backdoor_ds is not None:
print(f"[Mode Flex] Backdoor Dataset Size: {len(backdoor_ds)}")
ds_ls.append(backdoor_ds.add_column(DatasetLoader.IS_CLEAN, [False] * portion_sz(rate=self.__poison_rate, n=ds_n)))
else:
print(f"[Mode Flex] Backdoor Dataset Size: 0")
# self.__full_dataset[DatasetLoader.TRAIN] = self.__full_dataset[DatasetLoader.TRAIN].add_column(DatasetLoader.IS_CLEAN, [True] * train_n)
# self.__full_dataset[DatasetLoader.TEST] = self.__full_dataset[DatasetLoader.TEST].add_column(DatasetLoader.IS_CLEAN, [False] * test_n)
def trans(x):
if x[DatasetLoader.IS_CLEAN][0]:
return self.__transform_generator(self.__name, True, self.__R_trigger_only)(x)
return self.__transform_generator(self.__name, False, self.__R_trigger_only)(x)
self.__full_dataset = concatenate_datasets(ds_ls)
self.__full_dataset = self.__full_dataset.with_transform(trans)
print(f"[Mode Flex] Full Dataset Size: {len(self.__full_dataset)}")
def __extend_sz_dataset(self):
gen = torch.Generator()
gen.manual_seed(self.__seed)
def portion_sz(rate: float, n: int):
return int(n * float(rate))
def slice_ds(dataset, rate: float, ds_size: int):
if float(rate) == 0.0:
return None
elif float(rate) == 1.0:
return dataset
elif float(rate) > 1.0:
mul: int = int(rate // 1)
mod: float = float(rate - mul)
cat_ds = [slice_ds(dataset, rate=1.0, ds_size=ds_size) for i in range(mul)]
if mod > 0:
cat_ds.append(slice_ds(dataset, rate=mod, ds_size=ds_size))
return concatenate_datasets(cat_ds)
else:
return dataset.train_test_split(test_size=portion_sz(rate=rate, n=ds_size))[DatasetLoader.TEST]
def trans(x):
# print(f"x[DatasetLoader.IS_CLEAN] len: {len(x[DatasetLoader.IS_CLEAN])}")
if x[DatasetLoader.IS_CLEAN][0]:
return self.__transform_generator(self.__name, True, x[DatasetLoader.R_trigger_only][0])(x)
return self.__transform_generator(self.__name, False, x[DatasetLoader.R_trigger_only][0])(x)
ds_ls: List = []
ds_n = len(self.__dataset)
ext_backdoor_n = int(ds_n * float(self.__ext_poison_rate))
print(f"Total Dataset Size: {ds_n}")
clean_dataset = ext_backdoor_dataset = backdoor_dataset = None
# Apply transformations
if float(self.__ext_poison_rate) == 0.0:
clean_dataset = self.__dataset
ext_backdoor_dataset = None
elif float(self.__ext_poison_rate) == 1.0:
clean_dataset = None
ext_backdoor_dataset = self.__dataset
else:
full_dataset: datasets.DatasetDict = self.__dataset.train_test_split(test_size=ext_backdoor_n)
clean_dataset = full_dataset[DatasetLoader.TRAIN]
ext_backdoor_dataset = full_dataset[DatasetLoader.TEST]
if clean_dataset != None:
clean_n = len(clean_dataset)
clean_dataset = clean_dataset.add_column(DatasetLoader.IS_CLEAN, [True] * clean_n).add_column(DatasetLoader.R_trigger_only, [False] * clean_n)
print(f"[Mode Extend] Clean Dataset Size: {len(clean_dataset)}, {clean_dataset[1].keys()}")
clean_dataset = clean_dataset.with_transform(trans)
ds_ls.append(clean_dataset)
else:
print(f"[Mode Extend] Clean Dataset Size: 0")
# print(f"TRAIN IS_CLEAN N: {len(self.__full_dataset[DatasetLoader.TRAIN].filter(lambda x: x[DatasetLoader.IS_CLEAN]))}")
if ext_backdoor_dataset != None:
ext_backdoor_n = len(ext_backdoor_dataset)
ext_backdoor_dataset = ext_backdoor_dataset.add_column(DatasetLoader.IS_CLEAN, [False] * ext_backdoor_n).add_column(DatasetLoader.R_trigger_only, [self.__ext_R_trigger_only] * ext_backdoor_n)
print(f"[Mode Extend] Extend Backdoor Dataset Size: {len(ext_backdoor_dataset)}, {ext_backdoor_dataset[1].keys()}")
ext_backdoor_dataset = ext_backdoor_dataset.with_transform(trans)
ds_ls.append(ext_backdoor_dataset)
else:
print(f"[Mode Extend] Extend Backdoor Dataset Size: 0")
# print(f"TEST !IS_CLEAN N: {len(self.__full_dataset[DatasetLoader.TEST].filter(lambda x: not x[DatasetLoader.IS_CLEAN]))}")
backdoor_dataset = slice_ds(dataset=self.__dataset, rate=float(self.__poison_rate), ds_size=ds_n)
if backdoor_dataset is not None:
backdoor_n = portion_sz(rate=self.__poison_rate, n=ds_n)
backdoor_dataset = backdoor_dataset.add_column(DatasetLoader.IS_CLEAN, [False] * backdoor_n).add_column(DatasetLoader.R_trigger_only, [self.__R_trigger_only] * backdoor_n)
print(f"[Mode Extend] Backdoor Dataset Size: {len(backdoor_dataset)}, {backdoor_dataset[1].keys()}")
backdoor_dataset = backdoor_dataset.with_transform(trans)
ds_ls.append(backdoor_dataset)
else:
print(f"[Mode Extend] Backdoor Dataset Size: 0")
# self.__full_dataset[DatasetLoader.TRAIN] = self.__full_dataset[DatasetLoader.TRAIN].add_column(DatasetLoader.IS_CLEAN, [True] * train_n)
# self.__full_dataset[DatasetLoader.TEST] = self.__full_dataset[DatasetLoader.TEST].add_column(DatasetLoader.IS_CLEAN, [False] * test_n)
self.__full_dataset = concatenate_datasets(ds_ls)
# self.__full_dataset = self.__full_dataset.with_transform(trans)
print(f"[Mode Extend] Full Dataset Size: {len(self.__full_dataset)}")
def prepare_dataset(self, mode: str="FIXED", R_trigger_only: bool=False, ext_R_trigger_only: bool=False, R_gaussian_aug: float=0.0) -> 'DatasetLoader':
self.__R_trigger_only = R_trigger_only
self.__ext_R_trigger_only = ext_R_trigger_only
self.__R_gaussian_aug = R_gaussian_aug
# Filter specified classes
if self.__label != None:
self.__dataset = self.__dataset.filter(lambda x: x[DatasetLoader.LABEL] in self.__label)
if mode == DatasetLoader.MODE_FIXED:
if self.__clean_rate != 1.0 or self.__clean_rate != None:
Log.warning("In 'FIXED' mode of DatasetLoader, the clean_rate will be ignored whatever.")
self.__fixed_sz_dataset()
elif mode == DatasetLoader.MODE_FLEX:
self.__flex_sz_dataset()
elif mode == DatasetLoader.MODE_EXTEND:
self.__extend_sz_dataset()
elif mode == DatasetLoader.MODE_NONE:
self.__full_dataset = self.__dataset
else:
raise NotImplementedError(f"Argument mode: {mode} isn't defined")
# Special Handling for LatentDataset
if self.__name == self.CELEBA_HQ_LATENT:
self.__full_dataset.set_poison(target_key=self.__target_type, poison_key=self.__trigger_type, raw='raw', poison_rate=self.__poison_rate, use_latent=True).set_use_names(target=DatasetLoader.TARGET, poison=DatasetLoader.PIXEL_VALUES, raw=DatasetLoader.IMAGE)
# Note the minimum and the maximum values
print(f"{self.__full_dataset[1].keys()}")
ex = self.__full_dataset[1][DatasetLoader.TARGET]
print(f"Dataset Len: {len(self.__full_dataset)}")
if len(ex) == 1:
print(f"Note that CHANNEL 0 - vmin: {torch.min(ex[0])} and vmax: {torch.max(ex[0])}")
elif len(ex) == 3:
print(f"Note that CHANNEL 0 - vmin: {torch.min(ex[0])} and vmax: {torch.max(ex[0])} | CHANNEL 1 - vmin: {torch.min(ex[1])} and vmax: {torch.max(ex[1])} | CHANNEL 2 - vmin: {torch.min(ex[2])} and vmax: {torch.max(ex[2])}")
return self
def get_dataset(self) -> datasets.Dataset:
return self.__full_dataset
def save_dataset(self, file: str):
self.__full_dataset.save_to_disk(file)
def get_dataloader(self, batch_size: int=None, shuffle: bool=None, num_workers: int=None, collate_fn: callable=None) -> torch.utils.data.DataLoader:
datasets = self.get_dataset()
if batch_size == None:
batch_size = self.__batch_size
if shuffle == None:
shuffle = self.__shuffle
if num_workers == None:
num_workers = 8
if collate_fn != None:
return DataLoader(datasets, batch_size=batch_size, shuffle=shuffle, pin_memory=True, num_workers=num_workers, collate_fn=collate_fn)
return DataLoader(datasets, batch_size=batch_size, shuffle=shuffle, pin_memory=True, num_workers=num_workers)
def get_mask(self, trigger: torch.Tensor) -> torch.Tensor:
return torch.where(trigger > self.__vmin, 0, 1)
def __transform_generator(self, dataset_name: str, clean: bool, R_trigger_only: bool) -> Callable[[torch.Tensor], torch.Tensor]:
if dataset_name == self.MNIST:
img_key = "image"
elif dataset_name == self.CIFAR10:
img_key = "img"
if dataset_name == self.CELEBA:
img_key = "image"
if dataset_name == self.CELEBA_HQ:
img_key = "image"
# define function
def clean_transforms(examples) -> DatasetDict:
if dataset_name == self.MNIST:
trans = self.__get_transform()
examples[DatasetLoader.IMAGE] = torch.stack([trans(image.convert("L")) for image in examples[img_key]])
else:
# trans = self.__get_transform(prev_trans=[transforms.PILToTensor()])
trans = self.__get_transform()
# trans = Compose([transforms.PILToTensor(), transforms.Lambda(lambda t: t / 255)])
examples[DatasetLoader.IMAGE] = torch.stack([trans(image) for image in examples[img_key]])
# examples[DatasetLoader.PIXEL_VALUES] = torch.tensor(np.array([np.asarray(image) / 255 for image in examples[img_key]])).permute(0, 3, 1, 2)
if img_key != DatasetLoader.IMAGE:
del examples[img_key]
examples[DatasetLoader.PIXEL_VALUES_TRIGGER] = torch.full_like(examples[DatasetLoader.IMAGE], 0)
examples[DatasetLoader.PIXEL_VALUES] = torch.full_like(examples[DatasetLoader.IMAGE], 0)
examples[DatasetLoader.TARGET] = torch.clone(examples[DatasetLoader.IMAGE])
data_shape = examples[DatasetLoader.PIXEL_VALUES].shape
repeat_times = (data_shape[0], *([1] * len(data_shape[1:])))
examples[DatasetLoader.TRIGGER] = self.__trigger.repeat(*repeat_times)
# examples[DatasetLoader.IS_CLEAN] = torch.tensor([True] * len(examples[DatasetLoader.PIXEL_VALUES]))
if DatasetLoader.LABEL in examples:
examples[DatasetLoader.LABEL] = torch.tensor([torch.tensor(x, dtype=torch.float) for x in examples[DatasetLoader.LABEL]])
else:
examples[DatasetLoader.LABEL] = torch.tensor([torch.tensor(-1, dtype=torch.float) for i in range(len(examples[DatasetLoader.PIXEL_VALUES]))])
# print(f"examples[img_key] Type: {type(examples[img_key])}")
# examples[img_key] = torch.tensor(np.array([np.asarray(image) / 255 for image in examples[img_key]])).permute(2, 0, 1)
# examples[img_key] = torch.stack([self.__get_transform()(np.asarray(image)) for image in examples[img_key]])
return examples
def backdoor_transforms(examples) -> DatasetDict:
examples = clean_transforms(examples)
data_shape = examples[DatasetLoader.PIXEL_VALUES].shape
repeat_times = (data_shape[0], *([1] * len(data_shape[1:])))
masks = self.get_mask(self.__trigger).repeat(*repeat_times)
# print(f"masks shape: {masks.shape} | examples[DatasetLoader.PIXEL_VALUES] shape: {examples[DatasetLoader.PIXEL_VALUES].shape} | self.__trigger.repeat(*repeat_times) shape: {self.__trigger.repeat(*repeat_times).shape}")
# examples[DatasetLoader.PIXEL_VALUES] = masks * examples[DatasetLoader.IMAGE] + (1 - masks) * self.__trigger.repeat(*repeat_times)
examples[DatasetLoader.PIXEL_VALUES_TRIGGER] = self.__trigger.repeat(*repeat_times)
if R_trigger_only:
examples[DatasetLoader.PIXEL_VALUES] = self.__trigger.repeat(*repeat_times)
else:
examples[DatasetLoader.PIXEL_VALUES] = masks * examples[DatasetLoader.IMAGE] + (1 - masks) * self.__trigger.repeat(*repeat_times)
# print(f"self.__target.repeat(*repeat_times) shape: {self.__target.repeat(*repeat_times).shape}")
examples[DatasetLoader.TARGET] = self.__target.repeat(*repeat_times)
# examples[DatasetLoader.IS_CLEAN] = torch.tensor([False] * data_shape[0])
return examples
if clean:
return clean_transforms
return backdoor_transforms
def get_poisoned(self, imgs) -> torch.Tensor:
data_shape = imgs.shape
repeat_times = (data_shape[0], *([1] * len(data_shape[1:])))
masks = self.get_mask(self.__trigger).repeat(*repeat_times)
return masks * imgs + (1 - masks) * self.__trigger.repeat(*repeat_times)
def get_inpainted(self, imgs, mask: torch.Tensor) -> torch.Tensor:
data_shape = imgs.shape
repeat_times = (data_shape[0], *([1] * len(data_shape[1:])))
notthing_tensor = torch.full_like(imgs, fill_value=torch.min(imgs))
masks = mask.repeat(*repeat_times)
return masks * imgs + (1 - masks) * notthing_tensor
def get_inpainted_boxes(self, imgs, up: int, low: int, left: int, right: int) -> torch.Tensor:
masked_val = 0
unmasked_val = 1
mask = torch.full_like(imgs[0], fill_value=unmasked_val)
if len(mask.shape) == 3:
mask[:, up:low, left:right] = masked_val
elif len(mask.shape) == 2:
mask[up:low, left:right] = masked_val
return self.get_inpainted(imgs=imgs, mask=mask)
def get_inpainted_by_type(self, imgs: torch.Tensor, inpaint_type: str) -> torch.Tensor:
if inpaint_type == DatasetLoader.INPAINT_LINE:
half_dim = imgs.shape[-1] // 2
up = half_dim - half_dim
low = half_dim + half_dim
left = half_dim - half_dim // 10
right = half_dim + half_dim // 20
return self.get_inpainted_boxes(imgs=imgs, up=up, low=low, left=left, right=right)
elif inpaint_type == DatasetLoader.INPAINT_BOX:
half_dim = imgs.shape[-1] // 2
up_left = half_dim - half_dim // 3
low_right = half_dim + half_dim // 3
return self.get_inpainted_boxes(imgs=imgs, up=up_left, low=low_right, left=up_left, right=low_right)
else:
raise NotImplementedError(f"inpaint: {inpaint_type} is not implemented")
def show_sample(self, img: torch.Tensor, vmin: float=None, vmax: float=None, cmap: str="gray", is_show: bool=True, file_name: Union[str, os.PathLike]=None, is_axis: bool=False) -> None:
cmap_used = self.__cmap if cmap == None else cmap
vmin_used = self.__vmin if vmin == None else vmin
vmax_used = self.__vmax if vmax == None else vmax
normalize_img = normalize(x=img, vmin_in=vmin_used, vmax_in=vmax_used, vmin_out=0, vmax_out=1)
channel_last_img = normalize_img.permute(1, 2, 0).reshape(self.__image_size, self.__image_size, self.__channel)
plt.imshow(channel_last_img, vmin=0, vmax=1, cmap=cmap_used)
# plt.imshow(img.permute(1, 2, 0).reshape(self.__image_size, self.__image_size, self.__channel), vmin=None, vmax=None, cmap=cmap_used)
# plt.imshow(img)
if not is_axis:
plt.axis('off')
plt.tight_layout()
if is_show:
plt.show()
if file_name != None:
save_image(normalize_img, file_name)
@property
def len(self):
return len(self.get_dataset())
def __len__(self):
return self.len
@property
def num_batch(self):
return len(self.get_dataloader())
@property
def trigger(self):
return self.__trigger
@property
def target(self):
return self.__target
@property
def name(self):
return self.__name
@property
def root(self):
return self.__root
@property
def batch_size(self):
return self.__batch_size
@property
def channel(self):
return self.__channel
@property
def image_size(self):
return self.__image_size
class Backdoor():
CHANNEL_LAST = -1
CHANNEL_FIRST = -3
GREY_BG_RATIO = 0.3
STOP_SIGN_IMG = "static/stop_sign_wo_bg.png"
# STOP_SIGN_IMG = "static/stop_sign_bg_blk.jpg"
CAT_IMG = "static/cat_wo_bg.png"
GLASSES_IMG = "static/glasses.png"
TARGET_FA = "SHOE"
TARGET_TG = "NOSHIFT"
TARGET_BOX = "CORNER"
# TARGET_BOX_MED = "BOX_MED"
TARGET_SHIFT = "SHIFT"
TARGET_HAT = "BWHAT"
TARGET_FEDORA_HAT = "HAT"
TARGET_CAT = "CAT"
TRIGGER_GAP_X = TRIGGER_GAP_Y = 2
TRIGGER_NONE = "NONE"
TRIGGER_FA = "FASHION"
TRIGGER_FA_EZ = "FASHION_EZ"
TRIGGER_MNIST = "MNIST"
TRIGGER_MNIST_EZ = "MNIST_EZ"
TRIGGER_SM_BOX = "SM_BOX"
TRIGGER_XSM_BOX = "XSM_BOX"
TRIGGER_XXSM_BOX = "XXSM_BOX"
TRIGGER_XXXSM_BOX = "XXXSM_BOX"
TRIGGER_BIG_BOX = "BIG_BOX"
TRIGGER_BIG_BOX_MED = "BOX_18"
TRIGGER_SM_BOX_MED = "BOX_14"
TRIGGER_XSM_BOX_MED = "BOX_11"
TRIGGER_XXSM_BOX_MED = "BOX_8"
TRIGGER_XXXSM_BOX_MED = "BOX_4"
TRIGGER_GLASSES = "GLASSES"
TRIGGER_BIG_STOP_SIGN = "STOP_SIGN_18"
TRIGGER_SM_STOP_SIGN = "STOP_SIGN_14"
TRIGGER_XSM_STOP_SIGN = "STOP_SIGN_11"
TRIGGER_XXSM_STOP_SIGN = "STOP_SIGN_8"
TRIGGER_XXXSM_STOP_SIGN = "STOP_SIGN_4"
# GREY_NORM_MIN = 0
# GREY_NORM_MAX = 1
def __init__(self, root: str):
self.__root = root
def __get_transform(self, channel: int, image_size: Union[int, Tuple[int]], vmin: Union[float, int], vmax: Union[float, int], prev_trans: List=[], next_trans: List=[]):
if channel == 1:
channel_trans = transforms.Grayscale(num_output_channels=1)
elif channel == 3:
channel_trans = transforms.Lambda(lambda x: x.convert("RGB"))
trans = [channel_trans,
transforms.Resize(image_size),
transforms.ToTensor(),
# transforms.Lambda(lambda x: normalize(vmin_out=vmin, vmax_out=vmax, x=x)),
transforms.Lambda(lambda x: normalize(vmin_in=0.0, vmax_in=1.0, vmin_out=vmin, vmax_out=vmax, x=x)),
# transforms.Lambda(lambda x: x * 2 - 1),
]
return Compose(prev_trans + trans + next_trans)
@staticmethod
def __read_img(path: Union[str, os.PathLike]):
return Image.open(path)
@staticmethod
def __bg2grey(trig, vmin: Union[float, int], vmax: Union[float, int]):
thres = (vmax - vmin) * Backdoor.GREY_BG_RATIO + vmin
trig[trig <= thres] = thres
return trig
@staticmethod
def __bg2black(trig, vmin: Union[float, int], vmax: Union[float, int]):
thres = (vmax - vmin) * Backdoor.GREY_BG_RATIO + vmin
trig[trig <= thres] = vmin
return trig
@staticmethod
def __white2grey(trig, vmin: Union[float, int], vmax: Union[float, int]):
thres = vmax - (vmax - vmin) * Backdoor.GREY_BG_RATIO
trig[trig >= thres] = thres
return trig
@staticmethod
def __white2med(trig, vmin: Union[float, int], vmax: Union[float, int]):
thres = vmax - (vmax - vmin) * Backdoor.GREY_BG_RATIO
trig[trig >= 0.7] = (vmax - vmin) / 2
return trig
def __get_img_target(self, path: Union[str, os.PathLike], image_size: int, channel: int, vmin: Union[float, int], vmax: Union[float, int]):
img = Backdoor.__read_img(path)
trig = self.__get_transform(channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)(img)
return Backdoor.__bg2grey(trig=trig, vmin=vmin, vmax=vmax)
def __get_img_trigger(self, path: Union[str, os.PathLike], image_size: int, channel: int, trigger_sz: int, vmin: Union[float, int], vmax: Union[float, int], x: int=None, y: int=None):
# Padding of Left & Top
l_pad = t_pad = int((image_size - trigger_sz) / 2)
r_pad = image_size - trigger_sz - l_pad
b_pad = image_size - trigger_sz - t_pad
residual = image_size - trigger_sz
if x != None:
if x > 0:
l_pad = x
r_pad = residual - l_pad
else:
r_pad = -x
l_pad = residual - r_pad
if y != None:
if y > 0:
t_pad = y
b_pad = residual - t_pad
else:
b_pad = -y
t_pad = residual - b_pad
img = Backdoor.__read_img(path)
next_trans = [transforms.Pad(padding=[l_pad, t_pad, r_pad, b_pad], fill=vmin)]
trig = self.__get_transform(channel=channel, image_size=trigger_sz, vmin=vmin, vmax=vmax, next_trans=next_trans)(img)
# thres = (vmax - vmin) * 0.3 + vmin
# trig[trig <= thres] = vmin
trig[trig >= 0.999] = vmin
# print(f"trigger shape: {trig.shape}")
return trig
@staticmethod
def __roll(x: torch.Tensor, dx: int, dy: int):
shift = tuple([0] * len(x.shape[:-2]) + [dy] + [dx])
dim = tuple([i for i in range(len(x.shape))])
return torch.roll(x, shifts=shift, dims=dim)
@staticmethod
def __get_box_trig(b1: Tuple[int, int], b2: Tuple[int, int], channel: int, image_size: int, vmin: Union[float, int], vmax: Union[float, int], val: Union[float, int]):
if isinstance(image_size, int):
img_shape = (image_size, image_size)
elif isinstance(image_size, list):
img_shape = image_size
else:
raise TypeError(f"Argument image_size should be either an integer or a list")
trig = torch.full(size=(channel, *img_shape), fill_value=vmin)
trig[:, b1[0]:b2[0], b1[1]:b2[1]] = val
return trig
@staticmethod
def __get_white_box_trig(b1: Tuple[int, int], b2: Tuple[int, int], channel: int, image_size: int, vmin: Union[float, int], vmax: Union[float, int]):
return Backdoor.__get_box_trig(b1=b1, b2=b2, channel=channel, image_size=image_size, vmin=vmin, vmax=vmax, val=vmax)
@staticmethod
def __get_grey_box_trig(b1: Tuple[int, int], b2: Tuple[int, int], channel: int, image_size: int, vmin: Union[float, int], vmax: Union[float, int]):
return Backdoor.__get_box_trig(b1=b1, b2=b2, channel=channel, image_size=image_size, vmin=vmin, vmax=vmax, val=(vmin + vmax) / 2)
@staticmethod
def __get_trig_box_coord(x: int, y: int):
if x < 0 or y < 0:
raise ValueError(f"Argument x, y should > 0")
return (- (y + Backdoor.TRIGGER_GAP_Y), - (x + Backdoor.TRIGGER_GAP_X)), (- Backdoor.TRIGGER_GAP_Y, - Backdoor.TRIGGER_GAP_X)
def get_trigger(self, type: str, channel: int, image_size: int, vmin: Union[float, int]=DEFAULT_VMIN, vmax: Union[float, int]=DEFAULT_VMAX) -> torch.Tensor:
if type == Backdoor.TRIGGER_FA:
trans = self.__get_transform(channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
ds = FashionMNIST(root=self.__root, train=True, download=True, transform=trans)
return Backdoor.__roll(Backdoor.__bg2black(trig=ds[0][0], vmin=vmin, vmax=vmax), dx=0, dy=2)
elif type == Backdoor.TRIGGER_FA_EZ:
trans = self.__get_transform(channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
ds = FashionMNIST(root=self.__root, train=True, download=True, transform=trans)
# Backdoor image ID: 135, 144
# return ds[144][0]
return Backdoor.__roll(Backdoor.__bg2black(trig=ds[144][0], vmin=vmin, vmax=vmax), dx=0, dy=4)
elif type == Backdoor.TRIGGER_MNIST:
trans = self.__get_transform(channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
ds = MNIST(root=self.__root, train=True, download=True, transform=trans)
# Backdoor image ID: 3, 6, 8
# return ds[3][0]
return Backdoor.__roll(Backdoor.__bg2black(trig=ds[3][0], vmin=vmin, vmax=vmax), dx=10, dy=3)
elif type == Backdoor.TRIGGER_MNIST_EZ:
trans = self.__get_transform(channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
ds = MNIST(root=self.__root, train=True, download=True, transform=trans)
# Backdoor image ID: 3, 6, 8
# return ds[6][0]
return Backdoor.__roll(Backdoor.__bg2black(trig=ds[6][0], vmin=vmin, vmax=vmax), dx=10, dy=3)
elif type == Backdoor.TRIGGER_SM_BOX:
b1, b2 = Backdoor.__get_trig_box_coord(14, 14)
# trig = torch.full(size=(channel, image_size, image_size), fill_value=vmin)
# trig[:, b1[0]:b2[0], b1[1]:b2[1]] = vmax
# return trig
return Backdoor.__get_white_box_trig(b1=b1, b2=b2, channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
elif type == Backdoor.TRIGGER_XSM_BOX:
b1, b2 = Backdoor.__get_trig_box_coord(11, 11)
# trig = torch.full(size=(channel, image_size, image_size), fill_value=vmin)
# trig[:, b1[0]:b2[0], b1[1]:b2[1]] = vmax
# return trig
return Backdoor.__get_white_box_trig(b1=b1, b2=b2, channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
elif type == Backdoor.TRIGGER_XXSM_BOX:
b1, b2 = Backdoor.__get_trig_box_coord(8, 8)
# trig = torch.full(size=(channel, image_size, image_size), fill_value=vmin)
# trig[:, b1[0]:b2[0], b1[1]:b2[1]] = vmax
# return trig
return Backdoor.__get_white_box_trig(b1=b1, b2=b2, channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
elif type == Backdoor.TRIGGER_XXXSM_BOX:
b1, b2 = Backdoor.__get_trig_box_coord(4, 4)
# trig = torch.full(size=(channel, image_size, image_size), fill_value=vmin)
# trig[:, b1[0]:b2[0], b1[1]:b2[1]] = vmax
# return trig
return Backdoor.__get_white_box_trig(b1=b1, b2=b2, channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
elif type == Backdoor.TRIGGER_BIG_BOX:
b1, b2 = Backdoor.__get_trig_box_coord(18, 18)
# trig = torch.full(size=(channel, image_size, image_size), fill_value=vmin)
# trig[:, b1[0]:b2[0], b1[1]:b2[1]] = vmax
# return trig
return Backdoor.__get_white_box_trig(b1=b1, b2=b2, channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
elif type == Backdoor.TRIGGER_BIG_BOX_MED:
b1, b2 = Backdoor.__get_trig_box_coord(18, 18)
return Backdoor.__get_grey_box_trig(b1=b1, b2=b2, channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
elif type == Backdoor.TRIGGER_SM_BOX_MED:
b1, b2 = Backdoor.__get_trig_box_coord(14, 14)
# trig = torch.full(size=(channel, image_size, image_size), fill_value=vmin)
# trig[:, b1[0]:b2[0], b1[1]:b2[1]] = (vmax + vmin) / 2
# return trig
return Backdoor.__get_grey_box_trig(b1=b1, b2=b2, channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
elif type == Backdoor.TRIGGER_XSM_BOX_MED:
b1, b2 = Backdoor.__get_trig_box_coord(11, 11)
# trig = torch.full(size=(channel, image_size, image_size), fill_value=vmin)
# trig[:, b1[0]:b2[0], b1[1]:b2[1]] = (vmax + vmin) / 2
# return trig
return Backdoor.__get_grey_box_trig(b1=b1, b2=b2, channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
elif type == Backdoor.TRIGGER_XXSM_BOX_MED:
b1, b2 = Backdoor.__get_trig_box_coord(8, 8)
# trig = torch.full(size=(channel, image_size, image_size), fill_value=vmin)
# trig[:, b1[0]:b2[0], b1[1]:b2[1]] = (vmax + vmin) / 2
# return trig
return Backdoor.__get_grey_box_trig(b1=b1, b2=b2, channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
elif type == Backdoor.TRIGGER_XXXSM_BOX_MED:
b1, b2 = Backdoor.__get_trig_box_coord(4, 4)
# trig = torch.full(size=(channel, image_size, image_size), fill_value=vmin)
# trig[:, b1[0]:b2[0], b1[1]:b2[1]] = (vmax + vmin) / 2
# return trig
return Backdoor.__get_grey_box_trig(b1=b1, b2=b2, channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
elif type == Backdoor.TRIGGER_GLASSES:
trigger_sz = int(image_size * 0.625)
return self.__get_img_trigger(path=Backdoor.GLASSES_IMG, image_size=image_size, channel=channel, trigger_sz=trigger_sz, vmin=vmin, vmax=vmax)
elif type == Backdoor.TRIGGER_BIG_STOP_SIGN:
return self.__get_img_trigger(path=Backdoor.STOP_SIGN_IMG, image_size=image_size, channel=channel, trigger_sz=18, vmin=vmin, vmax=vmax, x=-2, y=-2)
elif type == Backdoor.TRIGGER_SM_STOP_SIGN:
return self.__get_img_trigger(path=Backdoor.STOP_SIGN_IMG, image_size=image_size, channel=channel, trigger_sz=14, vmin=vmin, vmax=vmax, x=-2, y=-2)
elif type == Backdoor.TRIGGER_XSM_STOP_SIGN:
return self.__get_img_trigger(path=Backdoor.STOP_SIGN_IMG, image_size=image_size, channel=channel, trigger_sz=11, vmin=vmin, vmax=vmax, x=-2, y=-2)
elif type == Backdoor.TRIGGER_XXSM_STOP_SIGN:
return self.__get_img_trigger(path=Backdoor.STOP_SIGN_IMG, image_size=image_size, channel=channel, trigger_sz=8, vmin=vmin, vmax=vmax, x=-2, y=-2)
elif type == Backdoor.TRIGGER_XXXSM_STOP_SIGN:
return self.__get_img_trigger(path=Backdoor.STOP_SIGN_IMG, image_size=image_size, channel=channel, trigger_sz=4, vmin=vmin, vmax=vmax, x=-2, y=-2)
elif type == Backdoor.TRIGGER_NONE:
# trig = torch.zeros(channel, image_size, image_size)
trig = torch.full(size=(channel, image_size, image_size), fill_value=vmin)
return trig
else:
raise ValueError(f"Trigger type {type} isn't found")
def __check_channel(self, sample: torch.Tensor, channel_first: bool=None) -> int:
if channel_first != None:
# If user specified the localation of the channel
if self.__channel_first:
if sample.shape[Backdoor.CHANNEL_FIRST] == 1 or sample.shape[Backdoor.CHANNEL_FIRST] == 3:
return Backdoor.CHANNEL_FIRST
elif sample.shape[Backdoor.CHANNEL_LAST] == 1 or sample.shape[Backdoor.CHANNEL_LAST] == 3:
return Backdoor.CHANNEL_LAST
warnings.warn(Log.warning("The specified Channel doesn't exist, determine channel automatically"))
print(Log.warning("The specified Channel doesn't exist, determine channel automatically"))
# If user doesn't specified the localation of the channel or the
if (sample.shape[Backdoor.CHANNEL_LAST] == 1 or sample.shape[Backdoor.CHANNEL_LAST] == 3) and \
(sample.shape[Backdoor.CHANNEL_FIRST] == 1 or sample.shape[Backdoor.CHANNEL_FIRST] == 3):
raise ValueError(f"Duplicate channel found, found {sample.shape[Backdoor.CHANNEL_LAST]} at dimension 2 and {sample.shape[Backdoor.CHANNEL_FIRST]} at dimension 0")
if sample.shape[Backdoor.CHANNEL_LAST] == 1 or sample.shape[Backdoor.CHANNEL_LAST] == 3:
return Backdoor.CHANNEL_LAST
elif sample.shape[Backdoor.CHANNEL_FIRST] == 1 or sample.shape[Backdoor.CHANNEL_FIRST] == 3:
return Backdoor.CHANNEL_FIRST
else:
raise ValueError(f"Invalid channel shape, found {sample.shape[Backdoor.CHANNEL_LAST]} at dimension 2 and {sample.shape[Backdoor.CHANNEL_FIRST]} at dimension 0")
def __check_image_size(self, sample: torch.Tensor, channel_loc: int):
image_size = list(sample.shape)[-3:]
del image_size[channel_loc]
return image_size
def get_target(self, type: str, trigger: torch.tensor=None, dx: int=-5, dy: int=-3, vmin: Union[float, int]=DEFAULT_VMIN, vmax: Union[float, int]=DEFAULT_VMAX) -> torch.Tensor:
channel_loc = self.__check_channel(sample=trigger, channel_first=None)
channel = trigger.shape[channel_loc]
image_size = self.__check_image_size(sample=trigger, channel_loc=channel_loc)
print(f"image size: {image_size}")
if type == Backdoor.TARGET_TG:
if trigger == None:
raise ValueError("trigger shouldn't be none")
return Backdoor.__bg2grey(trigger.clone().detach(), vmin=vmin, vmax=vmax)
elif type == Backdoor.TARGET_SHIFT:
if trigger == None:
raise ValueError("trigger shouldn't be none")
# t_trig = trigger.clone().detach()
# shift = tuple([0] * len(t_trig.shape[:-2]) + [dy] + [dx])
# dim = tuple([i for i in range(len(t_trig.shape))])
# # print(f"Shift: {shift} | t_trig: {t_trig.shape}")
# return torch.roll(t_trig, shifts=shift, dims=dim)
return Backdoor.__bg2grey(Backdoor.__roll(trigger.clone().detach(), dx=dx, dy=dy), vmin=vmin, vmax=vmax)
# elif type == Backdoor.TARGET_BOX:
# # z = torch.full_like(trigger, fill_value=vmin)
# # z[:, 0:10, 0:10] = vmax
# # return z
# b1 = (None, None)
# b2 = (10, 10)
# return Backdoor.__get_white_box_trig(b1=b1, b2=b2, channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
elif type == Backdoor.TARGET_BOX:
b1 = (None, None)
b2 = (10, 10)
return Backdoor.__bg2grey(trig=Backdoor.__get_grey_box_trig(b1=b1, b2=b2, channel=channel, image_size=image_size, vmin=vmin, vmax=vmax), vmin=vmin, vmax=vmax)
elif type == Backdoor.TARGET_FA:
trans = self.__get_transform(channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
ds = FashionMNIST(root=self.__root, train=True, download=True, transform=trans)
# return ds[0][0]
return Backdoor.__bg2grey(trig=ds[0][0], vmin=vmin, vmax=vmax)
elif type == Backdoor.TARGET_HAT:
# img = Backdoor.__read_img("static/hat.png")
# trig = self.__get_transform(channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)(img)
# return trig
return self.__get_img_target(path="static/hat.png", channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
elif type == Backdoor.TARGET_FEDORA_HAT:
# img = Backdoor.__read_img("static/fedora-hat.png")
# trig = self.__get_transform(channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)(img)
# return trig
return self.__get_img_target(path="static/fedora-hat.png", channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
elif type == Backdoor.TARGET_CAT:
# img = Backdoor.__read_img("static/cat.png")
# trig = self.__get_transform(channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)(img)
# return trig
return self.__get_img_target(path=Backdoor.CAT_IMG, channel=channel, image_size=image_size, vmin=vmin, vmax=vmax)
else:
raise NotImplementedError(f"Target type {type} isn't found")
def show_image(self, img: torch.Tensor):
plt.axis('off')
plt.tight_layout()
plt.imshow(img.permute(1, 2, 0).squeeze(), cmap='gray')
plt.show()
class ReplicateDataset(torch.utils.data.Dataset):
def __init__(self, val: torch.Tensor, n: int):
self.__val: torch.Tensor = val
self.__n: int = n
self.__one_vec = [1 for i in range(self.__n)]
def __len__(self):
return self.__n
def __getitem__(self, slc):
n: int = len(self.__one_vec[slc])
reps = ([len(self.__val)] + ([1] * n))
return torch.squeeze((self.__val.repeat(*reps)))
class ImagePathDataset(torch.utils.data.Dataset):
IMAGE_EXTENSIONS = {'bmp', 'jpg', 'jpeg', 'pgm', 'png', 'ppm', 'tif', 'tiff', 'webp'}
# TRANSFORM = [transforms.ToTensor()]
def __init__(self, path, transforms=None, njobs: int=-1):
self.__path = pathlib.Path(path)
self.__files = sorted([file for ext in ImagePathDataset.IMAGE_EXTENSIONS
for file in self.__path.glob('*.{}'.format(ext))])
self.__transforms = transforms
self.__njobs = njobs