-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathmeta_train.py
1000 lines (781 loc) · 44.5 KB
/
meta_train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# from livelossplot import PlotLosses
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import os
os.environ['KMP_DUPLICATE_LIB_OK']='True'
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
import numpy as np
from collections import OrderedDict
from torchmeta.utils.gradient_based import gradient_update_parameters
from torch.utils.data import DataLoader
import random
from torchvision import datasets
import torchvision
import torchvision.transforms as transforms
import torchvision.utils as vutils
import torch.utils.data as data
from os import listdir
from PIL import Image
from models.renderer import *
from models.networks import VGGLoss,Net_Des19
import argparse
from torch.utils.tensorboard import SummaryWriter
from util.util import EPSILON
from util.descriptor import TDLoss_2
from skimage.transform import resize
from collections import OrderedDict
from meta_model import *
from meta_utils import *
def meta_learn(opt, model, coords, train_data, val_data, device, FeaExtractor=None):
print(opt.resume_name)
load_model='pretrain' if opt.load_pretrain else opt.model
max_iters = opt.train_iter#500000 #@param {type:"integer"}
outer_lr = opt.outer_lr #@param
inner_lr = opt.inner_lr #@param
inner_steps = opt.inner_step #@param {type:"integer"}
val_steps = opt.val_step #@param {type:"integer"}
first_order = opt.first_order #@param {type:"boolean"}
if opt.no_inner:
val_step=1
inner_steps=0
print('######################### inner lr is %.8f #############################'%(inner_lr))
print('######################### outer lr is %.8f, decay_lr %d #############################'%(outer_lr, opt.decay_lr))
print('######################### use first order %s #############################'%(first_order))
model.train()
if opt.losstype=='L2':
print('######################### using L2 loss #######################')
criterion = torch.nn.MSELoss()
else:
print('######################### using L1 loss #######################')
criterion = torch.nn.L1Loss()
params = OrderedDict(model.meta_named_parameters())
if opt.use_TD_loss:
criterionTD = TDLoss_2(device, 0)
criterionVGG = VGGLoss(opt)
weights = [0.1, 0.1, 0.1, 0.1]
keys = [1, 2, 3,4]
if opt.netloss=='Des19Net':
criterionNet = MyLoss(opt, keys, weights, device)
if opt.optim=='adam':
if opt.netloss=='Des19Net' and not opt.no_learn_loss:
print('------------------ learn loss %.8f-----------------------' % opt.lossnet_lr)
if FeaExtractor is None:
print('------------------ FeaExtractor in inner loop or no FeaExtractor -----------------------')
opt_outer = torch.optim.Adam([
{'params':model.parameters(),'lr':outer_lr, 'betas':[opt.beta1,opt.beta2], 'eps':opt.eps_adam},
{'params':criterionNet.parameters(),'lr':opt.lossnet_lr},
])
else:
print('------------------ FeaExtractor in outer loop %.8f-----------------------'% opt.lossnet_lr)
opt_outer = torch.optim.Adam([
{'params':model.parameters(),'lr':outer_lr, 'betas':[opt.beta1,opt.beta2], 'eps':opt.eps_adam},
{'params':list(criterionNet.parameters())+list(FeaExtractor.parameters()),'lr':opt.lossnet_lr}
])
else:
print('------------------ no learn loss -----------------------')
if FeaExtractor is None:
opt_outer = torch.optim.Adam(model.parameters(), lr=outer_lr, betas=[opt.beta1,opt.beta2], eps=opt.eps_adam)
else:
print('------------------ FeaExtractor in outer loop %.8f-----------------------'% opt.lossnet_lr)
opt_outer = torch.optim.Adam([
{'params':model.parameters(),'lr':outer_lr, 'betas':[opt.beta1,opt.beta2], 'eps':opt.eps_adam},
{'params':FeaExtractor.parameters(),'lr':opt.lossnet_lr}
])
# print('# ------------------- using adam ----------------------- #')
elif opt.optim=='rprop':
opt_outer = torch.optim.Rprop(model.parameters(), lr=outer_lr, etas=[opt.eta1,opt.eta2], step_sizes=[opt.step_min,opt.step_max])
print('# ------------------- using rprop ----------------------- #')
save_train_img_path = '%s/%s/train_imgs'%(opt.model, opt.name)
save_val_img_path = '%s/%s/val_imgs'%(opt.model, opt.name)
save_model_path = '%s/%s/models'%(opt.model, opt.name)
save_loss_path = '%s/%s/loss'%(opt.model, opt.name)
if not os.path.exists(save_train_img_path):
os.makedirs(save_train_img_path)
if not os.path.exists(save_val_img_path):
os.makedirs(save_val_img_path)
if not os.path.exists(save_model_path):
os.makedirs(save_model_path)
if not os.path.exists(save_loss_path):
os.makedirs(save_loss_path)
writer = SummaryWriter(save_loss_path)
if opt.resume_name!='':
print('resuming from ', opt.resume_name)
if opt.model=='OutConSiren' and opt.load_pretrain:
load_network_pretrain(model, FeaExtractor, opt.load_iter,'%s/%s/models'%(load_model, opt.resume_name), cond_type=opt.cond_type, name='model')
else:
load_network(model,opt.load_iter,'%s/%s/models'%(load_model, opt.resume_name), name='model')
# load_network(opt_outer,opt.load_iter,'%s/%s/models'%(opt.model, opt.resume_name), name='optimizer')
if FeaExtractor is not None and not opt.load_pretrain:
load_network(FeaExtractor, opt.load_iter,'%s/%s/models'%(load_model, opt.resume_name), name='FeaExtractor')
if opt.netloss=='Des19Net' and not opt.load_pretrain:
load_network(criterionNet.lossnet, opt.load_iter,'%s/%s/models'%(load_model, opt.resume_name), name='des19net')
log_step = []
val_log_ren = OrderedDict()
val_log_fea = OrderedDict()
trainfea_log = []
trainren_log = []
weight_list = []
step = 1
Position_map=PositionMap(opt.res,opt.res,3).to(device)
lightinten =16
in_lirange=0.01
if opt.model=='Siren':
myinput = coords[0]
# print('input shape', myinput.shape)
LightPos = torch.tensor([[0,0,4.0]], device=device)
while step < max_iters:
for i,example in enumerate(train_data):
if step > max_iters:
break
if opt.vary_light:
LightPos = NLight_VaryH(opt.N_input, in_lirange, device) # 0.03 is good to make light around center
# hard-code light intensity
lightinten = LightPos[0][-1].item()**2 if not opt.fixed_inten else 16
if opt.netloss=='Des19Net':
criterionNet.set_gradient(True)
tr_rens = OrderedDict()
tr_feas = OrderedDict()
model.zero_grad()
image = example.to(device)
VG_ren_loss = torch.tensor(0., device=device)
L_ren_loss = torch.tensor(0., device=device)
TD_ren_loss = torch.tensor(0., device=device)
outer_fea_loss = torch.tensor(0., device=device)
outer_feavg_loss = torch.tensor(0., device=device)
for task_id in range(opt.batch_size):
image_tr = image[task_id].permute(1,2,0) #[H,W,C]
gt_ren_tr = SingleRender_NumberPointLight_FixedCamera(2*image_tr[:,:,0:3]-1, image_tr[:,:,3:6],image_tr[:,:,6:7], image_tr[:,:,9:12], LightPos, Position_map, device, CamLi_co=True, lightinten=lightinten, no_spec=opt.no_spec) #[B,N, W,H,C]
gt_ren_tr = (gt_ren_tr.clamp(0,1)+opt.EPSILON)**(1/2.2)
# if opt.N_input>1:
# gt_ren_tr = gt_ren_tr.clone().reshape(1, gt_ren_tr.shape[1], gt_ren_tr.shape[2], opt.N_input*3) #[1, H, W, 3*N]
# print('gt_ren_tr ', gt_ren_tr.shape)
if opt.model=='UNet':
myinput = 2*gt_ren_tr.permute(0,3,1,2)-1 # [-1,1] #[N,3,H,W]
elif opt.model=='InConSiren':
myinput = 2*gt_ren_tr.permute(0,3,1,2)-1 #[N,3,H,W] [-1,1]
myinput = [coords[0],myinput] # coords:[H,W,2] myinput: [N,3,H,W]
elif opt.model=='OutConSiren':
myfea_in = FeaExtractor(2*gt_ren_tr.permute(0,3,1,2)-1).squeeze(0).permute(1,2,0) #[N,3,H,W] --> [H,W,32]
myinput = torch.cat([coords[0], myfea_in],dim=-1)
params_cp = params
## inner loop
for in_step in range(inner_steps):
g_tr = model(myinput, params=params_cp) # [0,1]
if g_tr.dim()==4:
g_tr = g_tr.squeeze(0).permute(1,2,0) # [B,C,H,W] --> [H,W,C]
if 'N' in opt.fea and 'all' not in opt.fea:
g_tr = ProcessNormal(opt, g_tr)
else:
g_tr = paramize_out(opt,g_tr)
if in_step==0 or in_step==1 or in_step==2:
tr_feas[str(in_step)] = g_tr
out_ren_tr = SingleRender_NumberPointLight_FixedCamera(2*g_tr[:,:,0:3]-1, g_tr[:,:,3:6], g_tr[:,:,6:7], g_tr[:,:,9:12], LightPos, Position_map, device, CamLi_co=True, lightinten=lightinten, no_spec=opt.no_spec) #[B,N, W,H,C]
out_ren_tr = (out_ren_tr.clamp(0,1)+opt.EPSILON)**(1/2.2)
ren_loss = criterion(gt_ren_tr, out_ren_tr)*opt.Winner_ren
if opt.netloss=='Des19Net':
inner_loss, des19_svbrdf_tr = criterionNet(g_tr, 2*gt_ren_tr.permute(0,3,1,2)-1)
inner_loss += ren_loss
elif opt.netloss=='GT':
inner_loss = criterion(g_tr, image_tr)
inner_loss += ren_loss
else:
inner_loss = ren_loss
# if opt.Wvgg_inner>0:
# out_ren_tr_vg = VGGpreprocess(out_ren_tr.permute(0,3,1,2))
# gt_ren_tr_vg = VGGpreprocess(gt_ren_tr.permute(0,3,1,2))
# inner_ren_vgg_loss = criterionVGG(out_ren_tr_vg, gt_ren_tr_vg)*opt.Wvgg_inner
# inner_loss += inner_ren_vgg_loss
if opt.Wtd_inner>0:
out_ren_tr_td = out_ren_tr.permute(0,3,1,2)
gt_ren_tr_td = gt_ren_tr.permute(0,3,1,2)
inner_ren_td = criterionTD(out_ren_tr_td, gt_ren_tr_td)*opt.Wtd_inner
inner_loss += inner_ren_td
# downL2_loss=0
# if opt.WdownL2_inner>0:
# gt16 = F.interpolate(gt_ren_tr, size=(16,16), mode='bilinear', align_corners=True)
# out16 = F.interpolate(out_ren_tr, size=(16,16), mode='bilinear', align_corners=True)
# downL2_loss = criterion(gt16, out16)*opt.WdownL2_inner
# inner_loss += downL2_loss
if in_step==0 or in_step==1 or in_step==2:
tr_rens[str(in_step)] = out_ren_tr
model.zero_grad()
params_cp = gradient_update_parameters(model, inner_loss, params=params_cp,
step_size=inner_lr, first_order=first_order)
if in_step==0 and task_id==0:
params_cp1=params_cp
if in_step==1 and task_id==0:
params_cp2=params_cp
## outer loop
pred_tr = model(myinput, params=params_cp)
if pred_tr.dim()==4:
pred_tr = pred_tr.squeeze(0).permute(1,2,0)
if 'N' in opt.fea and 'all' not in opt.fea:
pred_tr = ProcessNormal(opt, pred_tr)
else:
pred_tr = paramize_out(opt,pred_tr)
tr_feas[str(inner_steps)] = pred_tr
pred_ren_tr = SingleRender_NumberPointLight_FixedCamera(2*pred_tr[:,:,0:3]-1, pred_tr[:,:,3:6], pred_tr[:,:,6:7], pred_tr[:,:,9:12], LightPos, Position_map, device, CamLi_co=True, lightinten=lightinten, no_spec=opt.no_spec) #[B,N, W,H,C]
pred_ren_tr = (pred_ren_tr.clamp(0,1)+opt.EPSILON)**(1/2.2)
tr_rens[str(inner_steps)] = pred_ren_tr
if not opt.no_fea_outer:
outer_fea_loss += 0.25*(criterion(pred_tr[:,:,0:3], image_tr[:,:,0:3])*opt.wN_outer
+ criterion(pred_tr[:,:,3:6], image_tr[:,:,3:6])*opt.wD_outer
+ criterion(pred_tr[:,:,6:9], image_tr[:,:,6:9])*opt.wR_outer
+ criterion(pred_tr[:,:,9:12], image_tr[:,:,9:12])*opt.wS_outer)
# add VGG loss to features
if opt.Wfea_vgg>0:
D_vg_pred = VGGpreprocess(pred_tr[:,:,3:6].unsqueeze(0).permute(0,3,1,2))
D_vg_gt = VGGpreprocess(image_tr[:,:,3:6].unsqueeze(0).permute(0,3,1,2))
D_vg = criterionVGG(D_vg_pred, D_vg_gt)*opt.Wfea_vgg
S_vg_pred = VGGpreprocess(pred_tr[:,:,9:12].unsqueeze(0).permute(0,3,1,2))
S_vg_gt = VGGpreprocess(image_tr[:,:,9:12].unsqueeze(0).permute(0,3,1,2))
S_vg = criterionVGG(S_vg_pred, S_vg_gt)*opt.Wfea_vgg
outer_feavg_loss += (D_vg + S_vg)*0.5
if opt.Nren_outerloss>0:
if opt.no_fea_outer:
rand_light_pos = LightPos
else:
rand_light_pos = NLight_VaryH(opt.Nren_outerloss, 0.05, device)
# lightinten_rand = rand_light_pos[0][-1].item()**2 if opt.rand_inlight=='jitter' else 16
lightinten_rand = rand_light_pos[0][-1].item()**2
gt_ren_sampled = SingleRender_NumberPointLight_FixedCamera(2*image_tr[:,:,0:3]-1, image_tr[:,:,3:6],image_tr[:,:,6:7], image_tr[:,:,9:12], rand_light_pos, Position_map, device, CamLi_co=True, lightinten=lightinten_rand, no_spec=opt.no_spec) #[B,N, W,H,C]
gt_ren_sampled = (gt_ren_sampled.clamp(0,1)+opt.EPSILON)**(1/2.2)
pred_ren_sampled = SingleRender_NumberPointLight_FixedCamera(2*pred_tr[:,:,0:3]-1, pred_tr[:,:,3:6], pred_tr[:,:,6:7], pred_tr[:,:,9:12], rand_light_pos, Position_map, device, CamLi_co=True, lightinten=lightinten_rand, no_spec=opt.no_spec) #[B,N, W,H,C]
pred_ren_sampled = (pred_ren_sampled.clamp(0,1)+opt.EPSILON)**(1/2.2)
if opt.use_TD_loss:
out_ren = pred_ren_sampled[0:1,...].permute(0,3,1,2)
gt_ren = gt_ren_sampled[0:1,...].permute(0,3,1,2)
out_16 = nn.functional.interpolate(out_ren, size = opt.dres, mode='bilinear', align_corners=True)
gt_16 = nn.functional.interpolate(gt_ren, size = opt.dres, mode='bilinear', align_corners=True)
dren_loss = criterion(out_16, gt_16)*opt.Wdren_outer
TDren_loss = criterionTD(out_ren, gt_ren)*opt.WTDren_outer
L_ren_loss+=dren_loss
TD_ren_loss+=TDren_loss
else:
L_ren_loss += criterion(pred_ren_sampled, gt_ren_sampled)
if opt.Wvgg_outer>0:
out_ren_te_vg = VGGpreprocess(pred_ren_sampled[0:1,...].permute(0,3,1,2))
gt_ren_te_vg = VGGpreprocess(gt_ren_sampled[0:1,...].permute(0,3,1,2))
VG_ren_loss += criterionVGG(out_ren_te_vg, gt_ren_te_vg)*opt.Wvgg_outer
outer_fea_loss.div_(opt.batch_size)
outer_feavg_loss.div_(opt.batch_size)
# outer_ren_loss.div_(opt.batch_size)
L_ren_loss.div_(opt.batch_size)
TD_ren_loss.div_(opt.batch_size)
VG_ren_loss.div_(opt.batch_size)
outer_ren_loss = L_ren_loss + VG_ren_loss + TD_ren_loss
outer_loss = outer_fea_loss + outer_feavg_loss + outer_ren_loss
opt_outer.zero_grad()
outer_loss.backward()
opt_outer.step()
if step % opt.save_freq == 0 or step==1:
if opt.netloss=='Des19Net':
criterionNet.set_gradient(False)
if not opt.no_val:
val_rens = OrderedDict()
val_feas = OrderedDict()
val_rens_loss = OrderedDict()
val_feas_loss = OrderedDict()
val_N_loss = OrderedDict()
val_D_loss = OrderedDict()
val_R_loss = OrderedDict()
val_S_loss = OrderedDict()
model.eval()
g_test = None
out_val = None
valfea_loss = 0
valren_loss = 0
for i,val_example in enumerate(val_data):
val_example, _ = val_example
val_img = val_example.to(device)
image_te = val_img[0].permute(1,2,0) # [256, 256, C]
gt_ren_te = SingleRender_NumberPointLight_FixedCamera(2*image_te[:,:,0:3]-1, image_te[:,:,3:6],image_te[:,:,6:7], image_te[:,:,9:12], LightPos, Position_map, device, CamLi_co=True, lightinten=lightinten, no_spec=opt.no_spec) #[B,N, W,H,C]
gt_ren_te = (gt_ren_te.clamp(0,1)+EPSILON)**(1/2.2)
if opt.model=='UNet':
myinput = 2*gt_ren_te.permute(0,3,1,2)-1 # #[B,C,H,W] [-1,1]
elif opt.model=='InConSiren':
myinput = 2*gt_ren_te.permute(0,3,1,2)-1 #[B,3,H,W] [-1,1]
myinput = [coords[0],myinput] # coords:[H,W,2] myinput: [B,3,H,W]
elif opt.model=='OutConSiren':
myfea_in = FeaExtractor(2*gt_ren_te.permute(0,3,1,2)-1).squeeze(0).permute(1,2,0) #[B,3,H,W] --> [H,W,32]
myinput = torch.cat([coords[0], myfea_in],dim=-1)
params_test = params
for in_step in range(val_steps+1):
g_te = model(myinput, params=params_test)
if g_te.dim()==4:
g_te = g_te.squeeze(0).permute(1,2,0)
if opt.fea=='all_N1' or opt.fea=='D+R' or opt.fea=='all_N2' or opt.fea=='all_N3':
g_te = paramize_out(opt,g_te)
elif opt.fea=='N' or opt.fea=='N2' or opt.fea=='N3':
g_te = ProcessNormal(opt, g_te)
elif opt.fea=='D':
g_te = g_te
elif opt.fea=='R':
g_te = g_te.repeat(1,1,3)
if in_step==0:
val_feas[str(in_step)] = g_te
elif in_step==1:
val_feas[str(in_step)] = g_te
elif in_step==2:
val_feas[str(in_step)] = g_te
if not opt.no_inner:
out_ren_te = SingleRender_NumberPointLight_FixedCamera(2*g_te[:,:,0:3]-1, g_te[:,:,3:6], g_te[:,:,6:7], g_te[:,:,9:12], LightPos, Position_map, device, CamLi_co=True, lightinten=lightinten, no_spec=opt.no_spec) #[B,N, W,H,C]
out_ren_te = (out_ren_te.clamp(0,1)+EPSILON)**(1/2.2)
ren_loss = criterion(gt_ren_te, out_ren_te)*opt.Winner_ren
if opt.netloss=='Des19Net':
# inner_loss, des19_svbrdf = criterionNet(2*out_ren_tr.permute(0,3,1,2)-1, 2*gt_ren_tr.permute(0,3,1,2)-1)
inner_loss, des19_svbrdf_te = criterionNet(g_te, 2*gt_ren_te.permute(0,3,1,2)-1)
inner_loss += ren_loss
elif opt.netloss=='GT':
inner_loss = criterion(g_te, image_te)
inner_loss += ren_loss
else:
inner_loss = ren_loss
if opt.Wtd_inner>0:
out_ren_te_td = out_ren_te.permute(0,3,1,2)
gt_ren_te_td = gt_ren_te.permute(0,3,1,2)
inner_loss += criterionTD(out_ren_te_td, gt_ren_te_td)*opt.Wtd_inner
if in_step==0:
val_rens[str(in_step)] = out_ren_te
elif in_step==1:
val_rens[str(in_step)] = out_ren_te
elif in_step==2:
val_rens[str(in_step)] = out_ren_te
model.zero_grad()
params_test = gradient_update_parameters(model, inner_loss, params=params_test,
step_size=inner_lr, first_order=True)
# compute losses
for ((k1,fea), (k2,ren)) in zip(val_feas.items(), val_rens.items()):
val_feas_loss[k1] = val_feas_loss[k1] + criterion(fea, image_te) if k1 in val_feas_loss else criterion(fea, image_te)
val_rens_loss[k1] = val_rens_loss[k1] + criterion(ren, gt_ren_te) if k1 in val_rens_loss else criterion(ren, gt_ren_te)
val_N_loss[k1] = val_N_loss[k1] + criterion(fea[:,:,0:3], image_te[:,:,0:3]) if k1 in val_N_loss else criterion(fea[:,:,0:3], image_te[:,:,0:3])
val_D_loss[k1] = val_D_loss[k1] + criterion(fea[:,:,3:6], image_te[:,:,3:6]) if k1 in val_D_loss else criterion(fea[:,:,3:6], image_te[:,:,3:6])
val_R_loss[k1] = val_R_loss[k1] + criterion(fea[:,:,6:9], image_te[:,:,6:9]) if k1 in val_R_loss else criterion(fea[:,:,6:9], image_te[:,:,6:9])
val_S_loss[k1] = val_S_loss[k1] + criterion(fea[:,:,9:12], image_te[:,:,9:12]) if k1 in val_S_loss else criterion(fea[:,:,9:12], image_te[:,:,9:12])
################# save imgs ################33
final_vis_tr = OrderedDict()
final_vis_te = OrderedDict()
if 'all' in opt.fea:
for key_fea, fea in val_feas.items():
final_vis_te[key_fea] = torch.cat((fea[:,:,0:3],(fea[:,:,3:6]+EPSILON)**(1/2.2),fea[:,:,6:9],(fea[:,:,9:12]+EPSILON)**(1/2.2)),dim=1)
final_vis_te['gt'] = torch.cat((image_te[:,:,0:3],image_te[:,:,3:6]**(1/2.2),image_te[:,:,6:9],image_te[:,:,9:12]**(1/2.2)),dim=1)
for key_fea, fea in tr_feas.items():
final_vis_tr[key_fea] = torch.cat((fea[:,:,0:3],(fea[:,:,3:6]+EPSILON)**(1/2.2),fea[:,:,6:9],(fea[:,:,9:12]+EPSILON)**(1/2.2)),dim=1)
final_vis_tr['gt'] = torch.cat((image_tr[:,:,0:3],image_tr[:,:,3:6]**(1/2.2),image_tr[:,:,6:9],image_tr[:,:,9:12]**(1/2.2)),dim=1)
if opt.netloss=='Des19Net':
final_vis_tr['des19'] = torch.cat((des19_svbrdf_tr[:,:,0:3],des19_svbrdf_tr[:,:,3:6]**(1/2.2),des19_svbrdf_tr[:,:,6:9],des19_svbrdf_tr[:,:,9:12]**(1/2.2),des19_svbrdf_tr[:,:,3:6]**(1/2.2)),dim=1)
final_vis_te['des19'] = torch.cat((des19_svbrdf_te[:,:,0:3],des19_svbrdf_te[:,:,3:6]**(1/2.2),des19_svbrdf_te[:,:,6:9],des19_svbrdf_te[:,:,9:12]**(1/2.2),des19_svbrdf_te[:,:,3:6]**(1/2.2)),dim=1)
for key_ren, val_ren in val_rens.items():
# final_vis_te[key_ren] = torch.cat((final_vis_te[key_ren], val_ren[save_index,...]),dim=1)
for j in range(opt.N_input):
final_vis_te[key_ren] = torch.cat((final_vis_te[key_ren], val_ren[j,...]),dim=1)
for j in range(opt.N_input):
final_vis_te['gt'] = torch.cat((final_vis_te['gt'], gt_ren_te[j,...]),dim=1)
for key_ren, tr_ren in tr_rens.items():
for j in range(opt.N_input):
final_vis_tr[key_ren] = torch.cat((final_vis_tr[key_ren], tr_ren[j,...]),dim=1)
for j in range(opt.N_input):
final_vis_tr['gt'] = torch.cat((final_vis_tr['gt'], gt_ren_tr[j,...]),dim=1)
if opt.Nren_outerloss>0:
final_vis_tr['gt'] = torch.cat((final_vis_tr['gt'], gt_ren_sampled[0,...]),dim=1)
final_vis_tr[str(inner_steps)] = torch.cat((final_vis_tr[str(inner_steps)], pred_ren_sampled[0,...]),dim=1)
# save images
fig = plt.figure(figsize=(30,10))
for i, key in enumerate(final_vis_te):
print(i, key)
plt.subplot(len(final_vis_te), 1, int(i)+1)
plt.imshow((torch.clamp(final_vis_te[key], 0, 1).cpu().detach().numpy()*255.0).astype(np.uint8))
fig.savefig(os.path.join(save_val_img_path,'%d.png'%(step)))
plt.close()
# save images
fig = plt.figure(figsize=(30,10))
for i, key in enumerate(final_vis_tr):
print(i, key)
plt.subplot(len(final_vis_tr), 1, int(i)+1)
plt.imshow((torch.clamp(final_vis_tr[key], 0, 1).cpu().detach().numpy()*255.0).astype(np.uint8))
fig.savefig(os.path.join(save_train_img_path,'%d.png'%(step)))
plt.close()
del final_vis_tr
del final_vis_te
del val_feas
del val_rens
################# save logs ################
log_step.append(step)
# weight_list.append(weight_fea)
for k,v in val_rens_loss.items():
val_N_loss[k] = val_N_loss[k].div_(len(val_data))
val_D_loss[k] = val_D_loss[k].div_(len(val_data))
val_R_loss[k] = val_R_loss[k].div_(len(val_data))
val_S_loss[k] = val_S_loss[k].div_(len(val_data))
val_feas_loss[k] = val_feas_loss[k].div_(len(val_data))
val_rens_loss[k] = val_rens_loss[k].div_(len(val_data))
if k in val_log_fea:
val_log_fea[k].append(val_feas_loss[k].data.cpu().numpy())
else:
val_log_fea[k] = [val_feas_loss[k].data.cpu().numpy()]
if k in val_log_ren:
val_log_ren[k].append(val_rens_loss[k].data.cpu().numpy())
else:
val_log_ren[k] = [val_rens_loss[k].data.cpu().numpy()]
# trainfea_log.append(outer_loss.data.cpu().numpy())
# trainren_log.append(outer_ren_loss.data.cpu().numpy())
# save_loss(weight_list, save_loss_path, log_step, save_name='weight_fea')
save_loss(val_log_fea, save_loss_path, log_step, save_name='val_fea')
save_loss(val_log_ren, save_loss_path, log_step, save_name='val_ren')
# save_loss(trainfea_log, save_loss_path, log_step, save_name='train_fea')
# save_loss(trainren_log, save_loss_path, log_step, save_name='train_ren')
writer.add_scalar(f'valfea_1step_loss',val_feas_loss[f'{opt.inner_step}'].item(), step)
writer.add_scalar(f'valren_1step_loss',val_rens_loss[f'{opt.inner_step}'].item(), step)
writer.add_scalars(f'val_1step_loss',{'N': val_N_loss[f'{opt.inner_step}'].item()}, step)
writer.add_scalars(f'val_1step_loss',{'D': val_D_loss[f'{opt.inner_step}'].item()}, step)
writer.add_scalars(f'val_1step_loss',{'R':val_R_loss[f'{opt.inner_step}'].item()}, step)
writer.add_scalars(f'val_1step_loss',{'S':val_S_loss[f'{opt.inner_step}'].item()}, step)
writer.add_scalar('trainfe_loss',outer_loss.item(), step)
writer.add_scalar('trainren_loss',outer_ren_loss.item(), step)
temp_lr = opt_outer.param_groups[0]['lr']
print(f'step: {step:d}, outer_lr: {temp_lr:.8f}, wN: {opt.wN_outer:f}, wS: {opt.wS_outer:f}, wD: {opt.wD_outer:f}, wR: {opt.wR_outer:f},trfea: {outer_fea_loss:.5f},trfeavg: {outer_feavg_loss:.5f},trren: {outer_ren_loss:.5f}, valfea: {val_feas_loss[str(inner_steps)]:.5f}, valren: {val_rens_loss[str(inner_steps)]:.5f}')
print(f'Ren downsampled {opt.dres} L2 loss: {L_ren_loss:.5f}, TD loss:{TD_ren_loss:.5f}, VG loss:{VG_ren_loss:.5f}')
model.train()
# for p in criterionNet.parameters():
# p.requires_grad = True
# ------------------------------ for debuging ------------------------------------
else:
print(step)
# ------------------------------ debuging ------------------------------------
if step%10000==0 or step==1:
model_dict = {}
model_dict['model']=model.cpu().state_dict()
model.cuda()
if opt.netloss=='Des19Net':
model_dict['des19net']=criterionNet.lossnet.cpu().state_dict()
criterionNet.lossnet.cuda()
if opt.model=='OutConSiren':
model_dict['FeaExtractor']=FeaExtractor.cpu().state_dict()
FeaExtractor.cuda()
# model_dict['optimizer'] = opt_outer.state_dict()
save_all_models(model_dict, save_model_path, step)
del model_dict
if step%opt.decay_lr==0 and step!=0:
opt_outer.param_groups[0]['lr'] *= 0.5
step += 1
def pretrain(opt, model, coords, train_data, val_data, device, FeaExtractor=None):
coords = coords.repeat(opt.batch_size,1,1,1)
max_iters = opt.train_iter#500000 #@param {type:"integer"}
outer_lr = opt.outer_lr #@param
inner_lr = opt.inner_lr #@param
inner_steps = opt.inner_step #@param {type:"integer"}
val_steps = opt.val_step #@param {type:"integer"}
first_order = opt.first_order #@param {type:"boolean"}
model.train()
if opt.losstype=='L2':
print('######################### using L2 loss #######################')
criterion = torch.nn.MSELoss().to(device)
else:
print('######################### using L1 loss #######################')
criterion = torch.nn.L1Loss().to(device)
criterionTD = TDLoss_2(device, 0)
params = OrderedDict(model.named_parameters())
criterionVGG = VGGLoss(opt)
weights = [0.1, 0.1, 0.1, 0.1]
keys = [1, 2, 3,4]
print('------------------ no learn loss -----------------------')
opt_outer = torch.optim.Adam(model.parameters(), lr=outer_lr, betas=[opt.beta1,opt.beta2], eps=opt.eps_adam)
load_model='pretrain' if opt.load_pretrain else opt.model
train_img_dir = '%s/%s/train_imgs'%(opt.model, opt.name)
val_img_dir = '%s/%s/val_imgs'%(opt.model, opt.name)
ckpt_dir = '%s/%s/models'%(opt.model, opt.name)
loss_path = '%s/%s/loss'%(opt.model, opt.name)
if not os.path.exists(train_img_dir):
os.makedirs(train_img_dir)
if not os.path.exists(val_img_dir):
os.makedirs(val_img_dir)
if not os.path.exists(ckpt_dir):
os.makedirs(ckpt_dir)
if not os.path.exists(loss_path):
os.makedirs(loss_path)
if opt.resume_name!='':
print('resuming from ', opt.resume_name)
load_network(model,opt.load_iter,'%s/%s/models'%(load_model, opt.resume_name), name='model')
# load_network(opt_outer,opt.load_iter,'%s/%s/models'%(opt.model, opt.resume_name), name='optimizer')
if FeaExtractor is not None and not opt.load_pretrain:
load_network(FeaExtractor, opt.load_iter,'%s/%s/models'%(load_model, opt.resume_name), name='FeaExtractor')
if opt.netloss=='Des19Net' and not opt.load_pretrain:
load_network(criterionNet.lossnet, opt.load_iter,'%s/%s/models'%(load_model, opt.resume_name), name='des19net')
writer = SummaryWriter(loss_path)
step = 1
Position_map=PositionMap(opt.res,opt.res,3).to(device)
resize = transforms.Resize((opt.res,opt.res))
LightPos = torch.tensor([[0,0,4.0]], device=device)
if opt.model=='Siren':
myinput = coords
while step < max_iters:
for i,example in enumerate(train_data):
if step > max_iters:
break
params = OrderedDict(model.named_parameters())
model.zero_grad()
image = example.to(device)
outer_ren_loss = torch.tensor(0., device=device)
outer_fea_loss = torch.tensor(0., device=device)
image_tr = image.permute(0,2,3,1) #[B,H,W,C]
gt_ren_tr = SingleRender_NumberPointLight_FixedCamera(2*image_tr[...,0:3]-1, image_tr[...,3:6],image_tr[...,6:7], image_tr[...,9:12], LightPos, Position_map, device, CamLi_co=True) #[B, W,H,C]
gt_ren_tr = (gt_ren_tr.clamp(0,1)+opt.EPSILON)**(1/2.2)
myinput = 2*gt_ren_tr.permute(0,3,1,2)-1 #[B,3,H,W] [-1,1]
myinput = [coords,myinput] # coords:[H,W,2] myinput: [B,3,H,W]
## outer loop
pred_tr = model(myinput, params = params)
pred_tr = paramize_out(opt,pred_tr)
pred_ren_tr = SingleRender_NumberPointLight_FixedCamera(2*pred_tr[...,0:3]-1, pred_tr[...,3:6], pred_tr[...,6:7], pred_tr[...,9:12], LightPos, Position_map, device, CamLi_co=True) #[B,N, W,H,C]
pred_ren_tr = (pred_ren_tr.clamp(0,1)+opt.EPSILON)**(1/2.2)
outer_fea_loss = 0.25*(criterion(pred_tr[...,0:3], image_tr[...,0:3])*opt.wN_outer
+ criterion(pred_tr[...,3:6], image_tr[...,3:6])
+ criterion(pred_tr[...,6:9], image_tr[...,6:9])*opt.wR_outer
+ criterion(pred_tr[...,9:12], image_tr[...,9:12])*opt.wS_outer)
# add VGG loss to features
if opt.Wfea_vgg>0:
D_vg_pred = VGGpreprocess(pred_tr[...,3:6].permute(0,3,1,2))
D_vg_gt = VGGpreprocess(image_tr[...,3:6].permute(0,3,1,2))
D_vg = criterionVGG(D_vg_pred, D_vg_gt)*opt.Wfea_vgg
S_vg_pred = VGGpreprocess(pred_tr[...,9:12].permute(0,3,1,2))
S_vg_gt = VGGpreprocess(image_tr[...,9:12].permute(0,3,1,2))
S_vg = criterionVGG(S_vg_pred, S_vg_gt)*opt.Wfea_vgg
outer_feavg_loss = (D_vg + S_vg)*0.5
if opt.Nren_outerloss>0:
rand_light_pos = NLight_VaryH(opt.Nren_outerloss, 0.05, device)
lightinten_rand = rand_light_pos[0][-1].item()**2
gt_ren_sampled = SingleRender_NumberPointLight_FixedCamera(2*image_tr[...,0:3]-1, image_tr[...,3:6],image_tr[...,6:7], image_tr[...,9:12], rand_light_pos, Position_map, device, CamLi_co=True, no_spec=opt.no_spec) #[B,N, W,H,C]
gt_ren_sampled = (gt_ren_sampled.clamp(0,1)+opt.EPSILON)**(1/2.2)
pred_ren_sampled = SingleRender_NumberPointLight_FixedCamera(2*pred_tr[...,0:3]-1, pred_tr[...,3:6], pred_tr[...,6:7], pred_tr[...,9:12], rand_light_pos, Position_map, device, CamLi_co=True, no_spec=opt.no_spec) #[B,N, W,H,C]
pred_ren_sampled = (pred_ren_sampled.clamp(0,1)+opt.EPSILON)**(1/2.2)
out_ren = pred_ren_sampled[0:1,...].permute(0,3,1,2)
gt_ren = gt_ren_sampled[0:1,...].permute(0,3,1,2)
out_16 = nn.functional.interpolate(out_ren, size = opt.dres, mode='bilinear', align_corners=True)
gt_16 = nn.functional.interpolate(gt_ren, size = opt.dres, mode='bilinear', align_corners=True)
L_ren_loss = criterion(out_16, gt_16)*opt.Wdren_outer
TD_ren_loss = criterionTD(out_ren, gt_ren)*opt.WTDren_outer
outer_ren_loss = L_ren_loss + TD_ren_loss
outer_loss = outer_fea_loss + outer_ren_loss + outer_feavg_loss
opt_outer.zero_grad()
outer_loss.backward()
opt_outer.step()
if step % opt.save_freq == 0 or step==1:
if not opt.no_val:
model.eval()
val_ren_loss = 0
val_fea_loss = 0
for i,val_example in enumerate(val_data):
val_example,_ = val_example
val_img = val_example.to(device)
gt_fea_te = val_img.permute(0,2,3,1) # [B, 256, 256, C]
gt_ren_te = SingleRender_NumberPointLight_FixedCamera(2*gt_fea_te[...,0:3]-1, gt_fea_te[...,3:6],gt_fea_te[...,6:7], gt_fea_te[...,9:12], LightPos, Position_map, device, CamLi_co=True) #[B,N, W,H,C]
gt_ren_te = (gt_ren_te.clamp(0,1)+EPSILON)**(1/2.2)
myinput = 2*gt_ren_te.permute(0,3,1,2)-1 #[B,3,H,W] [-1,1]
myinput = [coords[0:1], myinput] # coords:[H,W,2] myinput: [B,3,H,W]
out_fea_te = model(myinput, params=params)
out_fea_te = paramize_out(opt,out_fea_te)
out_ren_te = SingleRender_NumberPointLight_FixedCamera(2*out_fea_te[...,0:3]-1, out_fea_te[...,3:6], out_fea_te[...,6:7], out_fea_te[...,9:12], LightPos, Position_map, device, CamLi_co=True) #[B,N, W,H,C]
out_ren_te = (out_ren_te.clamp(0,1)+EPSILON)**(1/2.2)
val_ren_loss += criterion(gt_ren_te, out_ren_te).data
# compute losses
val_fea_loss += criterion(gt_fea_te, out_fea_te).data
################# save imgs ################33
fea_tr = torch.cat([image_tr[0,:,:,0:3],image_tr[0,:,:,3:6]**(1/2.2),image_tr[0,:,:,6:9],image_tr[0,:,:,9:12]**(1/2.2)], dim=1)
pred_tr = torch.cat([pred_tr[0,:,:,0:3],pred_tr[0,:,:,3:6]**(1/2.2),pred_tr[0,:,:,6:9],pred_tr[0,:,:,9:12]**(1/2.2)], dim=1)
save_image(fea_tr.detach(), os.path.join(train_img_dir,f'{step}gt_fea_tr.png'))
save_image(pred_tr.detach(), os.path.join(train_img_dir,f'{step}out_fea_tr.png'))
save_image(gt_ren_sampled[0,...].detach(), os.path.join(train_img_dir,f'{step}gt_ren_tr.png'))
save_image(pred_ren_sampled[0,...].detach(), os.path.join(train_img_dir,f'{step}out_ren_tr.png'))
fea_tr = torch.cat([gt_fea_te[0,:,:,0:3],gt_fea_te[0,:,:,3:6]**(1/2.2),gt_fea_te[0,:,:,6:9],gt_fea_te[0,:,:,9:12]**(1/2.2)], dim=1)
pred_tr = torch.cat([out_fea_te[0,:,:,0:3],out_fea_te[0,:,:,3:6]**(1/2.2),out_fea_te[0,:,:,6:9],out_fea_te[0,:,:,9:12]**(1/2.2)], dim=1)
save_image(fea_tr.detach(), os.path.join(val_img_dir,f'{step}gt_fea_te.png'))
save_image(pred_tr.detach(), os.path.join(val_img_dir,f'{step}out_fea_te.png'))
save_image(gt_ren_te[0,...].detach(), os.path.join(val_img_dir,f'{step}gt_ren_te.png'))
save_image(out_ren_te[0,...].detach(), os.path.join(val_img_dir,f'{step}out_ren_te.png'))
################# save logs ################
val_fea_loss = val_fea_loss.div_(len(val_data))
val_ren_loss = val_ren_loss.div_(len(val_data))
writer.add_scalar('val_fea',val_fea_loss.item(), step)
writer.add_scalar('val_ren',val_ren_loss.item(), step)
writer.add_scalar('val_total',val_ren_loss.item()+val_fea_loss.item(), step)
writer.add_scalar('trainfea',outer_fea_loss.item(), step)
writer.add_scalar('trainren',outer_ren_loss.item(), step)
print(f'step: {step:d}, trfea: {outer_fea_loss:.5f},trren: {outer_ren_loss:.5f}, valfea: {val_fea_loss:.5f}, valren: {val_ren_loss:.5f}')
model.train()
# for p in criterionNet.parameters():
# p.requires_grad = True
# ------------------------------ for debuging ------------------------------------
else:
print(f'step: {step:d},tr_outer_fea: {outer_loss.item():.4f},tr_outer_ren: {outer_ren_loss.item():.4f}')
# ------------------------------ debuging ------------------------------------
if step%50000==0 or step==1:
model_dict = {}
model_dict['model']=model.cpu().state_dict()
model.cuda()
save_all_models(model_dict, ckpt_dir, step)
del model_dict
step += 1
def save_args(opt):
args = vars(opt)
save_opt_path = '%s/%s'%(opt.model, opt.name)
if not os.path.exists(save_opt_path):
os.makedirs(save_opt_path)
file_name = os.path.join(save_opt_path, 'opt.txt')
with open(file_name, 'wt') as opt_file:
opt_file.write('------------ Options -------------\n')
for k, v in sorted(args.items()):
opt_file.write('%s: %s\n' % (str(k), str(v)))
opt_file.write('-------------- End ----------------\n')
return
def filter_args(args):
if args.Wvgg_outer>0:
args.Nren_outerloss = 1
if args.mode=='pretrain':
args.model='InConSiren'
args.save_freq=1000
# args.outer_lr=1e-5
args.decay_lr=300000
return args
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--res', type=int, default=256, help='resolution')
parser.add_argument('--batch_size', type=int, default=3, help='batch_size')
parser.add_argument('--n_threads', type=int, default=0, help='number of n_threads')
parser.add_argument('--inner_step', type=int, default=1, help='# of inner step')
parser.add_argument('--val_step', type=int, default=1, help='# of inner step')
parser.add_argument('--load_iter', type=int, default=500000, help='# of iteration')
parser.add_argument('--train_iter', type=int, default=1000000, help='# of training iteration')
parser.add_argument('--save_freq', type=int, default=2000, help='frequency of saving images')
# parser.add_argument('--Light_range', type=float, default=0.0, help='range of input light')
parser.add_argument('--EPSILON', type=float, default=1e-6, help='eps value for rendering')
parser.add_argument('--lambda_decay', type=float, default=1.0, help='weight_decay of feature weight or stage decay')
parser.add_argument('--clip', type=float, default=0, help='gradient clipping')
parser.add_argument('--clip_hook', action='store_true', help='gradient clipping using hook')
parser.add_argument('--first_order', action='store_true', help='use the first order gradient')
parser.add_argument('--no_val', action='store_true', help='no validation dataset')
parser.add_argument('--inner_as_outer_loss', action='store_true', help='outer loss same as inner loss')
parser.add_argument('--Nren_outerloss', type=int, default=-1, help='# of rendered loss as outer loss')
parser.add_argument('--reverse_outer', action='store_true', help='reverse the feature lambda to render lambda')
parser.add_argument('--no_learn_loss', action='store_true', help='DO NOT learn loss function')
parser.add_argument('--no_spec', action='store_true', help='DO NOT use specular map')
parser.add_argument('--gamma', action='store_true', help='gamma output D and S')
# for training
parser.add_argument('--inner_lr', type=float, default=0.001, help='the inner lr')
parser.add_argument('--vary_light', action='store_true', help='vary_light')
parser.add_argument('--wN_outer', type=float, default=1, help='weight of normal in the outer loss')
parser.add_argument('--wD_outer', type=float, default=1, help='weight of diffuse in the outer loss')
parser.add_argument('--wR_outer', type=float, default=1, help='weight of rougness in the outer loss')
parser.add_argument('--wS_outer', type=float, default=1, help='weight of specular in the outer loss')
parser.add_argument('--HN_factor', type=float, default=10.0, help='height to normal scale factor')
parser.add_argument('--lossnet_lr', type=float, default=1e-5, help='the lr of loss net')
parser.add_argument('--outer_lr', type=float, default=1e-6, help='the outer lr')
parser.add_argument('--decay_lr', type=int, default=200000, help='decay outer learning rate every')
parser.add_argument('--eps_adam', type=float, default=1e-8, help='eps of adam optimzer')
parser.add_argument('--losstype', type=str, default='L2', help='use L2 or L1 loss')
parser.add_argument('--beta1', type=float, default=0.5, help='beta1 of Adam optimizer')
parser.add_argument('--beta2', type=float, default=0.5, help='beta2 of Adam optimizer')
parser.add_argument('--eta1', type=float, default=0.5, help='eta1 of rprop optimizer')
parser.add_argument('--eta2', type=float, default=1.2, help='eta2 of rprop optimizer')
parser.add_argument('--step_min', type=float, default=1e-6, help='minimum step of rprop')
parser.add_argument('--step_max', type=float, default=50, help='maximum step of rprop')
parser.add_argument('--Wren', type=float, default=1.0, help='weight of rendering loss')
parser.add_argument('--N_input', type=int, default=1, help='number of input images')
parser.add_argument('--Wvgg_outer', type=float, default=-1, help='add vgg for rendering outer')
parser.add_argument('--WdownL2_inner', type=float, default=-1, help='weight of downL2 inner')
parser.add_argument('--Wvgg_inner', type=float, default=-1, help='add vgg for rendering inner')
parser.add_argument('--Wtd_inner', type=float, default=-1, help='add td for rendering inner')
parser.add_argument('--Winner_ren', type=float, default=1, help='weight of inner rendering loss')
parser.add_argument('--no_inner', action='store_true', help='not use inner step')
parser.add_argument('--Wfea_vgg', type=float, default=-1, help='add vgg of feature maps in outer loop')
parser.add_argument('--use_TD_loss', action='store_true', help='use TD loss and downsampled L1 loss')
parser.add_argument('--WTDren_outer', type=float, default=1, help='weight of TD for rendering outer')
parser.add_argument('--Wdren_outer', type=float, default=1, help='weight of downsampled L2 for rendering outer')
parser.add_argument('--dres', type=int, default=16, help='downsampled res for L1')
# for dataset
parser.add_argument('--const_weightfea', type=float, default=-1, help='const weight feature: -1: not use const | >=0: use const')
parser.add_argument('--train_root', type=str, default='D:/XilongZhou/Research/Research/dataset/DeepMaterialsData/ImageData/SynTrainData', help='input image name')
parser.add_argument('--val_root', type=str, default='D:/XilongZhou/Research/Research/dataset/DeepMaterialsData/ImageData/SynTestData2', help='input image name')
parser.add_argument('--exp', type=str, default='SVBRDF', help='input image name')
parser.add_argument('--name', type=str, default='SVBRDF', help='name of folder')
parser.add_argument('--mode', type=str, default='meta', help='meta || pretrain || test')
parser.add_argument('--optim', type=str, default='adam', help='which optimzer to use? adam || sgd || rprop')
parser.add_argument('--fea', type=str,default='all_N1', help='use all feature maps or just albedo')
parser.add_argument('--netloss', type=str, default='Des19Net',help='train network loss function')
parser.add_argument('--Des19Net_npy_path', type=str, default='../Dataset/Des19Net/Des19.npy', help='npy path for des19 ')
parser.add_argument('--resume_name', type=str, default='', help='resume training from folder wth name')
parser.add_argument('--load_pretrain', action='store_true', help='load the pretrained model')
parser.add_argument('--add_16L2', action='store_true', help='load the pretrained model')
parser.add_argument('--fixed_inten', action='store_true', help='fix light intensity or not')
parser.add_argument('--sc_des19', action='store_true', help='scale for des19 network')
parser.add_argument('--no_fea_outer', action='store_true', help='not using feature loss in outer loop')
# network architecture
parser.add_argument('--num_layers', type=int, default=5, help='# of layers')
parser.add_argument('--dim_c', type=int, default=8, help='# of hidden channels in UNet')
parser.add_argument('--extfea_c', type=int, default=32, help='# of output channels in feature extractor')
parser.add_argument('--model', type=str, default='InConSiren', help='use Siren or Fourier Feature')
parser.add_argument('--cond_type', type=str, default='unet', help='unet or conv')
parser.add_argument('--n_layer_unet', type=int, default=4, help='the # of layer in UNet')
parser.add_argument('--branch', action='store_true', help='use branch for network')
parser.add_argument('--no_coords', action='store_true', help='no_coords input')
parser.add_argument('--branch_len', type=int, default=2, help='use branch length for network')
# for debug
parser.add_argument('--file', type=str, default='filesall', help='name of folder')
parser.add_argument('--test_img', type=str,default='Real2', help='test images')
parser.add_argument('--meta_debug', type=str,default='', help='debugging')
opt = parser.parse_args()
opt = filter_args(opt)
save_args(opt)
torch.set_num_threads(8)
num_val_exs = 5
CROP_SIZE = opt.res
train_data_root = 'data/celeba'
val_data_root = 'data/celeba_val'
mytraindata_root = opt.train_root
myvaldata_root = opt.val_root
Mytrain_set = DataLoaderHelper(mytraindata_root, opt)
Mytrain_data = DataLoader(dataset=Mytrain_set, batch_size=opt.batch_size, num_workers=opt.n_threads, shuffle=True, drop_last=True)
Myval_set = DataLoaderHelper_test(myvaldata_root, opt)
Myval_data = DataLoader(dataset=Myval_set, batch_size=1, num_workers=0, shuffle=True, drop_last=False)
if opt.exp=='SVBRDF':
train_dataloader = Mytrain_data
val_dataloader = Myval_data
##########################################################################################################
##########################################################################################################
device = torch.device('cuda')
x1 = torch.linspace(0, 1, opt.res+1)[:-1]
coords = torch.stack(torch.meshgrid([x1,x1]), dim=-1)[None,...]
coords = coords.to(device)
if opt.fea=='all_N1':
out_nc = 8
elif opt.fea=='all_N2':
out_nc = 9
if opt.no_spec:
out_nc = out_nc - 3
FeaExtractNet = None
if opt.model=='Siren':
model = SirenNet(2, 256, out_nc, opt.num_layers, w0_initial=200., w0=200.,
final_activation=lambda x: x + .5).to(device)
# Featuer Extractor not in inner loop
elif opt.model=='OutConSiren':
FeaExtractNet = FeatureExtractor(opt.extfea_c, cond_type=opt.cond_type).to(device)
model = SirenNet(2+opt.extfea_c, 256, out_nc, opt.num_layers, w0_initial=200., w0=200.,
final_activation=lambda x: x + .5).to(device)
# Featuer Extractor in inner loop
elif opt.model=='InConSiren':
print('InConSiren model...')
model = ConSirenNet(opt, 2, 256, out_nc, opt.num_layers, w0_initial=200., w0=200.,
final_activation=lambda x: x + .5, cond_type=opt.cond_type, N_in=opt.N_input, n_layer_unet=opt.n_layer_unet).to(device)
elif opt.model=='UNet':
model = MaterialMetaUNet(layer_n=opt.n_layer_unet).to(device)
print(model)
if FeaExtractNet is not None:
print(FeaExtractNet)
print('output nc:', out_nc)
print('model N params', sum(p.numel() for n,p in model.meta_named_parameters() if p.requires_grad))
if opt.mode=='meta':
meta_learn(opt, model, coords, Mytrain_data, Myval_data, device, FeaExtractNet)
elif opt.mode=='pretrain':
pretrain(opt, model, coords, Mytrain_data, Myval_data, device, FeaExtractNet)