-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.lua
1319 lines (1207 loc) · 59.2 KB
/
train.lua
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
--------------------------------------------------------
-- Torch Implementation of Attending to Discriminative Certainty for Domain Adaptation(CVPR2019)
--- Written by jointly Vinod Kumar Kurmi ([email protected]) and Shanu Kumar
require 'cutorch'
require 'nngraph'
print("pass")
require 'math'
require 'cunn'
require 'optim'
require 'gnuplot'
require 'loadcaffe'
require 'image';
require 'torch';
require 'nn';
require 'xlua'
require 'loadcaffe'
require 'cudnn'
require 'nnlr' --- for layer wise learnig rate
require 'distributions';
bayesian_crossentropy = require 'bayesian_crossentropy';
require 'cephes'
----------------------------------------------------
local c = require 'trepl.colorize'
local data_tm = torch.Timer()
----------------------------------------------------
opt = {
manual_seed=1, -- Seed
batchSize = 20, -- batch Size
Test_batchSize = 40, -- Test time batch size, it may change after the last epoch of test data
start_Batch_IndexTest=1, --batch index at time of testing
loadSize = 256, -- resize the loaded image to loadsize maintaining aspect ratio. -- see donkey_folder.lua
fineSize = 224, -- size of random crops
nc = 3, -- # of channels in input
nThreads = 1, -- # of data loading threads to use
gpu = 1, -- gpu = 0 is CPU mode. gpu=X is GPU mode on GPU X
save='logs/', -- Saving the logs of trainining
net1_freeze='yes', -- For not updating the first 3 Conv layer
--momentum
number_of_testclass=31, -- number of class in test dataset, in general it is =source class but we can use less class also.
lamda=1, -- Lamda value for gradeint reversal value.(fix)
momentum=0.9,
baseLearningRate=0.0002,
max_epoch=6000,
gamma=0.001, -- for inverse policy : base_lr * (1 + gamma * iter) ^ (- power)
power=0.75, -- for inverse policy : base_lr * (1 + gamma * iter) ^ (- power)
max_epoch_grl=10000, -- For progress in process , calculate the lamda for grl
alpha=10, -- LR schdular (2nd way)
num_sim = 100, -- monte_carlo_simulations
alpha_0_cls = 100,
alpha_0_dis = 15,
al_wt_for_dis=100,
al_wt_for_clw=100,
}
cutorch.manualSeed(opt.manual_seed)
torch.manualSeed(opt.manual_seed)
--=====================Tuning Parameters===================================
local prev_accuracy = 0
local max_accuracy = 0
batchSize =opt.batchSize
opt.save=opt.save .. 'aleatoric_batchsize_' .. opt.batchSize
torch.setnumthreads(1)
torch.setdefaulttensortype('torch.FloatTensor')
--==============Ploting Fuction=============================================================================
confusion = optim.ConfusionMatrix({'letter_tray','paper_notebook','printer','bike_helmet','desk_lamp','mobile_phone',
'desk_chair','pen','phone','headphones','ring_binder','tape_dispenser','bookcase','back_pack','laptop_computer','stapler',
'ruler','mouse','projector','trash_can','monitor','file_cabinet','speaker','punchers','desktop_computer','bottle',
'mug','keyboard','scissors','bike','calculator'})
print('Will save at '..opt.save)
paths.mkdir(opt.save)
-- testLogger = optim.Logger(paths.concat(opt.save, 'test.log'))
-- testLogger:setNames{'% mean class accuracy (train set)', '% mean class accuracy (test set)'}
-- testLogger.showPlot = false
-------------logger for accuracy------------
acc_Logger = optim.Logger(paths.concat(opt.save, 'acc.log'))
acc_Logger:setNames{'Train acc', 'Test acc'}
acc_Logger.showPlot = false
----------------logger for loss----------------
error_Logger = optim.Logger(paths.concat(opt.save, 'error.log'))
error_Logger:setNames{'Train error', 'Test error', 'Train error cross', 'Test error cross'}
error_Logger.showPlot = false
---------------logger for entropy-----------------
entropy_Logger = optim.Logger(paths.concat(opt.save, 'entropy.log'))
entropy_Logger:setNames{'Train source entropy', 'Train target entropy', 'Test entropy'}
entropy_Logger.showPlot = false
---------------logger for differential entropy-----------------
entropy_diff_Logger = optim.Logger(paths.concat(opt.save, 'diff_entropy.log'))
entropy_diff_Logger:setNames{'Train source diff entropy', 'Train target diff entropy', 'Test diff entropy'}
entropy_diff_Logger.showPlot = false
---------------logger for var-----------------
var_Logger = optim.Logger(paths.concat(opt.save, 'var.log'))
var_Logger:setNames{'Train source var', 'Train target var', 'Test var'}
var_Logger.showPlot = false
----------------logger for brier score----------------
brier_Logger = optim.Logger(paths.concat(opt.save, 'brier.log'))
brier_Logger:setNames{'Train brier', 'Test brier'}
brier_Logger.showPlot = false
----------------logger for domain aleatoric loss----------------
error_domain_aleatoric_Logger = optim.Logger(paths.concat(opt.save, 'error_domain_aleatoric.log'))
error_domain_aleatoric_Logger:setNames{'Train source aleatoric error', 'Train target aleatoric error', 'Test target aleatoric error'}
error_domain_aleatoric_Logger.showPlot = false
----------------logger for domain aleatoric uncertainty----------------
aleatoric_domain_uncertainty_Logger = optim.Logger(paths.concat(opt.save, 'aleatoric_domain_uncertainty.log'))
aleatoric_domain_uncertainty_Logger:setNames{'Train source aleatoric domain uncertainty', 'Train target aleatoric domain uncertainty', 'Test target aleatoric domain uncertainty'}
aleatoric_domain_uncertainty_Logger.showPlot = false
----------------logger for classification aleatoric loss----------------
error_aleatoric_Logger = optim.Logger(paths.concat(opt.save, 'error_aleatoric.log'))
error_aleatoric_Logger:setNames{'Train source aleatoric error', 'Test target aleatoric error'}
error_aleatoric_Logger.showPlot = false
----------------logger for classification aleatoric uncertainty----------------
aleatoric_uncertainty_Logger = optim.Logger(paths.concat(opt.save, 'aleatoric_uncertainty.log'))
aleatoric_uncertainty_Logger:setNames{'Train source aleatoric uncertainty', 'Train target aleatoric uncertainty', 'Test target aleatoric uncertainty'}
aleatoric_uncertainty_Logger.showPlot = false
--==========================================================================================------------------- -------------------------Path Initilication ----------------
net_orignal=torch.load('../../../../../Pretrained_models/pretrained_netwrok_resnet/resnet-50.t7')
-- prototxt_name = '../pretrained_network/deploy.prototxt'
-- binary_name = '../pretrained_network/bvlc_alexnet.caffemodel'
-- net_orignal = loadcaffe.load(prototxt_name, binary_name,'cudnn');
-- print(' net_orignal', net_orignal)
---------------------------------------------------------
-- create Train data loader
local DataLoader = paths.dofile('data/data.lua')
local data = DataLoader.new(opt.nThreads, opt)
print("Train Dataset Size: ", data:size())
-- create Val data loader
local DataLoaderVal = paths.dofile('data/data_target.lua')
local dataVal = DataLoaderVal.new(opt.nThreads, opt)
print("Val Dataset Size: ", dataVal:size())
-- create Test data loader
local DataLoaderTest = paths.dofile('data_test/data.lua')
local dataTest = DataLoaderTest.new(0, opt)
print("test new Dataset Size: ", dataTest:size())
----------------------------------------------------
--===FUNCTIONS==============
function uti(filename)
local net = torch.load(filename)
net:apply(function(m) if m.weight then
m.gradWeight = m.weight:clone():zero();
m.gradBias = m.bias:clone():zero(); end end)
return net
end
function lgamma(input)
local temp = (cephes.lgam(input:float())):type(input:type())
return temp:resizeAs(input)
end
function digamma(input)
local temp = (cephes.digamma(input:float())):type(input:type())
return temp:resizeAs(input)
end
function check_accuracy(scores, targets)
local num_test = (#targets)[1]
local no_correct = 0
local confidences, indices = torch.sort(scores, true)
local predicted_classes = indices[{{},{1}}]:long()
targets = targets:long()
no_correct = no_correct + ((torch.squeeze(predicted_classes):eq(targets)):sum())
local accuracy = no_correct / num_test
return accuracy
end
function check_accuracyTest(scores, targets)
local num_test = (#targets)[1]
local no_correct = 0
local confidences, indices = torch.sort(scores, true)
local predicted_classes = indices[{{},{1}}]:long()
targets = targets:long()
if num_test==1 then
no_correct = no_correct + ((predicted_classes:eq(targets)):sum())
else
no_correct = no_correct + ((torch.squeeze(predicted_classes):eq(targets)):sum())
end
local accuracy = no_correct
return accuracy
end
function round(what, precision)
return math.floor(what*math.pow(10,precision)+0.5) / math.pow(10,precision)
end
function TableConcat(t1,t2)
for i=1,#t2 do
t1[#t1+1] = t2[i]
end
return t1
end
--===============================feature extractor======================================================
function build_feature_extractor()
net1= nn.MapTable()
net2= nn.MapTable()
net3= nn.MapTable()
netB= nn.MapTable()
net11= nn.Sequential()
net22= nn.Sequential()
net33= nn.Sequential()
net88= nn.Sequential()
netBB= nn.Sequential()
-- Layer by layer copy from the pretrained Alexnet Netwrok
for i, module in ipairs(net_orignal.modules) do
if(i<5) then
net11:add(module):learningRate('weight', 1)
:learningRate('bias', 2)
:weightDecay('weight', 1)
:weightDecay('bias', 0) --conv4
elseif (i==5) then
net11:add(module):learningRate('weight', 1)
:learningRate('bias', 2)
:weightDecay('weight', 1)
:weightDecay('bias', 0) --conv4
elseif (i==6) then
net11:add(module):learningRate('weight', 1)
:learningRate('bias', 2)
:weightDecay('weight', 1)
:weightDecay('bias', 0) --FC6
elseif (i==7) then
net22:add(module):learningRate('weight', 1)
:learningRate('bias', 2)
:weightDecay('weight', 1)
:weightDecay('bias', 0) --FC6
elseif (i==8) then
net22:add(module):learningRate('weight', 1)
:learningRate('bias', 2)
:weightDecay('weight', 1)
:weightDecay('bias', 0) --FC6
elseif (i>8 and i<11 ) then
net33:add(module):learningRate('weight', 1)
:learningRate('bias', 2)
:weightDecay('weight', 1)
:weightDecay('bias', 0) --FC6
else
net88:add(module):learningRate('weight', 1)
:learningRate('bias', 2)
:weightDecay('weight', 1)
:weightDecay('bias', 0) --FC6
end
end
net1:add(net11)
net2:add(net22)
net3:add(net33)
-- Bottlenec Network------
netBB:add(nn.Linear( 2048, 256)):learningRate('weight', 10)
:learningRate('bias', 20)
:weightDecay('weight', 1)
:weightDecay('bias', 0)
netBB:add(nn.ReLU(true))
netB:add(netBB)
return net1,net2,net3,netB
end
function build_classifier()
-- Classifier Network------------------
local net4 = nn.MapTable()
local net44 = nn.Sequential()
net44:add( nn.Linear(256, 128)):learningRate('weight', 10)
:learningRate('bias', 20)
net44:add(nn.ReLU(true))
------stochastic inference----------
net44:add(nn.Dropout(0.5, nil, nil, true))
-- net44:add(nn.LogSoftMax())
net4:add(net44)
return net4
end
function build_classifier_logit()
local net5 = nn.MapTable()
local net55 = nn.Sequential()
net55:add( nn.Linear(128, 31)):learningRate('weight', 10)
:learningRate('bias', 20)
:weightDecay('weight', 1)
:weightDecay('bias', 0) --FC6
net5:add(net55)
return net5
end
function build_classifier_variance()
local netv = nn.MapTable()
local netvv = nn.Sequential()
netvv:add(nn.Linear(128, 1)):learningRate('weight', 10)
:learningRate('bias', 20)
:weightDecay('weight', 1)
:weightDecay('bias', 0) --FC6
netvv:add(nn.SoftPlus())
netv:add(netvv)
return netv
end
function build_discriminator()
local netD = nn.MapTable()
local netDD = nn.Sequential()
module = nn.GradientReversal(lambda)
netDD:add(module)
netDD:add( nn.Linear( 256, 1024)):learningRate('weight', 10)
:learningRate('bias', 20)
netDD:add(nn.ReLU(true))
netDD:add(nn.Dropout(0.5, nil, nil, false))
netDD:add( nn.Linear( 1024, 1024)):learningRate('weight', 10)
:learningRate('bias', 20)
netDD:add(nn.ReLU(true))
netDD:add(nn.Dropout(0.5, nil, nil, false))
-- netDD:add( nn.Linear( 1024, 2)):learningRate('weight', 10)
-- :learningRate('bias', 20)
--netDD:add(nn.Sigmoid()) -- removed this layer if we are using the nn.CrossEntropyCriterion()
netD:add(netDD)
--Initially Lamda set =0
module:setLambda(0)
return netD
end
function build_discriminator_logit()
local netDL = nn.MapTable()
local netDDL = nn.Sequential()
netDDL:add(nn.Linear(1024, 2)):learningRate('weight', 10)
:learningRate('bias', 20)
netDL:add(netDDL)
return netDL
end
function build_discriminator_var()
local netDV = nn.MapTable()
local netDDV = nn.Sequential()
netDDV:add(nn.Linear(1024, 1)):learningRate('weight', 10)
:learningRate('bias', 20)
netDDV:add(nn.SoftPlus())
netDV:add(netDDV)
return netDV
end
-- ================= Models ====================
net4 = build_classifier()
net5 = build_classifier_logit() -- compute logits
netV = build_classifier_variance() -- compute variance
net1, net2, net3, netB = build_feature_extractor()
netD = build_discriminator()
netDL = build_discriminator_logit()
netDV = build_discriminator_var()
softmax = nn.MapTable()
softmax:add(nn.SoftMax())
logsoftmax = nn.MapTable()
logsoftmax:add(nn.LogSoftMax())
onehot = nn.MapTable()
onehot:add(nn.OneHot(31))
onehot_dis = nn.MapTable()
onehot_dis:add(nn.OneHot(2))--============ Criterion=================
local criterion = nn.CrossEntropyCriterion()
local criterionNLL = nn.ClassNLLCriterion()
local criterionNLL_parallel = nn.ParallelCriterion():add(criterionNLL):add(criterionNLL)
local criterionCrossE = nn.CrossEntropyCriterion()
local criterionCrossE_parallel = nn.ParallelCriterion():add(criterionCrossE):add(criterionCrossE)
-- local criterion_bayesian = autograd.nn.AutoCriterion('AutoMSE')(bayesian_categorical_crossentropy)
local criterion_bayesian_class = bayesian_crossentropy.criterion(opt.num_sim, 31)
local criterion_bayesian_dis = bayesian_crossentropy.criterion(opt.num_sim, 2)
--==========================================
-----------------------------------------------------------------------------------------------
if opt.gpu >=0 then
net1:cuda()
net2:cuda()
net3:cuda()
netB:cuda()
net4:cuda()
net5:cuda()
netV:cuda()
netD:cuda()
netDL:cuda()
netDV:cuda()
criterion:cuda()
criterionNLL_parallel:cuda()
criterionCrossE_parallel:cuda()
softmax:cuda()
onehot:cuda()
onehot_dis:cuda()
logsoftmax:cuda()
criterion_bayesian_dis:cuda()
criterion_bayesian_class:cuda()
end
--=== Different Learning rate for weigth and bias
local temp_baseWeightDecay=0.001 --no meaningin my case
local learningRates_Net1, weightDecays_Net1 = net1:getOptimConfig(opt.baseLearningRate,temp_baseWeightDecay)
local learningRates_Net2, weightDecays_Net2 = net2:getOptimConfig(opt.baseLearningRate, temp_baseWeightDecay)
local learningRates_Net3, weightDecays_Net3 = net3:getOptimConfig(opt.baseLearningRate, temp_baseWeightDecay)
local learningRates_NetB, weightDecays_NetB = netB:getOptimConfig(opt.baseLearningRate, temp_baseWeightDecay)
local learningRates_NetD, weightDecays_NetD = netD:getOptimConfig(opt.baseLearningRate, temp_baseWeightDecay)
local learningRates_Net4, weightDecays_Net4 = net4:getOptimConfig(opt.baseLearningRate, temp_baseWeightDecay)
local learningRates_Net5, weightDecays_Net5 = net5:getOptimConfig(opt.baseLearningRate, temp_baseWeightDecay)
local learningRates_NetDL, weightDecays_NetDL = netDL:getOptimConfig(opt.baseLearningRate, temp_baseWeightDecay)
local learningRates_NetDV, weightDecays_NetDV = netDV:getOptimConfig(opt.baseLearningRate, temp_baseWeightDecay)
local learningRates_NetV, weightDecays_NetV = netV:getOptimConfig(opt.baseLearningRate, temp_baseWeightDecay)
--===========Parameters===================================
parameters1, gradParameters1 = net1:getParameters()
parameters2, gradParameters2 = net2:getParameters()
parameters3, gradParameters3 = net3:getParameters()
parametersB, gradParametersB = netB:getParameters()
parametersD, gradParametersD = netD:getParameters()
parameters4, gradParameters4 = net4:getParameters()
parameters5, gradParameters5 = net5:getParameters()
parametersDL, gradParametersDL = netDL:getParameters()
parametersDV, gradParametersDV = netDV:getParameters()
parametersV, gradParametersV = netV:getParameters()
--============================================================
local method = 'xavier'
net4 = require('misc/weight-init')(net4, method)
net5 = require('misc/weight-init')(net5, method)
netV = require('misc/weight-init')(netV, method)
netB = require('misc/weight-init')(netB, method)
netD = require('misc/weight-init')(netD, method)
netDL = require('misc/weight-init')(netDL, method)
netDV = require('misc/weight-init')(netDV, method)
----------------------------------------------------------------------------------------------------
print('=> New Model')
print(model)
print('net1', net1)
print('net2', net2)
print('net3', net3)
print('netB', netB)
print("netD", netD)
print("net4", net4)
print("net5", net5)
print("netV", netV)
print("netDL", netDL)
print("netDV", netDV)
print(criterion)
collectgarbage()
local updated_learningrate=opt.baseLearningRate
--===================Training Fuctions======================================
function train()
net1:training()
net2:training()
net3:training()
net4:training()
net5:training()
netV:training()
netB:training()
netD:training()
netDL:training()
netDV:training()
epoch = epoch or 1
if(epoch>1) then
print(c.blue '==>'.." online epoch # " .. epoch .. ' [batchSize = ' .. batchSize .. ']')
local p=epoch/opt.max_epoch_grl
local baseWeightDecay = torch.pow((1 + epoch * opt.gamma), (-1 * opt.power)) -- need to chanage
updated_learningrate=opt.baseLearningRate*baseWeightDecay
print('Learnig Rate',updated_learningrate)
--opt.lamda=(2*torch.pow(1+torch.exp(-10*p),-1))-1
print('Lamda', opt.lamda)
module:setLambda(opt.lamda)
end
local avg_loss = 0
local avg_loss_cross = 0
local avg_aleatoric_loss_sr = 0 --- for source
local avg_aleatoric_domain_loss_sr = 0 --- for source
local avg_aleatoric_domain_loss_tg = 0 --- for target
local avg_aleatoric_domain_uncertainty_sr = 0 --- for source
local avg_aleatoric_domain_uncertainty_tg = 0 --- for target
local avg_aleatoric_uncertainty_sr = 0
local avg_aleatoric_uncertainty_tg = 0
local avg_acc = 0
local avg_entropy_sr = 0
local avg_entropy_tg = 0
local avg_brier = 0
local avg_diff_entropy_sr = 0
local avg_diff_entropy_tg = 0
local avg_dir_var_sr = 0
local avg_dir_var_tg = 0
local count = 0
for i = 1, data:size(), opt.batchSize do
-----Classifier Network-------------------
data_tm:reset(); data_tm:resume()
local batchInputs_source,label = data:getBatch()
local batchInputs_target = dataVal:getBatch()
local SlabelDomain = torch.Tensor(opt.batchSize):fill(1)
local TlabelDomain = torch.Tensor(opt.batchSize):fill(2)
local smooth_factor_cls = torch.Tensor(opt.batchSize, 31):fill((opt.alpha_0_cls - 31)/opt.alpha_0_cls)
local bias_cls = torch.Tensor(opt.batchSize, 31):fill(1/opt.alpha_0_cls)
-- local smooth_factor_dis = torch.Tensor(opt.batchSize, 2):fill((opt.alpha_0_dis - 31)/opt.alpha_0_dis)
-- local bias_dis = torch.Tensor(opt.batchSize, 2):fill(1/opt.alpha_0_dis)
if opt.gpu >=0 then
label=label:cuda()
SlabelDomain = SlabelDomain:cuda()
TlabelDomain = TlabelDomain:cuda()
batchInputs_source = batchInputs_source:cuda()
batchInputs_target = batchInputs_target:cuda()
smooth_factor_cls = smooth_factor_cls:cuda()
bias_cls = bias_cls:cuda()
-- smooth_factor_dis = smooth_factor_dis:cuda()
-- bias_dis = bias_dis:cuda()
end
-------- forwardNetwork -------------------------
outputs1 = net1:forward({batchInputs_source,batchInputs_target})
outputs2 = net2:forward(outputs1)
outputs3 = net3:forward(outputs2)
outputsB = netB:forward(outputs3)
-- outputs4 = net4:forward(outputsB)
-- outputs5 = net5:forward(outputs4) -- predicted class logits
-- outputsV = netV:forward(outputs4) -- predicted variance
outputsD = netD:forward(outputsB)
outputsDL = netDL:forward(outputsD) -- predicted domain logits
outputsDV = netDV:forward(outputsD) -- predicted domain variance
--------------------- backward Network ------------------------------------------
gradParametersD:zero()
gradParametersDL:zero()
gradParametersDV:zero()
gradParameters4:zero()
gradParameters5:zero()
gradParametersV:zero()
gradParametersB:zero()
gradParameters3:zero()
gradParameters2:zero()
gradParameters1:zero()
---Aleatoric local attenstion from discriminator using uncertainty (certainty)
local outputsB_source = outputsB[1]:clone()
local outputsB_target = outputsB[2]:clone()
gradParametersD:zero()
local outputsB_source1 = outputsB[1]:clone()
local outputsB_target1 = outputsB[2]:clone()
local dgradOutputs_modDV_aletoric = netDV:backward(outputsD, outputsDV)
local dgradOutputs_modD_aletoric = netD:backward({outputsB_source1, outputsB_target1}, dgradOutputs_modDV_aletoric)
outputsDV_clone1= outputsDV[1]:clone()
outputsDV_clone2= outputsDV[2]:clone()
local activations_source =outputsB[1]:clone()
local activations_target =outputsB[2]:clone()
activations_source:squeeze()
activations_target:squeeze()
local gradients_source = dgradOutputs_modD_aletoric[1]:squeeze()
local gradients_target = dgradOutputs_modD_aletoric[2]:squeeze()
local map_source = torch.cmul(activations_source, gradients_source)
map_source = map_source:cmul(torch.gt(map_source,0):typeAs(map_source))
map_source[map_source:le(0)] = -1000
local soft_source=nn.SoftMax():cuda():forward(map_source)
local one_tensor = torch.CudaTensor(outputsDV_clone1:size()):fill(1)
local weight_DV1 = torch.csub(one_tensor, outputsDV_clone1)
local soft_after_mul_source = soft_source:cmul(weight_DV1:view(opt.batchSize,1):expand(opt.batchSize,256))
local soft_clamp_source = torch.clamp(soft_after_mul_source, 0, 1)
attentive_weight_source=torch.CudaTensor() --Declaration of dgradOutputsS for source class
attentive_weight_source:resize(outputsB[1]:size())
attentive_weight_source:fill(1)
attentive_weight_source = attentive_weight_source + soft_clamp_source
effective_feature_source = torch.cmul(outputsB_source, attentive_weight_source)
local map_target = torch.cmul(activations_target, gradients_target)
map_target = map_target:cmul(torch.gt(map_target,0):typeAs(map_target))
map_target[map_target:le(0)] = -1000
local soft_target=nn.SoftMax():cuda():forward(map_target)
local one_tensor = torch.CudaTensor(outputsDV_clone2:size()):fill(1)
local weight_DV2 = torch.csub(one_tensor, outputsDV_clone2)
local soft_after_mul_target = soft_target:cmul(weight_DV2:view(opt.batchSize,1):expand(opt.batchSize,256))
local soft_clamp_target = torch.clamp(soft_after_mul_target, 0, 1)
attentive_weight_target=torch.CudaTensor() --Declaration of dgradOutputsS for source class
attentive_weight_target:resize(outputsB[2]:size())
attentive_weight_target:fill(1)
attentive_weight_target = attentive_weight_target + soft_clamp_target
effective_feature_target = torch.cmul(outputsB_target, attentive_weight_target)
--------------------------------------------------------------------------------------------------------------------
gradParametersD:zero()
gradParametersDL:zero()
gradParametersDV:zero()
gradParameters4:zero()
gradParameters5:zero()
gradParametersV:zero()
gradParametersB:zero()
gradParameters3:zero()
gradParameters2:zero()
gradParameters1:zero()
outputsD = netD:forward(outputsB)
outputsDL = netDL:forward(outputsD) -- predicted domain logits
outputsDV = netDV:forward(outputsD) -- predicted domain variance
-------- domain cross entropy ----------
errDomain = criterionCrossE_parallel:forward(outputsDL, {SlabelDomain,TlabelDomain})
--- domain classification loss grad -------------
dgradOutputsDomain_Classifier = criterionCrossE_parallel:backward(outputsDL, {SlabelDomain,TlabelDomain})
------- Gradient for aleatoric loss (domain classification + variance) -----------
local dgradOutputsDomain_Bayes = {}
------ aleatoric domain loss for source -------------------
err_aleatoric_dom_source = criterion_bayesian_dis:forward({outputsDL[1], SlabelDomain, outputsDV[1], torch.CudaTensor(opt.batchSize):fill(1), torch.CudaTensor(1):fill(opt.num_sim)})[1]
---- gradient for aleatoric loss (source domain batch) ------
dgradOutputsDomain_Bayes[1] = criterion_bayesian_dis:backward({outputsDL[1], SlabelDomain, outputsDV[1], torch.CudaTensor(opt.batchSize):fill(1), torch.CudaTensor(1):fill(opt.num_sim)}, torch.CudaTensor(1):fill(0.2))
------ aleatoric domain loss for target -------------------
err_aleatoric_dom_target = criterion_bayesian_dis:forward({outputsDL[2], TlabelDomain, outputsDV[2], torch.CudaTensor(opt.batchSize):fill(1), torch.CudaTensor(1):fill(opt.num_sim)})[1]
---- gradient for aleatoric loss (target domain batch) ------
dgradOutputsDomain_Bayes[2] = criterion_bayesian_dis:backward({outputsDL[2], TlabelDomain, outputsDV[2], torch.CudaTensor(opt.batchSize):fill(1), torch.CudaTensor(1):fill(opt.num_sim)}, torch.CudaTensor(1):fill(0.2))
---- gradient for domain classifier (gradient domain classification + gradient domain variance)
dgradOutputsDomain = {torch.add(dgradOutputsDomain_Classifier[1], dgradOutputsDomain_Bayes[1][1]), torch.add(dgradOutputsDomain_Classifier[2], dgradOutputsDomain_Bayes[2][1])}
---- Optimization Domain Confusion Branch (netDL) -------
feval_netDL = function(x)
dgradOutputs_modDL = netDL:backward(outputsD, dgradOutputsDomain)
return err, gradParametersDL
end
optim.sgd(feval_netDL, parametersDL, {
learningRates = learningRates_NetDL,
weightDecays = weightDecays_NetDL,
learningRate = updated_learningrate,
momentum = opt.momentum,
})
-----Optimzation NetDV --------
dgradOutputsDV = {dgradOutputsDomain_Bayes[1][3], dgradOutputsDomain_Bayes[2][3]}
feval_netDV = function(x)
dgradOutputs_modDV = netDV:backward(outputsD, dgradOutputsDV)
return err, gradParametersDV
end
optim.sgd(feval_netDV, parametersDV, {
learningRates = learningRates_NetDV,
weightDecays = weightDecays_NetDV,
learningRate = updated_learningrate,
momentum = opt.momentum,
})
---- Optimization NetD -------
local total_gradD = {}
total_gradD[1] = dgradOutputs_modDL[1] + dgradOutputs_modDV[1]
total_gradD[2] = dgradOutputs_modDL[2] + dgradOutputs_modDV[2]
feval_netD = function(x)
dgradOutputs_modD = netD:backward(outputsB, total_gradD)
return err, gradParametersD
end
optim.sgd(feval_netD, parametersD, {
learningRates = learningRates_NetD,
weightDecays = weightDecays_NetD,
learningRate = updated_learningrate,
momentum = opt.momentum,
})
outputs4 = net4:forward({effective_feature_source, effective_feature_target})
outputs5 = net5:forward(outputs4) -- predicted class logits
outputsV = netV:forward(outputs4) -- predicted variance
err_aleatoric_cl_sr = criterion_bayesian_class:forward({outputs5[1], label, outputsV[1], torch.CudaTensor(opt.batchSize):fill(1), torch.CudaTensor(1):fill(opt.num_sim)})[1]
-------------Gradient for classification -----------
local dgradOutputsS = torch.CudaTensor() --Declaration of dgradOutputsS for source class
dgradOutputsS:resize(outputs5[1]:size())
dgradOutputsS:zero()
------------ cross entropy for classification ------------------------------
--- smooth the true labels as its diffcult to match 1-hot continous distributions
local smooth_label_cls = torch.add(torch.cmul(onehot:forward({label})[1], smooth_factor_cls), bias_cls)
--- all the labels are possible as labels are smoothed
local all_label_cls = torch.CudaTensor(31)
for c = 1, 31 do
all_label_cls[c] = c
end
err = 0
for b = 1, opt.batchSize do
local criterion_weighted = nn.CrossEntropyCriterion(smooth_label_cls[b])
criterion_weighted:cuda()
local temp_logits = outputs5[1][b]:view(1, outputs5[1][b]:size()[1])
err = err + criterion_weighted:forward(torch.expand(temp_logits, temp_logits:size()[2], 31) , all_label_cls)
dgradOutputsS[b] = torch.sum(criterion_weighted:backward(torch.expand(temp_logits, temp_logits:size()[2], 31) , all_label_cls), 1)
criterion_weighted = nil
end
------- Gradient for aleatoric loss (classification + variance) -----------
dgradOutputsBV = criterion_bayesian_class:backward({outputs5[1], label, outputsV[1], torch.CudaTensor(opt.batchSize):fill(1), torch.CudaTensor(1):fill(opt.num_sim)}, torch.CudaTensor(1):fill(0.2))
-- Zero gradient for Target data Classification(we dont have target label)
local zeros = torch.CudaTensor()
zeros:resize(dgradOutputsS:size())
zeros:zero()
dgradOutputs = {torch.add(dgradOutputsS, dgradOutputsBV[1]), zeros}
-- dgradOutputs = {dgradOutputsS, zeros}
-- dgradOutputs = {dgradOutputsS, zeros}
---- Optimization Net5 -------
feval_net5 = function(x)
dgradOutputs_mod5 = net5:backward(outputs4, dgradOutputs)
return err, gradParameters5
end
optim.sgd(feval_net5, parameters5, {
learningRates = learningRates_Net5,
weightDecays = weightDecays_Net5,
learningRate = updated_learningrate,
momentum = opt.momentum,
})
-----Optimzation NetV --------
dgradOutputsV = dgradOutputsBV[3] -- gradient for variance of Source data
local zerov = torch.CudaTensor() -- Zero gradient for Target data Variance
zerov:resize(dgradOutputsV:size())
zerov:zero()
dgradOutputV = {dgradOutputsV, zerov}
feval_netV = function(x)
dgradOutputs_modV = netV:backward(outputs4, dgradOutputV)
return err, gradParametersV
end
optim.sgd(feval_netV, parametersV, {
learningRates = learningRates_NetV,
weightDecays = weightDecays_NetV,
learningRate = updated_learningrate,
momentum = opt.momentum,
})
---- Optimization Net4 -------
net4:zeroGradParameters()
gradParameters4:zero()
local total_grad4 = {}
total_grad4[1] = dgradOutputs_mod5[1] + dgradOutputs_modV[1]
total_grad4[2] = dgradOutputs_mod5[2] + dgradOutputs_modV[2]
feval_net4 = function(x)
dgradOutputs_mod4 = net4:backward({effective_feature_source, effective_feature_target}, total_grad4)
return err, gradParameters4
end
optim.sgd(feval_net4, parameters4, {
learningRates = learningRates_Net4,
weightDecays = weightDecays_Net4,
learningRate = updated_learningrate,
momentum = opt.momentum,
})
---- Optimization netB (bottleneck_Branch)-------
local total_grad={}
total_grad[1] = torch.cdiv(dgradOutputs_mod4[1], attentive_weight_source) + dgradOutputs_modD[1]
total_grad[2] = torch.cdiv(dgradOutputs_mod4[2], attentive_weight_target) + dgradOutputs_modD[2]
feval_netB = function(x)
dgradOutputs_modB = netB:backward(outputs3, total_grad)
return err, gradParametersB
end
optim.sgd(feval_netB, parametersB, {
learningRates = learningRates_NetB,
weightDecays = weightDecays_NetB,
learningRate = updated_learningrate,
momentum = opt.momentum,
})
---- Optimization net3(FC6,FC7) Branch -------
gradParameters3:zero()
feval_net3 = function(x)
dgradOutputs_mod3= net3:backward(outputs2,dgradOutputs_modB)
return err, gradParameters3
end
optim.sgd(feval_net3, parameters3, {
learningRates = learningRates_Net3,
weightDecays = weightDecays_Net3,
learningRate = updated_learningrate,
momentum = opt.momentum,
})
---- Optimization net2(Conv4 -Pool5) Branch -------
gradParameters2:zero()
feval_net2 = function(x)
dgradOutputs_mod2= net2:backward(outputs1,dgradOutputs_mod3)
return err, gradParameters2
end
optim.sgd(feval_net2, parameters2, {
learningRates = learningRates_Net2,
weightDecays = weightDecays_Net2,
learningRate = updated_learningrate,
momentum = opt.momentum,
})
---- if required then Net1 optimization----
if opt.net1_freeze =='no' then
gradParameters1:zero()
feval_net1 = function(x)
model.modules[1]:backward((batchInputs_source),dgradOutputs_mod2)
return gradOutputs_mod1, gradParameters1
end
optim.sgd(feval_net1, parameters1, {
learningRates = learningRates_Net1,
weightDecays = weightDecays_Net1,
learningRate = updated_learningrate,
momentum = opt.momentum,
})
end
local prob = softmax:forward({outputs5[1], outputs5[2]})
local prob_sr = prob[1]
local prob_tg = prob[2]
local _, max_ids_sr = torch.max(prob_sr,2)
local _, max_ids_tg = torch.max(prob_tg,2)
local log_prob = logsoftmax:forward({outputs5[1]:clone(), outputs5[2]:clone()})
local log_prob_sr = log_prob[1]
local log_prob_tg = log_prob[2]
local train_acc = check_accuracy(prob_sr, label)
-- average classification loss ---------
avg_loss = avg_loss + err/opt.batchSize
avg_loss_cross = avg_loss_cross + criterion:forward(outputs5[1], label)
-- average aleatoric source loss ---------
avg_aleatoric_loss_sr = avg_aleatoric_loss_sr + err_aleatoric_cl_sr
-- average aleatoric classification unceratinty for source -------
avg_aleatoric_uncertainty_sr = avg_aleatoric_uncertainty_sr + torch.mean(outputsV[1])
-- average aleatoric classification unceratinty for target -------
avg_aleatoric_uncertainty_tg = avg_aleatoric_uncertainty_tg + torch.mean(outputsV[2])
-- average aleatoric domain loss for source ---------
avg_aleatoric_domain_loss_sr = avg_aleatoric_domain_loss_sr + err_aleatoric_dom_source
-- average aleatoric domain loss for target ---------
avg_aleatoric_domain_loss_tg = avg_aleatoric_domain_loss_tg + err_aleatoric_dom_target
-- average aleatoric unceratinty for source-------
avg_aleatoric_domain_uncertainty_sr = avg_aleatoric_domain_uncertainty_sr + torch.mean(outputsDV[1])
-- average aleatoric unceratinty for target-------
avg_aleatoric_domain_uncertainty_tg = avg_aleatoric_domain_uncertainty_tg + torch.mean(outputsDV[2])
--- average training accuracy ----------
avg_acc = avg_acc + train_acc
--- average brier score ----------------
avg_brier = avg_brier + torch.squeeze(((torch.pow(prob_sr - onehot:forward({label})[1], 2):mean(2))/31):mean(1))
--- average entropy for source --------------------
avg_entropy_sr = avg_entropy_sr - torch.squeeze(((torch.cmul(prob_sr, log_prob_sr)):sum(2)):mean(1))
--- average entropy for target --------------------
avg_entropy_tg = avg_entropy_tg - torch.squeeze(((torch.cmul(prob_tg, log_prob_tg)):sum(2)):mean(1))
--- concentration for source domain
local cnt_sr = torch.mul(prob_sr, opt.alpha_0_cls)
--- concentration for target domain
local cnt_tg = torch.mul(prob_tg, opt.alpha_0_cls)
--- average differential entropy for source --------------------
avg_diff_entropy_sr = avg_diff_entropy_sr + torch.mean(torch.sum(lgamma(cnt_sr), 2) - lgamma(torch.sum(cnt_sr, 2)) - torch.cmul(31 - torch.sum(cnt_sr, 2), digamma(torch.sum(cnt_sr, 2))) - torch.sum(torch.cmul(torch.csub(cnt_sr, 1), digamma(cnt_sr)), 2))
--- average differential entropy for target --------------------
avg_diff_entropy_tg = avg_diff_entropy_tg + torch.mean(torch.sum(lgamma(cnt_tg), 2) - lgamma(torch.sum(cnt_tg, 2)) - torch.cmul(31 - torch.sum(cnt_tg, 2), digamma(torch.sum(cnt_tg, 2))) - torch.sum(torch.cmul(torch.csub(cnt_tg, 1), digamma(cnt_tg)), 2))
--- average dirichlet variance for source --------------------
local con0_sr = torch.expand(torch.sum(cnt_sr, 2):view(opt.batchSize, 1), opt.batchSize, 31)
local var_sr = torch.cdiv(torch.cmul(cnt_sr, torch.csub(con0_sr, cnt_sr)), torch.cmul(torch.pow(con0_sr, 2), torch.add(con0_sr, 1)))
avg_dir_var_sr = avg_dir_var_sr + torch.mean(var_sr:gather(2, max_ids_sr:view(max_ids_sr:size()[1], 1)))
--- average dirichlet variance for target --------------------
local con0_tg = torch.expand(torch.sum(cnt_tg, 2):view(opt.batchSize, 1), opt.batchSize, 31)
local var_tg = torch.cdiv(torch.cmul(cnt_tg, torch.csub(con0_tg, cnt_tg)), torch.cmul(torch.pow(con0_tg, 2), torch.add(con0_tg, 1)))
avg_dir_var_tg = avg_dir_var_tg + torch.mean(var_tg:gather(2, max_ids_tg:view(max_ids_tg:size()[1], 1)))
train_acc = nil
err = nil
smooth_label = nil
err_aleatoric_cl_sr = nil
err_aleatoric_cl_tg = nil
err_aleatoric_dom_source = nil
err_aleatoric_dom_target = nil
count = count + 1
end
epoch = epoch + 1
return avg_loss/count, avg_loss_cross/count, avg_acc/count, avg_brier/count, avg_entropy_sr/count, avg_entropy_tg/count, avg_diff_entropy_sr/count, avg_diff_entropy_tg/count, avg_dir_var_sr/count, avg_dir_var_tg/count, avg_aleatoric_loss_sr/count, avg_aleatoric_uncertainty_sr/count, avg_aleatoric_uncertainty_tg/count, avg_aleatoric_domain_loss_sr/count, avg_aleatoric_domain_loss_tg/count, avg_aleatoric_domain_uncertainty_sr/count, avg_aleatoric_domain_uncertainty_tg/count
end
--==============================Testing===================================================================
function test()
-- disable flips, dropouts and batch normalization
net1:evaluate()
net2:evaluate()
net3:evaluate()
netB:evaluate()
net4:evaluate()
net5:evaluate()
netV:evaluate()
netD:evaluate()
netDL:evaluate()
netDV:evaluate()
----------------scores---------------------------
local test_err = 0
local test_err_cross = 0
local test_acc = 0
local test_brier_score = 0
local test_entropy = 0
local test_var = 0
local test_diff_entropy = 0
local count = 0
local test_aleatoric_loss = 0
local test_aleatoric_uncertainty = 0
local test_aleatoric_domain_loss_tg = 0
local test_aleatoric_domain_uncertainty_tg = 0
opt.Test_batchSize = opt.batchSize
----------------testing----------------------------
for i = 1,dataTest:size(), opt.Test_batchSize do
data_tm:reset(); data_tm:resume()
opt.start_Batch_IndexTest=i
local batchInputs_test, validLabel = dataTest:getBatch(opt.start_Batch_IndexTest,dataTest:size())
local probs_sims = torch.CudaTensor(validLabel:size()[1], 31):zero()
local log_probs_sims = torch.CudaTensor(validLabel:size()[1], 31):zero()
local outputs_sims = torch.Tensor(validLabel:size()[1], 31):zero()
local outputsDL_sims = torch.Tensor(validLabel:size()[1], 2):zero()
local smooth_factor = torch.Tensor(validLabel:size()[1], 31):fill((opt.alpha_0_cls - 31)/opt.alpha_0_cls)
local bias = torch.Tensor(validLabel:size()[1], 31):fill(1/opt.alpha_0_cls)
local variance_sims = torch.Tensor(validLabel:size()[1], 1):zero()
local domain_variance_sims = torch.Tensor(validLabel:size()[1], 1):zero()
if opt.gpu >=0 then
probs_sims = probs_sims:cuda()
outputs_sims = outputs_sims:cuda()
batchInputs_test = batchInputs_test:cuda()
validLabel = validLabel:cuda()
domain_variance_sims = domain_variance_sims:cuda()
outputsDL_sims = outputsDL_sims:cuda()
variance_sims = variance_sims:cuda()
log_probs_sims = log_probs_sims:cuda()
smooth_factor = smooth_factor:cuda()
bias = bias:cuda()
end
for j = 1, opt.num_sim do
----------input forward-------------------------------
local outputs1 = net1:forward({batchInputs_test:cuda()})
local outputs2 = net2:forward(outputs1)
local outputs3 = net3:forward(outputs2)
local outputsB = netB:forward(outputs3)
local outputs4 = net4:forward(outputsB)
local outputs5 = net5:forward(outputs4) -- predicted class logits
local outputsV = netV:forward(outputs4) -- predicted variance
local outputsD = netD:forward(outputsB)
local outputsDL = netDL:forward(outputsD) -- predicted domain logits
local outputsDV = netDV:forward(outputsD) -- predicted domain variance
local probs = softmax:forward({outputs5[1]})[1]
local log_probs = logsoftmax:forward({outputs5[1]})[1]
probs_sims:add(probs)
outputs_sims:add(outputs5[1])
domain_variance_sims:add(outputsDV[1])
outputsDL_sims:add(outputsDL[1])
log_probs_sims:add(log_probs)
variance_sims:add(outputsV[1])
end
------------loss and accuracy per model---------------
local avg_probs = probs_sims/opt.num_sim
local avg_log_probs = log_probs_sims/opt.num_sim
local avg_logits = outputs_sims/opt.num_sim
local _, avg_max_ids = torch.max(avg_probs, 2)
---------- classification Loss -------------------------------
local smooth_label = torch.add(torch.cmul(onehot:forward({validLabel})[1], smooth_factor), bias)
--- all the labels are possible as labels are smoothed
local all_label = torch.CudaTensor(31)
for c = 1, 31 do
all_label[c] = c
end
err = 0
for b = 1, validLabel:size()[1] do
local criterion_weighted = nn.CrossEntropyCriterion(smooth_label[b])
criterion_weighted:cuda()
local temp_logits = avg_logits[b]:view(1, avg_logits[b]:size()[1])
err = err + criterion_weighted:forward(torch.expand(temp_logits, temp_logits:size()[2], 31) , all_label)
criterion_weighted = nil
end
test_err = test_err + err/(validLabel:size()[1])
test_err_cross = test_err_cross + criterion:forward(avg_logits, validLabel)
------------- classification accuracy -------------------
local test_batch_acc = check_accuracyTest(avg_probs, validLabel)
test_acc = test_acc + test_batch_acc
------------- brier score---------------------------------
test_brier_score = test_brier_score + torch.squeeze(((torch.pow(avg_probs - onehot:forward({validLabel})[1], 2):mean(2))/31):mean(1))
------------- entropy --------------------------------------
test_entropy = test_entropy - torch.squeeze(((torch.cmul(avg_probs, avg_log_probs)):sum(2)):mean(1))
--- concentration ------------
local cnt = torch.mul(avg_probs, opt.alpha_0_cls)
--- average differential entropy --------------------
test_diff_entropy = test_diff_entropy + torch.mean(torch.sum(lgamma(cnt), 2) - lgamma(torch.sum(cnt, 2)) - torch.cmul(31 - torch.sum(cnt, 2), digamma(torch.sum(cnt, 2))) - torch.sum(torch.cmul(torch.csub(cnt, 1), digamma(cnt)), 2))
--- average dirichlet variance for source --------------------
local con0 = torch.expand(torch.sum(cnt, 2):view(validLabel:size()[1], 1), validLabel:size()[1], 31)
local var = torch.cdiv(torch.cmul(cnt, torch.csub(con0, cnt)), torch.cmul(torch.pow(con0, 2), torch.add(con0, 1)))
test_var = test_var + torch.mean(var:gather(2, avg_max_ids:view(var:size()[1], 1)))