forked from eladhoffer/quantized.pytorch
-
Notifications
You must be signed in to change notification settings - Fork 1
/
Generator_m.py
1558 lines (1376 loc) · 63.5 KB
/
Generator_m.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
import os
import models
import shutil
import sys
import time
from datetime import datetime
import torch as th
import torch.nn as nn
import torch.backends.cudnn as cudnn
import torch.nn.functional as F
from torch.nn.utils import clip_grad_norm_
from torch import optim
import torchvision as tv
from torchvision.utils import save_image
import numpy as np
import matplotlib.pyplot as plt
from scipy import linalg
from PIL import Image
from ast import literal_eval
from apex import amp
from apex.parallel import SyncBatchNorm,convert_syncbn_model
from data import _DATASET_META_DATA, get_dataset
from utils.log import ResultsLog
from utils.absorb_bn import is_bn, search_absorbe_bn, get_bn_params
from utils.misc import _META, AutoArgParser, GaussianSmoothing
from utils.meters import AverageMeter, accuracy, ConfusionMeter
from utils.mixup import MixUp
from preprocess import get_transform
_VERSION=17
############## cfg
cudnn.benchmark=True
def settings():
record = 0
preset = ''
preset_sub_model = ''
if preset!='':
ckt_path = 'results/resnet44_cifar10/model_best.pth.tar'
dataset = 'cifar10'
model_config = "{'dataset': dataset,'depth':44}"
model='resnet'
else:
ckt_path=''
dataset=''
model_config=''
model=''
use_amp=0
result_root_dir = 'results'
exp_tag=''
measure='' #'imagine-cifar10-r44-dd_only_r500'
measure_steps=0 #500
measure_ds_limit=0
measure_seed=0
mixup=0
record_output_stats=0
measure_fid=''
fid_batch_size=512
fid_self=False
#split for measuring
split='train'
#reference split for fid
split_fid=split
output_temp=1.
adversarial=0
epsilon=0.1
smooth_sigma = 1
smooth_kernel = 5
report_freq = 20
batch_dup_factor = 4
use_stats_loss = 1
use_dd_loss = 0
use_prior_loss = 0
calc_stats_loss = 1
calc_cont_loss = 0
calc_smooth_loss = 1
dd_loss_mode='exp'
stats_loss_mode='kl'
stat_scale = 1.
cls_scale = 0.0004
smooth_scale = 1. #20
smooth_scale_decay={} #{100:0.2,5000:0.5,1000:0.5,1500:0.5}
batch_size = 256
betas = (0.9, 0.999)
lr = 0.1
replay_latent = 2000
n_samples_to_generate = 20
DEBUG_SHOW = 0
gen_resize_ratio=1.0
masking = False
SGD=0
target_stats=0
nclasses=-1
return locals()
parser=AutoArgParser()
parser.add_argument('-d',default=[0],type=int,nargs='+')
parser.add_argument('-lr_drop_replays',default=[800, 1500, 3000],type=int,nargs='+')
parser.add_argument('-snapshots_replay',default= [1000],type=int,nargs='+')
parser.add_argument('-stats_targets',default='pre_bn',choices=['pre_bn','post_bn','middle_bn'])
parser.auto_add(settings())
class _MODEL_META(_META):
_ATTRS = ['ds_meta','dataset','ckt_path','model','factory_method','model_config']
def __init__(self, **kwargs):
super().__init__(**kwargs)
_R18_IMGNT_Q4W4A=_MODEL_META(
ds_meta=_DATASET_META_DATA['imagenet'],
dataset='imagenet',
ckt_path='results/resnet18-Q4w4a_{}/model_best.pth.tar',
ckt_models={'CE':'imagenet_CE','dist':'imagenet_distilled',
'BNSKL':'imaginet_no_dd_kl_r1000',
'BNSKL_I':'imaginet_dd-exp_kl_r1000'},
model='resnet',
factory_method=models.resnet,
model_config={'depth': 18, 'conv1':{'a': 8,'w':8},'fc':{'a': 8,'w':8},
'activations_numbits':4,'weights_numbits':4, 'bias_quant':False,'quantize':True}
)
_R44_C10=_MODEL_META(
ds_meta = _DATASET_META_DATA['cifar10'],
dataset = 'cifar10',
ckt_path ='results/resnet44_cifar10/model_best.pth.tar',
model='resnet',
factory_method=models.resnet,
model_config={'dataset':'cifar10'}
)
_R44_C100=_MODEL_META(
ds_meta=_DATASET_META_DATA['cifar100'],
dataset='cifar100',
ckt_path ='results/resnet44_ba_m40_cifar100/model_best.pth.tar',
model='resnet',
factory_method=models.resnet,
model_config={'dataset':'cifar100','depth':44}
)
_R28_10_C100=_MODEL_META(
ds_meta=_DATASET_META_DATA['cifar100'],
dataset='cifar100',
ckt_path ='results/wresnet28-10_ba_m10_cifar100/checkpoint.pth.tar',
model='resnet',
factory_method=models.resnet,
model_config={'dataset':'cifar100','depth':28 ,'width':[160,320,640]}
)
_R18_IMGNT=_MODEL_META(
ds_meta=_DATASET_META_DATA['imagenet'],
dataset='imagenet',
ckt_path ='',
model='resnet',
factory_method=tv.models.resnet18,
model_config={'pretrained':True}
)
_DNS121_IMGNT=_MODEL_META(
ds_meta=_DATASET_META_DATA['imagenet'],
dataset='imagenet',
ckt_path ='',
model='densenet',
factory_method=getattr(tv.models,'densenet121'),
model_config={'pretrained':True}
)
_MBL2_IMGNT=_MODEL_META(
ds_meta=_DATASET_META_DATA['imagenet'],
dataset='imagenet',
ckt_path ='',
model='mobilenet_v2',
factory_method=getattr(tv.models,'mobilenet_v2'),
model_config={'pretrained':True}
)
_VGG16_BN_IMGNT=_MODEL_META(
ds_meta=_DATASET_META_DATA['imagenet'],
dataset='imagenet',
ckt_path ='',
model='vgg16_bn',
factory_method=getattr(tv.models,'vgg16_bn'),
model_config={'pretrained':True},
get_final_layer=lambda m: list(m._modules['classifier']._modules)[-1]
)
_VGG11_BN_IMGNT=_MODEL_META(
ds_meta=_DATASET_META_DATA['imagenet'],
dataset='imagenet',
ckt_path ='',
model='vgg11_bn',
factory_method=getattr(tv.models,'vgg11_bn'),
model_config={'pretrained':True},
get_final_layer=lambda m: list(m._modules['classifier']._modules)[-1]
)
_MODEL_CONFIGS={
'r44_cifar100':_R44_C100,
'r44_cifar10':_R44_C10,
'wr28-10_cifar100':_R28_10_C100,
'r18_imagenet':_R18_IMGNT,
'r18_q4w4a_imagenet':_R18_IMGNT_Q4W4A,
'densenet121_imagenet':_DNS121_IMGNT,
'mobilenet_v2_imagenet':_MBL2_IMGNT,
'vgg_16-bn_imagenet':_VGG16_BN_IMGNT,
'vgg_11-bn_imagenet': _VGG11_BN_IMGNT,
}
class GenLoader():
def __init__(self, input_data, nclasses, target_mode='random', limit=10):
self.data = input_data
self.nclasses = nclasses
self.target_mode = target_mode
self.limit = limit
class GenIterator():
_TARGET_MODES = ['random', 'running']
def __init__(self, input_data, nclasses, target_mode='random', limit=10):
assert target_mode in GenLoader.GenIterator._TARGET_MODES
self.data = input_data
self.nclasses = nclasses
self.target_mode = target_mode
self.iter = 0
self.iter_limit = limit
def reset_buffer(self):
# reset buffer
with th.no_grad():
self.data.detach()
self.data.normal_()
def __next__(self):
self.reset_buffer()
if self.iter > self.iter_limit:
raise StopIteration
if self.target_mode == 'running':
target = (th.arange(self.data.shape[0], device=self.data.device) + self.iter * self.data.shape[
0]) % self.nclasses
else:
# random mode should normally be used unless batch size is much larger then number of classes
target = th.randint(0, self.nclasses, (self.data.shape[0],), device=self.data.device)
self.iter += 1
return self.data, target
def __len__(self):
return self.iter_limit
def __iter__(self):
return GenLoader.GenIterator(self.data, self.nclasses, self.target_mode, self.limit)
class GussianSmoothingLoss(nn.Module):
def __init__(self,sigma=1,kernel_size=3,channels=3):
super(GussianSmoothingLoss,self).__init__()
self.smoothing_op=GaussianSmoothing(channels=channels, kernel_size=kernel_size, sigma=sigma)
self.loss_criterion=nn.MSELoss()
def forward(self, input,show=False):
smoothed=self.smoothing_op(input)
if show:
plot_grid(input)
plot_grid(smoothed)
# save_image(smoothed[:16].detach().cpu(), f'smoothed.jpg', nrow=4)
# save_image(image[:16].detach().cpu(), f'not_smoothed.jpg', nrow=4)
return self.loss_criterion(input, smoothed)
class ContinuityLoss(nn.Module):
def __init__(self,criterion=nn.MSELoss(reduction=None),diag=True,beta=0.1):
super(ContinuityLoss,self).__init__()
self.criterion=criterion
self.beta=beta
self.diag=diag
def forward(self,image):
lateral1 = image[:,:,:-1,:].contiguous().view(image.size(0),-1)
lateral2 = image[:,:,1:,:].contiguous().view(image.size(0),-1)
lat=self.criterion(lateral1,lateral2)
horizontal1 = image[:, :, :, :-1].contiguous().view(image.size(0),-1)
horizontal2 = image[:, :, :, 1:].contiguous().view(image.size(0),-1)
hor=self.criterion(horizontal1,horizontal2)
if self.diag:
diagonal1 = image[:, :, :-1, :-1].contiguous().view(image.size(0),-1)
diagonal2 = image[:, :, 1:, 1:].contiguous().view(image.size(0),-1)
diag=self.criterion(diagonal1,diagonal2)
# s=th.cat([lateral1,horizontal1,diagonal1],-1)
# t=th.cat([lateral2,horizontal2,diagonal2],-1)
else:
# s = th.cat([lateral1, horizontal1], -1)
# t = th.cat([lateral2, horizontal2], -1)
diag=0
return (hor+lat+diag).pow(self.beta).sum()
#return self.criterion(s,t)
class RandCropResize(nn.Module):
"""Randomly mask out one or more patches from an image.
Args:
n_holes (int): Number of patches to cut out of each image.
length (int): The length (in pixels) of each square patch.
"""
def __init__(self,min=0.33):
super(RandCropResize,self).__init__()
self.min = min
def forward(self, img):
"""
Args:
img (Tensor): Tensor image of size (C, H, W).
Returns:
Tensor: Image with n_holes of dimension length x length cut out of it.
"""
b=img.size(0)
h = img.size(2)
w = img.size(3)
min_h=int(self.min * h)
min_w=int(self.min * w)
hight = th.randint(min_h,h,(b,))
width = th.randint(min_w,w,(b,))
#center position
y = th.randint(min_h//2,h,(b,))
x = th.randint(min_w//2,w,(b,))
# size of crop around center
y1 = th.clamp(y - hight // 2, 0, h)
y2 = th.clamp(y + hight // 2, 0, h)
x1 = th.clamp(x - width // 2, 0, w)
x2 = th.clamp(x + width // 2, 0, w)
#
# img_ = img[:,y1: y2, x1: x2]
# img = th.nn.functional.upsample(img_,size=img.shape[1:],mode='bilinear').squeeze()
im = []
for i, (y1_, y2_, x1_, x2_) in enumerate(zip(y1, y2, x1, x2)):
im += [th.nn.functional.interpolate(img[i:i + 1, :, y1_: y2_, x1_: x2_], size=img.shape[2:],mode='bilinear')]
img=th.cat(im)
#mask = mask.expand_as(img)
#img = img * mask
return img
class Flip(nn.Module):
"""Randomly mask out one or more patches from an image.
Args:
prob (int): flip probability
"""
def __init__(self,prob=0.2):
super(Flip,self).__init__()
self.flip = prob
def forward(self, img):
"""
Args:
img (Tensor): Tensor image of size (C, H, W).
Returns:
flipped image w.p. flip
"""
#lateral
if th.rand(1) < self.flip:
img = img.flip(1)
#horizontal
if th.rand(1) < self.flip:
img = img.flip(2)
return img
class Cutout(nn.Module):
"""Randomly mask out one or more patches from an image.
Args:
n_holes (int): Number of patches to cut out of each image.
length (int): The length (in pixels) of each square patch.
"""
def __init__(self, max_num_holes=5,ratio=1/3):
super(Cutout,self).__init__()
self.max_num_holes = max_num_holes
self.ratio = ratio
def forward(self, img):
"""
Args:
img (Tensor): Tensor image of size (C, H, W).
Returns:
Tensor: Image with n_holes of dimension length x length cut out of it.
"""
b = img.size(0)
h = img.size(2)
w = img.size(3)
mask = th.ones((b,h,w),device=img.device)
for i in range(b):
for n in range(th.randint(self.max_num_holes,(1,))):
hight = th.randint(1,int(h * self.ratio),(1,))
width = th.randint(1,int(w * self.ratio),(1,))
y = th.randint(h,(1,))
x = th.randint(w,(1,))
y1 = th.clamp(y - hight // 2, 0, h)
y2 = th.clamp(y + hight // 2, 0, h)
x1 = th.clamp(x - width // 2, 0, w)
x2 = th.clamp(x + width // 2, 0, w)
mask[i:i+1,y1: y2, x1: x2] = 0.
mask = mask.unsqueeze(1).expand_as(img)
img = img * mask
return img
class AugmentImage(nn.Module):
_AUGS=['flip','crop','cutout']
def __init__(self,augmentations = {'flip':{},'crop':{},'cutout':{}}):
super(AugmentImage,self).__init__()
self.augs=[]
if 'flip' in augmentations:
self.flip = Flip(**augmentations['flip'])
self.augs.append(self.flip)
# else:
# self.Flipper = None
if 'crop' in augmentations:
self.crop = RandCropResize(**augmentations['crop'])
self.augs.append(self.crop)
# else:
# self.Cropper = None
if 'cutout' in augmentations:
self.cutout = Cutout(**augmentations['cutout'])
self.augs.append(self.cutout)
# else:
# self.Cutter = None
self.n_augs=len(augmentations)
def forward(self, input):
rand_aug_ids=th.randperm(self.n_augs).tolist()[:th.randint(self.n_augs,(1,)).item()]
i=input
for aug_id in rand_aug_ids:
aug=self.augs[aug_id]
i = aug(i)
return i
class AugmentBatch(nn.Module):
def __init__(self,dup_factor=8,aug_conf={}):
super(AugmentBatch,self).__init__()
self.dups=dup_factor
self.img_augmenter=AugmentImage(**aug_conf)
def forward(self,samples_,labels_,other=None):
samp = []
labs = [labels_]*self.dups
for _ in range(self.dups):
samp+=[self.img_augmenter(samples_)]
samples = th.cat(samp)
labels = th.cat(labs)
if other is None:
return samples,labels
others = th.cat([other] * self.dups)
return samples, (labels, others)
def freeze_params(model):
for param in model.parameters(True):
param.requires_grad = False
# P(mu1,var1)||P(mu2,var2)
def Gaussian_KL(mu1,var1,mu2,var2,epsilon=1e-5):
var1=var1.clamp(min=epsilon)
var2=var2.clamp(min=epsilon)
return 1/2*(-1 + th.log(var2/var1)+(var1+(mu1-mu2).pow(2))/var2)
def Gaussian_sym_KL(mu1,sigma1,mu2,sigma2,epsilon=1e-5):
return 0.5*(Gaussian_KL(mu1,sigma1,mu2,sigma2,epsilon)+Gaussian_KL(mu2,sigma2,mu1,sigma1,epsilon))
def reversedKLDNormal(mu1,var1,epsilon=1e-5):
return 0.5*(var1+mu1.pow(2) - th.log(var1.clamp(min=epsilon)) -1)
def forwardKLDNormal(mu2,var2,epsilon=1e-5):
return 0.5*(th.log(var2.clamp(min=epsilon)) + (1 + mu2.pow(2))/var2.clamp(min=epsilon) - 1)
def symKLDNormal(mu,sigma,epsilon=1e-5):
return 0.5*(forwardKLDNormal(mu,sigma,epsilon)+reversedKLDNormal(mu,sigma,epsilon))
def calc_stats_loss(loss_stat,stat_dict=None,running_dict=None,mode='kl',ref_stats_dict=None,inputs=None,verbose=0,epsilon=1e-8,pre_bn=True,record=None):
# used to aggregate running stats from multiple devices
def _get_stats_from_running_dict(stats_dict):
batch_statistics = {}
for k, v in stats_dict.items():
mean = v[0] / v[2]
var = v[1] / v[2] - mean.pow(2)
## var = E(x^2)-(EX)^2: sum_p2/n -sum/n
batch_statistics[k]=(mean, var)
if record is not None:
record.insert(k+'_batch_mean',mean)
record.insert(k+'_batch_var',var)
stats_dict.clear()
return batch_statistics
target_mean_key, target_var_key = None, None
batch_statistics = stat_dict or _get_stats_from_running_dict(running_dict)
if inputs is not None:
in_mu, in_std = inputs.mean((0, 2, 3)), inputs.var((0,2,3))
batch_statistics['inputs']=(in_mu, in_std)
if mode == 'mse':
calc_stats = lambda m1,m2,v1,v2: m1.sub(m2).pow(2).mean() + v1.sub(v2).pow(2).mean()
else:
if ref_stats_dict:
if pre_bn:
target_mean_key, target_var_key='running_mean', 'running_var'
else:
target_mean_key, target_var_key = 'bias', 'weight'
if mode == 'kl':
calc_stats = lambda m1, m2, v1, v2 : Gaussian_KL(m2,v2,m1,v1,epsilon).mean()
elif mode == 'sym':
calc_stats = lambda m1, m2, v1, v2: Gaussian_sym_KL(m1,v1,m2,v2,epsilon).mean()
else:
if mode == 'kl':
calc_stats = lambda m1, m2, v1, v2: reversedKLDNormal(m2,v2,epsilon).mean()
elif mode == 'sym':
calc_stats = lambda m1, m2, v1, v2: symKLDNormal(m2,v2,epsilon).mean()
for i,(k, (m, v)) in enumerate(batch_statistics.items()):
if ref_stats_dict and k in ref_stats_dict:
ref_dict = ref_stats_dict[k]
m_ref, v_ref = ref_dict[target_mean_key],ref_dict[target_var_key]
else:
m_ref, v_ref = th.zeros_like(m), th.ones_like(v)
moments_distance = calc_stats(m_ref, m, v_ref, v)
if verbose > 0:
with th.no_grad():
zero_sigma = (v < 1e-5).sum()
if zero_sigma > 0 or moments_distance > 5*loss_stat/i:
print(f'high divergence in layer {k}: {moments_distance}'
f'\nmu:{m.mean():0.4f}<s:r>{m_ref.mean():0.4f}\tsigma:{v.mean():0.4f}<s:r>{v_ref.mean():0.4f}\tsmall sigmas:{zero_sigma}/{len(v)}')
loss_stat = loss_stat + moments_distance
ret_val = loss_stat / len(batch_statistics)
return ret_val
def plot_grid(image_tensor,samples= 16,nrow=4,denorm_meta=None):
g=image_tensor[:samples].detach().cpu()
if denorm_meta:
mean = th.tensor(denorm_meta.mean, requires_grad=False).reshape((1, 3, 1, 1))
std = th.tensor(denorm_meta.std, requires_grad=False).reshape((1,3, 1, 1))
g = g*std+mean
g = tv.utils.make_grid(g, nrow=nrow)
plt.imshow(g.permute((1, 2, 0)))
#plt.waitforbuttonpress()
pass
# collects stats accross devices as well
def layer_stats_hook_dict(stat_dict,trance_name,device,pre_bn=True):
def stat_record(m,inputs , outputs):
target_activations = inputs if pre_bn else [outputs.clone()]
sum = target_activations[0].sum((0, 2, 3))
sum_p2 = target_activations[0].pow(2).sum((0, 2, 3))
n= target_activations[0][:, 0, :, :].numel()
_sum=sum.to(device)
_sum_p2=sum_p2.to(device)
if trance_name not in stat_dict:
stat_dict[trance_name]=(_sum,_sum_p2,n)
else:
sum_, sum_p2_, n_ = stat_dict[trance_name]
stat_dict[trance_name]=(_sum+sum_, _sum_p2+sum_p2_,n+n_)
return stat_record
def layer_stats_hook(stat_list,device):
def stat_record(m,inputs,outputs):
mean = inputs[0].mean((0,2,3)).to(device)
std = inputs[0].transpose(1,0).contiguous().view((inputs[0].size(1),-1)).std(1).to(device)
stat_list.append((mean,std))
return stat_record
def sqrt_newton_schulz_autograd(A, numIters=100, dtype=th.float64,debug=False,eps=1e-6):
## based on https://github.com/msubhransu/matrix-sqrt
def compute_error(A, sA):
normA = th.sqrt(th.sum(th.sum(A * A, dim=1), dim=1))
error = A - th.bmm(sA, sA)
error = th.sqrt((error * error).sum(dim=1).sum(dim=1)) / normA
return th.mean(error)
source_type=A.dtype
A+= th.eye(A.size(0),dtype=A.dtype,device=A.device) * eps
A=A.type(dtype)
if A.dim()==2:
A=A.unsqueeze(0)
batchSize = A.size(0)
dim = A.size(1)
normA = A.mul(A).sum((1, 2)).sqrt()
Y = A.div(normA.view(batchSize, 1, 1).expand_as(A))
I = th.eye(dim,dtype=dtype,device=A.device).unsqueeze(0).repeat(batchSize, 1, 1)
Z = th.eye(dim,dtype=dtype,device=A.device).unsqueeze(0).repeat(batchSize, 1, 1)
for i in range(numIters):
T_ = 0.5 * (3.0 * I - Z.bmm(Y))
Y = Y.bmm(T_)
Z = T_.bmm(Z)
sA = Y * th.sqrt(normA).view(batchSize, 1, 1).expand_as(A)
if debug:
error = compute_error(A, sA)
print('sqrtm error:', error,
'max error wrt numpy:',abs(sA.squeeze().cpu().numpy()-linalg.sqrtm(A.squeeze().cpu().numpy()).real).max())
return sA.type(source_type)
def np_calculate_frechet_distance(mu1, sigma1, mu2, sigma2, eps=1e-8):
"""Numpy implementation of the Frechet Distance.
The Frechet distance between two multivariate Gaussians X_1 ~ N(mu_1, C_1)
and X_2 ~ N(mu_2, C_2) is
d^2 = ||mu_1 - mu_2||^2 + Tr(C_1 + C_2 - 2*sqrt(C_1*C_2)).
Stable version by Dougal J. Sutherland.
Params:
-- mu1 : Numpy array containing the activations of a layer of the
inception net (like returned by the function 'get_predictions')
for generated samples.
-- mu2 : The sample mean over activations, precalculated on an
representative data set.
-- sigma1: The covariance matrix over activations for generated samples.
-- sigma2: The covariance matrix over activations, precalculated on an
representative data set.
Returns:
-- : The Frechet Distance.
"""
mu1 = np.atleast_1d(mu1)
mu2 = np.atleast_1d(mu2)
sigma1 = np.atleast_2d(sigma1)
sigma2 = np.atleast_2d(sigma2)
assert mu1.shape == mu2.shape, \
'Training and test mean vectors have different lengths'
assert sigma1.shape == sigma2.shape, \
'Training and test covariances have different dimensions'
diff = mu1 - mu2
# Product might be almost singular
covmean, _ = linalg.sqrtm(sigma1.dot(sigma2), disp=False)
if not np.isfinite(covmean).all():
msg = ('fid calculation produces singular product; '
'adding %s to diagonal of cov estimates') % eps
print(msg)
offset = np.eye(sigma1.shape[0]) * eps
covmean = linalg.sqrtm((sigma1).dot(sigma2)+ offset)
# Numerical error might give slight imaginary component
if np.iscomplexobj(covmean):
if not np.allclose(np.diagonal(covmean).imag, 0, atol=1e-3):
m = np.max(np.abs(covmean.imag))
raise ValueError('Imaginary component {}'.format(m))
covmean = covmean.real
tr_covmean = np.trace(covmean)
tr= np.trace(sigma1) + np.trace(sigma2) - 2 * tr_covmean
diff_l2=diff.dot(diff)
return (diff_l2 + tr),covmean,tr,diff_l2
def np_calculate_activation_statistics(features):
mu = np.mean(features.cpu().numpy(), axis=0)
sigma = np.cov(features.cpu().numpy(), rowvar=False)
return mu, sigma
class InceptionDistance(nn.Module):
#todo: FID results do not match previously reported scores, statistics as well as matrix factorization are almost
# identical to numpy reference
debug_diff = lambda x, y: abs(x - y.cpu().numpy()).mean()
def __init__(self,model=None,ref_inputs=None,accumulate=False,model_grad=False,input_size=None,ref_path=None):
super().__init__()
self.input_size= input_size or ref_inputs.shape[2:]
if model is None:
self.input_size=(299,299)
self._I3 =tv.models.inception_v3(True)
else:
assert hasattr(model,'fc')
self._I3 = model
assert self.input_size is not None and type(self.input_size[0]) == int
if not model_grad:
freeze_params(self._I3)
self.pool_output=None
self.fc_output=None
self.mu=None
self.sigma=None
self.accumulate=accumulate
if accumulate:
self.collected_activations=None
self.collected_ref_activations=None
def feature_recording_hook(m,inputs,outputs):
assert inputs[0] is not None
print('updated logged features with input size',inputs[0].size())
self.pool_output=inputs[0]
self.fc_output=outputs[0]
self._I3.fc.register_forward_hook(feature_recording_hook)
self.ref_path = ref_path + '.pth' if ref_path is not None else None
if ref_inputs is None:
self.ref_mu=None
self.ref_sigma=None
else:
self.to(ref_inputs.device)
print('calc initial ref stats')
with th.no_grad():
self._update_features(ref_inputs,ref=True)
self.update_ref_stats(self.pool_output,debug=True)
def measure_or_load_ref_stats(self,ref_loader=None,limit=None,ref_path=None):
ref_path= ref_path or self.ref_path
assert ref_path is not None
if not os.path.exists(ref_path + '.pth'):
print(f'reference statistics are not found at {ref_path}.pth\ncollecting reference data')
mode=self.accumulate
self.accumulate=True
with th.no_grad():
for i, (fid_ref_b, _) in enumerate(ref_loader):
if i * fid_ref_b.size(0) >= limit:
break
fid_ref_b = fid_ref_b.to(self.device)
self.forward(ref=fid_ref_b, report=False)
self.accumulate=mode
else:
self.ref_mu, self.ref_sigma = th.load(ref_path)
def update_ref_stats(self,ref_feat,debug=False):
print('overwriting referece staistics with new ones', ref_feat.shape[0])
self.ref_mu, self.ref_sigma = self._calc_stat(ref_feat)
if self.ref_path is not None and not os.path.exists(self.ref_path):
print(f'saving reference measurements as {self.ref_path} based on {ref_feat.shape[0]} samples')
th.save([self.ref_mu, self.ref_sigma],self.ref_path)
if debug:
self.np_ref_mu, self.np_ref_sigma = np_calculate_activation_statistics(ref_feat)
def to(self,*args,**kwargs):
print(args,kwargs)
self._I3.to(*args,**kwargs)
def train(self, mode=True):
self._I3.train(mode)
def _update_features(self,inputs,ref=False):
if ref:
with th.no_grad():
train= self._I3.training
self._I3.eval()
if inputs.shape[2:]!=self.input_size:
inputs=F.interpolate(inputs,self.input_size)
self._I3(inputs)
self._I3.train(train)
if self.accumulate:
if self.collected_ref_activations is None:
self.collected_ref_activations=self.pool_output.cpu()
else:
self.collected_ref_activations = th.cat([self.collected_ref_activations,self.pool_output.cpu()])
else:
if inputs.shape[2:] != self.input_size:
inputs = F.interpolate(inputs, self.input_size)
self._I3(inputs)
if self.accumulate:
if self.collected_activations is None:
self.collected_activations=self.pool_output.cpu(),self.fc_output.clone().cpu()
else:
self.collected_activations = (th.cat([self.collected_activations[0],self.pool_output.cpu()]),
th.cat([self.collected_activations[1],self.fc_output.clone().cpu()]))
def _calc_stat(self,feat_2d):
assert feat_2d.dim() == 2
#assume first dim is batch,second dim is features
print(f'compute stats for {feat_2d.size(1)} variables based on {feat_2d.size(0)} examples')
mu=feat_2d.mean(0,keepdim=True)
z_mean_feat=feat_2d-mu
sigma= z_mean_feat.transpose(1,0).matmul(z_mean_feat)/(z_mean_feat.size(0)-1)
return mu,sigma
def _calc_Frechet_from_stats(self,mu,sigma,ref_mu,ref_sigma):
A = sigma.matmul(ref_sigma)
B = sigma + ref_sigma - 2 * sqrt_newton_schulz_autograd(A, debug=False).squeeze()
tr = th.trace(B)
diff = (ref_mu - mu).squeeze()
diff_l2 = diff.dot(diff)
FId = th.clamp(diff_l2 + tr, 0)
return FId
def _calc_Frechet(self,features,ref_features=None,debug=False):
mu,sigma = self._calc_stat(features)
if ref_features is not None:
with th.no_grad():
ref_mu, ref_sigma = self._calc_stat(ref_features)
else:
if self.ref_mu is None and self.collected_ref_activations is not None:
self.update_ref_stats(self.collected_ref_activations,debug=debug)
else:
assert self.ref_mu is not None
assert self.ref_sigma is not None
# using precomputed values
ref_mu, ref_sigma=self.ref_mu,self.ref_sigma
ref_mu, ref_sigma = ref_mu.to(mu.device), ref_sigma.to(mu.device)
FID=self._calc_Frechet_from_stats(mu,sigma,ref_mu, ref_sigma)
return FID
def _calc_IS(self,features):
return None
def forward(self,inputs=None,ref=None,debug=False,report=True,reset_buffers_on_report=True):
# report - is used to report on the statistics of the collected activations,
# False: statistics are only on the most recent inputs.
# reset_buffers_on_report-only when module accomulation is on
# False:statistics of inputs and references are computed from scratch after every invocation, while both buffers keep tracking history.
# True: history will reset after reporting accumulated results. Reference statistics are kept as long as no new referece inputs are passed.
if ref is not None:
# calculate fresh reference features for current batch
self._update_features(ref,True)
ref=self.pool_output
if inputs is not None:
# update prediction if inputs are provided, none inputs are usefull when the same model model is used
# for inference and measuring IS/FID on same data.
self._update_features(inputs)
if not report:
return
if self.accumulate and report:
assert self.collected_activations is not None, 'first run some input data'
if self.collected_ref_activations is not None:
# update reference stats
self.update_ref_stats(self.collected_ref_activations,debug=debug)
else:
assert self.ref_mu is not None and self.ref_sigma is not None, 'please provided reference statistics'
FId, IS = self._calc_Frechet(self.collected_activations[0],debug=debug), \
self._calc_IS(self.collected_activations[1])
if reset_buffers_on_report:
# reset collected activations
self.collected_activations = None
self.collected_ref_activations = None
return FId, IS
else:
assert inputs is not None
# report FID on the input batch.
FId,IS=self._calc_Frechet(self.pool_output,ref,debug=debug), self._calc_IS(self.fc_output)
return FId,IS
def layer_c_norm_hook(norm_ranking_dict,device,collect):
#collect norms per channel accross batch and spatial dimensions, compute softmax to normalize according to channel activity
def c_norm_hook(m,inputs,outputs):
if collect:
norms = inputs[0].transpose(1,0).contiguous().view((inputs[0].size(1),-1)).norm(1).to(device)
norm_ranking_dict[m] = F.softmax(norms)
return c_norm_hook
def generate_channel_mask(norm_ranking_dict,batch_size=1):
suppression_mask_dict={}
for m,nr in norm_ranking_dict:
sampler=th.distributions.Bernoulli(nr)
suppression_mask_dict[m] = sampler.sample((batch_size,)).reshape(batch_size,-1,1,1)
return suppression_mask_dict
def layer_output_mask(suppression_mask_dict,device):
def c_mask_output_hook(m,inputs,outputs):
outputs[0]=outputs[0]*suppression_mask_dict[m]
return c_mask_output_hook
## from pytorch tutorial
def fgsm_attack(image, epsilon, data_grad):
# Collect the element-wise sign of the data gradient
data_grad = data_grad.sign()
#sign_data_grad = th.clamp(data_grad,0,1)
# Create the perturbed image by adjusting each pixel of the input image
perturbed_image = image + epsilon*data_grad
# Adding clipping to maintain [0,1] range
# Return the perturbed image
#perturbed_image = th.clamp(perturbed_image, 0, 1)
return perturbed_image
def forward(model, data_loader, inp_shape, args, device, batch_augment = None, normalize_inputs=None,
optimizer=None, smooth_loss=None, cont_loss=None, recorded_stats_dict={},reference_stats_dict={},
adversarial=False, mixer=None, FID=None, log=None, save_path='tmp', time_stamp=None, use_amp=False):
# for m in model.modules():
# assert m.training == False
# for n,p in model.named_parameters():
# assert p.requires_grad==False,n
CE = nn.CrossEntropyLoss().to(device)
CE_meter = AverageMeter()
output_meter = AverageMeter()
output_meter_hard = AverageMeter()
batch_time = AverageMeter()
generator_time = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
loss_avg = AverageMeter()
loss_cont_avg = AverageMeter()
loss_stat_avg = AverageMeter()
loss_smoothness = AverageMeter()
confusion = ConfusionMeter()
hot_one_map = np.eye(args.nclasses)
hot_one_map_th=th.tensor(hot_one_map,device=device)
n_iter = (args.n_samples_to_generate * args.nclasses) // args.batch_size
end = time.time()
if args.record > 0:
from utils.misc import Recorder
r = Recorder(model, recording_mode=['inputs'],recursive=True,
include_matcher_fn=lambda n,m: isinstance(m,th.nn.BatchNorm2d) or isinstance(m,th.nn.Linear))
for step, (inputs_, labels_) in enumerate(data_loader):
labels_ = labels_.to(device)
inputs_ = inputs_.to(device)
if optimizer is None and args.measure_ds_limit > 0 and step * inputs_.size(0) >= args.measure_ds_limit * args.nclasses:
break
##currently broken
# if FID:
# with th.no_grad():
# #aggregate target dataset activations
# FID(inputs,report=False)
# print('FID:',FId.item())
## for reporting topk results
topk = min(5, args.nclasses)
if adversarial:
##todo optional: choose arbitrary adv_targets (targeted attack)
#adv_targets = (labels_ + th.randint_like(labels_, 1, args.nclasses - 2)) % args.nclasses
adv_targets = None
inputs_.requires_grad = True
inputs_.retain_grad()
inp_ptr = inputs_
else:
adv_targets = None
if not time_stamp:
time_stamp = datetime.now().strftime('%Y-%m-%d_%H-%M-%S')
for n_replay in range(args.replay_latent):
if optimizer and n_replay % 20 == 0:
with th.no_grad():
#inputs_.div_(inputs_.norm()) #var((0,2,3),keepdim=True)
#inputs_.data.div_(inputs_.data.min()-inputs_.data.max())
# inputs_.data.mul_(255).add_(0.5).clamp_(0, 255).floor_().div_(255)
# inputs_.add_(-inputs_.min())
# inputs_.div_(inputs_.max())
inputs_.mul_(255).add_(0.5).clamp_(0, 255).floor_().div_(255)
#inputs_ = inputs_.sigmoid()
prior_loss = th.zeros(1).to(device)
loss_stat = th.zeros(1).to(device)
loss = th.zeros(1).to(device)
if smooth_loss:
loss_smooth = smooth_loss(inputs_).mean()
if optimizer and n_replay in args.smooth_scale_decay:
args.smooth_scale = args.smooth_scale * args.smooth_scale_decay[n_replay]
print(f'reducing blur by factor {args.smooth_scale_decay[n_replay]}')
loss_smoothness.update(loss_smooth.item())
prior_loss = prior_loss + loss_smooth * args.smooth_scale
if cont_loss:
loss_cont = cont_loss(inputs_).mean()
loss_cont_avg.update(loss_cont.item())
prior_loss = prior_loss + loss_cont
if normalize_inputs:
inputs = (inputs_ - normalize_inputs['mean']) / normalize_inputs['std']
else:
inputs = inputs_
if batch_augment:
inputs, labels = batch_augment(inputs, labels_, adv_targets)
if adv_targets:
labels, adv_targets = labels
else:
labels = labels_
if args.DEBUG_SHOW:
plot_grid(inputs)
if (inputs.shape[2:]) != inp_shape[1:]:
inputs = nn.functional.interpolate(inputs, mode='bilinear', size=inp_shape[1:])
if mixer:
inputs = mixer(inputs, [0.5, inputs.size(0), True])
generator_time.update(time.time()-end)
if args.record <= step:
r.master_record_enable = False
out = model(inputs)
if args.record > step:
r.insert('inputs',inputs)
r.insert('labels', labels)
r.insert('outputs', out)
if args.measure:
soft_out = F.softmax(out.detach() / args.output_temp).cpu().numpy()
output_meter.update(soft_out.mean(0))
output_meter_hard.update(hot_one_map[soft_out.argmax(1)].mean(0))