-
Notifications
You must be signed in to change notification settings - Fork 2
/
run.py
1116 lines (982 loc) · 50 KB
/
run.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, sys, copy, glob, json, time, random, argparse, gc
from shutil import copyfile
from tqdm import tqdm, trange
from copy import deepcopy
import mmcv
import imageio
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from lib import utils, dvgo, dcvgo, dmpigo
from lib.load_data import load_data
from segment_3d import kmeans, query_kmeans
from lib.region_grower import dev_region_grower_mask
from torch_efficient_distloss import flatten_eff_distloss
from lib.grid import DenseGrid, TensoRFGrid, get_dense_grid_batch_processing
def config_parser():
'''Define command line arguments
'''
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--config', required=True,
help='config file path')
parser.add_argument("--seed", type=int, default=777,
help='Random seed')
parser.add_argument("--no_reload", action='store_true',
help='do not reload weights from saved ckpt')
parser.add_argument("--no_reload_optimizer", action='store_true',
help='do not reload optimizer state from saved ckpt')
parser.add_argument("--ft_path", type=str, default='',
help='specific weights npy file to reload for coarse network')
parser.add_argument("--export_bbox_and_cams_only", type=str, default='',
help='export scene bbox and camera poses for debugging and 3d visualization')
parser.add_argument("--export_coarse_only", type=str, default='')
# testing options
parser.add_argument("--render_only", action='store_true',
help='do not optimize, reload weights and render out render_poses path')
parser.add_argument("--render_test", action='store_true')
parser.add_argument("--render_train", action='store_true')
parser.add_argument("--render_video", action='store_true')
parser.add_argument("--render_segment", action='store_true')
parser.add_argument("--render_video_flipy", action='store_true')
parser.add_argument("--render_video_rot90", default=0, type=int)
parser.add_argument("--render_video_factor", type=float, default=0,
help='downsampling factor to speed up rendering, set 4 or 8 for fast preview')
parser.add_argument("--dump_images", action='store_true')
parser.add_argument("--eval_ssim", action='store_true')
parser.add_argument("--eval_lpips_alex", action='store_true')
parser.add_argument("--eval_lpips_vgg", action='store_true')
# logging/saving options
parser.add_argument("--i_print", type=int, default=500,
help='frequency of console printout and metric loggin')
parser.add_argument("--i_weights", type=int, default=5000,
help='frequency of weight ckpt saving')
# arguments for feature distillation
parser.add_argument("--freeze_density", action='store_true',
help='freeze density grid')
parser.add_argument("--freeze_rgb", action='store_true',
help='freeze rgb grid and mlp')
parser.add_argument("--freeze_feature", action='store_true',
help='freeze feature grid')
parser.add_argument("--only_distill_loss", action='store_true',
help='train on only loss of features')
parser.add_argument("--weighted_distill_loss", action='store_true',
help='train on weighted loss')
parser.add_argument("--distill_active", action='store_true',
help='enable learning with feature distillation.')
parser.add_argument("--distill_lambda", type=float, default=1e-3,
help='weight for feature distillation loss.')
parser.add_argument("--no_viewdirs_distill", action='store_true',
help='disable viewing direction for distillation layer.')
parser.add_argument("--segment", action='store_true',
help='interactively set threshold and re-render until stopped.')
parser.add_argument("--fv", type=str,
help='file path for the feature vector to use for segmentation')
parser.add_argument("--stop_at", type=int,
help='at what iteration to stop training.')
parser.add_argument("--dino_dim", type=int, default=64,
help='number of dimensions to use for segmentaiton in the interactive mode.')
return parser
@torch.no_grad()
def render_viewpoints(model, render_poses, HW, Ks, ndc, render_kwargs,
gt_imgs=None, savedir=None, dump_images=False, cfg=None,
render_factor=0, render_video_flipy=False, render_video_rot90=0,
eval_ssim=False, eval_lpips_alex=False, eval_lpips_vgg=False, distill_active=False, render_fct=0.0):
'''Render images for the given viewpoints; run evaluation if gt given.
'''
assert len(render_poses) == len(HW) and len(HW) == len(Ks)
if render_factor!=0:
HW = np.copy(HW)
Ks = np.copy(Ks)
HW = (HW/render_factor).astype(int)
Ks[:, :2, :3] /= render_factor
rgbs = []
features = []
depths = []
bgmaps = []
psnrs = []
ssims = []
lpips_alex = []
lpips_vgg = []
for i, c2w in enumerate(tqdm(render_poses)):
H, W = HW[i]
K = Ks[i]
c2w = torch.Tensor(c2w)
rays_o, rays_d, viewdirs = dvgo.get_rays_of_a_view(
H, W, K, c2w, ndc, inverse_y=render_kwargs['inverse_y'],
flip_x=cfg.data.flip_x, flip_y=cfg.data.flip_y)
keys = ['rgb_marched', 'depth', 'alphainv_last']
if distill_active: keys.append('f_marched')
rays_o = rays_o.flatten(0,-2)
rays_d = rays_d.flatten(0,-2)
viewdirs = viewdirs.flatten(0,-2)
render_result_chunks = [
{k: v for k, v in model(ro, rd, vd, distill_active=distill_active, render_fct=render_fct, **render_kwargs).items() if k in keys}
for ro, rd, vd in zip(rays_o.split(8192, 0), rays_d.split(8192, 0), viewdirs.split(8192, 0))
]
render_result = {
k: torch.cat([ret[k] for ret in render_result_chunks]).reshape(H,W,-1)
for k in render_result_chunks[0].keys()
}
rgb = render_result['rgb_marched'].cpu().numpy()
if distill_active:
feature = render_result['f_marched'].cpu()
else:
feature = None
depth = render_result['depth'].cpu().numpy()
bgmap = render_result['alphainv_last'].cpu().numpy()
rgbs.append(rgb)
if distill_active:
features.append(feature)
depths.append(depth)
bgmaps.append(bgmap)
if i==0:
print('Testing', rgb.shape)
if gt_imgs is not None and render_factor==0:
p = -10. * np.log10(np.mean(np.square(rgb - gt_imgs[i])))
psnrs.append(p)
if eval_ssim:
ssims.append(utils.rgb_ssim(rgb, gt_imgs[i], max_val=1))
if eval_lpips_alex:
lpips_alex.append(utils.rgb_lpips(rgb, gt_imgs[i], net_name='alex', device=c2w.device))
if eval_lpips_vgg:
lpips_vgg.append(utils.rgb_lpips(rgb, gt_imgs[i], net_name='vgg', device=c2w.device))
if len(psnrs):
print('Testing psnr', np.mean(psnrs), '(avg)')
if eval_ssim: print('Testing ssim', np.mean(ssims), '(avg)')
if eval_lpips_vgg: print('Testing lpips (vgg)', np.mean(lpips_vgg), '(avg)')
if eval_lpips_alex: print('Testing lpips (alex)', np.mean(lpips_alex), '(avg)')
if render_video_flipy:
for i in range(len(rgbs)):
rgbs[i] = np.flip(rgbs[i], axis=0)
depths[i] = np.flip(depths[i], axis=0)
bgmaps[i] = np.flip(bgmaps[i], axis=0)
if render_video_rot90 != 0:
for i in range(len(rgbs)):
rgbs[i] = np.rot90(rgbs[i], k=render_video_rot90, axes=(0,1))
depths[i] = np.rot90(depths[i], k=render_video_rot90, axes=(0,1))
bgmaps[i] = np.rot90(bgmaps[i], k=render_video_rot90, axes=(0,1))
if savedir is not None and dump_images:
for i in trange(len(rgbs)):
rgb8 = utils.to8b(rgbs[i])
filename = os.path.join(savedir, '{:03d}.png'.format(i))
imageio.imwrite(filename, rgb8)
for i in trange(len(features)):
filename = os.path.join(savedir, '{:03d}.pt'.format(i))
feature = features[i].to(torch.device("cpu"))
torch.save(feature, filename)
rgbs = np.array(rgbs)
depths = np.array(depths)
bgmaps = np.array(bgmaps)
# features = np.array(torch.stack(features))
if len(features): features = np.stack(features)
return rgbs, depths, bgmaps, features
def seed_everything(args):
'''Seed everything for better reproducibility.
(some pytorch operation is non-deterministic like the backprop of grid_samples)
'''
torch.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
def load_everything(args, cfg):
'''Load images / poses / camera settings / data split.
'''
cfg.data['distill_active'] = args.distill_active
data_dict = load_data(cfg.data)
# remove useless field
kept_keys = {
'hwf', 'HW', 'Ks', 'near', 'far', 'near_clip',
'i_train', 'i_val', 'i_test', 'irregular_shape',
'poses', 'render_poses', 'images', 'features'}
for k in list(data_dict.keys()):
if k not in kept_keys:
data_dict.pop(k)
# construct data tensor
if data_dict['irregular_shape']:
data_dict['images'] = [torch.FloatTensor(im, device='cpu') for im in data_dict['images']]
if args.distill_active:
data_dict['features'] = [torch.FloatTensor(im, device='cpu') for im in data_dict['features']]
else:
data_dict['images'] = torch.FloatTensor(data_dict['images'], device='cpu')
if args.distill_active:
if not isinstance(data_dict['features'], torch.Tensor):
data_dict['features'] = torch.FloatTensor(data_dict['features'], device='cpu')
data_dict['poses'] = torch.Tensor(data_dict['poses'])
return data_dict
def _compute_bbox_by_cam_frustrm_bounded(cfg, HW, Ks, poses, i_train, near, far):
xyz_min = torch.Tensor([np.inf, np.inf, np.inf])
xyz_max = -xyz_min
for (H, W), K, c2w in zip(HW[i_train], Ks[i_train], poses[i_train]):
rays_o, rays_d, viewdirs = dvgo.get_rays_of_a_view(
H=H, W=W, K=K, c2w=c2w,
ndc=cfg.data.ndc, inverse_y=cfg.data.inverse_y,
flip_x=cfg.data.flip_x, flip_y=cfg.data.flip_y)
if cfg.data.ndc:
pts_nf = torch.stack([rays_o+rays_d*near, rays_o+rays_d*far])
else:
pts_nf = torch.stack([rays_o+viewdirs*near, rays_o+viewdirs*far])
xyz_min = torch.minimum(xyz_min, pts_nf.amin((0,1,2)))
xyz_max = torch.maximum(xyz_max, pts_nf.amax((0,1,2)))
return xyz_min, xyz_max
def _compute_bbox_by_cam_frustrm_unbounded(cfg, HW, Ks, poses, i_train, near_clip):
# Find a tightest cube that cover all camera centers
xyz_min = torch.Tensor([np.inf, np.inf, np.inf])
xyz_max = -xyz_min
for (H, W), K, c2w in zip(HW[i_train], Ks[i_train], poses[i_train]):
rays_o, rays_d, viewdirs = dvgo.get_rays_of_a_view(
H=H, W=W, K=K, c2w=c2w,
ndc=cfg.data.ndc, inverse_y=cfg.data.inverse_y,
flip_x=cfg.data.flip_x, flip_y=cfg.data.flip_y)
pts = rays_o + rays_d * near_clip
xyz_min = torch.minimum(xyz_min, pts.amin((0,1)))
xyz_max = torch.maximum(xyz_max, pts.amax((0,1)))
center = (xyz_min + xyz_max) * 0.5
radius = (center - xyz_min).max() * cfg.data.unbounded_inner_r
xyz_min = center - radius
xyz_max = center + radius
return xyz_min, xyz_max
def compute_bbox_by_cam_frustrm(args, cfg, HW, Ks, poses, i_train, near, far, **kwargs):
print('compute_bbox_by_cam_frustrm: start')
if cfg.data.unbounded_inward:
xyz_min, xyz_max = _compute_bbox_by_cam_frustrm_unbounded(
cfg, HW, Ks, poses, i_train, kwargs.get('near_clip', None))
else:
xyz_min, xyz_max = _compute_bbox_by_cam_frustrm_bounded(
cfg, HW, Ks, poses, i_train, near, far)
print('compute_bbox_by_cam_frustrm: xyz_min', xyz_min)
print('compute_bbox_by_cam_frustrm: xyz_max', xyz_max)
print('compute_bbox_by_cam_frustrm: finish')
return xyz_min, xyz_max
@torch.no_grad()
def compute_bbox_by_coarse_geo(model_class, model_path, thres):
print('compute_bbox_by_coarse_geo: start')
eps_time = time.time()
model = utils.load_model(model_class, model_path)
interp = torch.stack(torch.meshgrid(
torch.linspace(0, 1, model.world_size[0]),
torch.linspace(0, 1, model.world_size[1]),
torch.linspace(0, 1, model.world_size[2]),
), -1)
dense_xyz = model.xyz_min * (1-interp) + model.xyz_max * interp
density = model.density(dense_xyz)
alpha = model.activate_density(density)
mask = (alpha > thres)
active_xyz = dense_xyz[mask]
xyz_min = active_xyz.amin(0)
xyz_max = active_xyz.amax(0)
print('compute_bbox_by_coarse_geo: xyz_min', xyz_min)
print('compute_bbox_by_coarse_geo: xyz_max', xyz_max)
eps_time = time.time() - eps_time
print('compute_bbox_by_coarse_geo: finish (eps time:', eps_time, 'secs)')
return xyz_min, xyz_max
def create_new_model(cfg, cfg_model, cfg_train, xyz_min, xyz_max, stage, coarse_ckpt_path):
model_kwargs = copy.deepcopy(cfg_model)
num_voxels = model_kwargs.pop('num_voxels')
if len(cfg_train.pg_scale):
num_voxels = int(num_voxels / (2**len(cfg_train.pg_scale)))
model_kwargs['f_num_voxels'] = int(model_kwargs['f_num_voxels'] / (2**len(cfg_train.pg_scale)))
if cfg.data.ndc:
#print(f'scene_rep_reconstruction ({stage}): \033[96muse multiplane images\033[0m')
#model = dmpigo.DirectMPIGO(
# xyz_min=xyz_min, xyz_max=xyz_max,
# num_voxels=num_voxels,
# **model_kwargs)
print(f'scene_rep_reconstruction ({stage}): \033[96muse dense voxel grid\033[0m')
model = dvgo.DirectVoxGO(
xyz_min=xyz_min, xyz_max=xyz_max,
num_voxels=num_voxels,
mask_cache_path=coarse_ckpt_path,
**model_kwargs)
elif cfg.data.unbounded_inward:
print(f'scene_rep_reconstruction ({stage}): \033[96muse contraced voxel grid (covering unbounded)\033[0m')
model = dcvgo.DirectContractedVoxGO(
xyz_min=xyz_min, xyz_max=xyz_max,
num_voxels=num_voxels,
**model_kwargs)
else:
print(f'scene_rep_reconstruction ({stage}): \033[96muse dense voxel grid\033[0m')
model = dvgo.DirectVoxGO(
xyz_min=xyz_min, xyz_max=xyz_max,
num_voxels=num_voxels,
mask_cache_path=coarse_ckpt_path,
**model_kwargs)
model = model.to(device)
optimizer = utils.create_optimizer_or_freeze_model(model, cfg_train, global_step=0)
return model, optimizer
def load_existed_model(args, cfg, cfg_train, reload_ckpt_path, device):
if cfg.data.ndc:
model_class = dvgo.DirectVoxGO
#model_class = dmpigo.DirectMPIGO
elif cfg.data.unbounded_inward:
model_class = dcvgo.DirectContractedVoxGO
else:
model_class = dvgo.DirectVoxGO
model = utils.load_model(model_class, reload_ckpt_path).to(device)
optimizer = utils.create_optimizer_or_freeze_model(model, cfg_train, global_step=0)
model, optimizer, start = utils.load_checkpoint(
model, optimizer, reload_ckpt_path, args.no_reload_optimizer)
return model, optimizer, start
def scene_rep_reconstruction(args, cfg, cfg_model, cfg_train, xyz_min, xyz_max, data_dict, stage, coarse_ckpt_path=None):
# init
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
if abs(cfg_model.world_bound_scale - 1) > 1e-9:
xyz_shift = (xyz_max - xyz_min) * (cfg_model.world_bound_scale - 1) / 2
xyz_min -= xyz_shift
xyz_max += xyz_shift
HW, Ks, near, far, i_train, i_val, i_test, poses, render_poses, images, features = [
data_dict[k] for k in [
'HW', 'Ks', 'near', 'far', 'i_train', 'i_val', 'i_test', 'poses', 'render_poses', 'images', 'features'
]
]
# find whether there is existing checkpoint path
last_ckpt_path = os.path.join(cfg.basedir, cfg.expname, f'{stage}_last.tar')
if args.no_reload:
reload_ckpt_path = None
elif args.ft_path:
reload_ckpt_path = args.ft_path
elif os.path.isfile(last_ckpt_path):
reload_ckpt_path = last_ckpt_path
else:
reload_ckpt_path = None
# init model and optimizer
if reload_ckpt_path is None:
print(f'scene_rep_reconstruction ({stage}): train from scratch')
model, optimizer = create_new_model(cfg, cfg_model, cfg_train, xyz_min, xyz_max, stage, coarse_ckpt_path)
start = 0
if cfg_model.maskout_near_cam_vox:
model.maskout_near_cam_vox(poses[i_train,:3,3], near)
else:
print(f'scene_rep_reconstruction ({stage}): reload from {reload_ckpt_path}')
model, optimizer, start = load_existed_model(args, cfg, cfg_train, reload_ckpt_path, device)
if args.freeze_density:
for param in model.named_parameters():
if 'density' in param[0]:
param[1].requires_grad = False
if args.freeze_rgb:
for param in model.named_parameters():
if 'rgbnet' in param[0] or ('f_k0' not in param[0] and 'k0' in param[0]):
param[1].requires_grad = False
if args.freeze_feature:
for param in model.named_parameters():
if 'f_k0' in param[0]:
param[1].requires_grad = False
# init rendering setup
render_kwargs = {
'near': data_dict['near'],
'far': data_dict['far'],
'bg': 1 if cfg.data.white_bkgd else 0,
'rand_bkgd': cfg.data.rand_bkgd,
'stepsize': cfg_model.stepsize,
'inverse_y': cfg.data.inverse_y,
'flip_x': cfg.data.flip_x,
'flip_y': cfg.data.flip_y,
}
# init batch rays sampler
def gather_training_rays():
if data_dict['irregular_shape']:
rgb_tr_ori = [images[i].to('cpu' if cfg.data.load2gpu_on_the_fly else device) for i in i_train]
if args.distill_active:
f_tr_ori = [features[i].to('cpu' if cfg.data.load2gpu_on_the_fly else device) for i in i_train]
else:
rgb_tr_ori = images[i_train].to('cpu' if cfg.data.load2gpu_on_the_fly else device)
if args.distill_active:
f_tr_ori = features[i_train].to('cpu' if cfg.data.load2gpu_on_the_fly else device)
if cfg_train.ray_sampler == 'in_maskcache':
if args.distill_active:
rgb_tr, f_tr, rays_o_tr, rays_d_tr, viewdirs_tr, imsz = dvgo.get_training_rays_in_maskcache_sampling(
rgb_tr_ori=rgb_tr_ori,
f_tr_ori = f_tr_ori,
train_poses=poses[i_train],
HW=HW[i_train], Ks=Ks[i_train],
ndc=cfg.data.ndc, inverse_y=cfg.data.inverse_y,
flip_x=cfg.data.flip_x, flip_y=cfg.data.flip_y,
model=model, render_kwargs=render_kwargs)
else:
rgb_tr, rays_o_tr, rays_d_tr, viewdirs_tr, imsz = dvgo.get_training_rays_in_maskcache_sampling(
rgb_tr_ori=rgb_tr_ori,
train_poses=poses[i_train],
HW=HW[i_train], Ks=Ks[i_train],
ndc=cfg.data.ndc, inverse_y=cfg.data.inverse_y,
flip_x=cfg.data.flip_x, flip_y=cfg.data.flip_y,
model=model, render_kwargs=render_kwargs)
elif cfg_train.ray_sampler == 'flatten':
if args.distill_active:
rgb_tr, f_tr, rays_o_tr, rays_d_tr, viewdirs_tr, imsz = dvgo.get_training_rays_flatten(
rgb_tr_ori=rgb_tr_ori,
f_tr_ori=f_tr_ori,
train_poses=poses[i_train],
HW=HW[i_train], Ks=Ks[i_train], ndc=cfg.data.ndc, inverse_y=cfg.data.inverse_y,
flip_x=cfg.data.flip_x, flip_y=cfg.data.flip_y)
else:
rgb_tr, rays_o_tr, rays_d_tr, viewdirs_tr, imsz = dvgo.get_training_rays_flatten(
rgb_tr_ori=rgb_tr_ori,
train_poses=poses[i_train],
HW=HW[i_train], Ks=Ks[i_train], ndc=cfg.data.ndc, inverse_y=cfg.data.inverse_y,
flip_x=cfg.data.flip_x, flip_y=cfg.data.flip_y)
else:
rgb_tr, rays_o_tr, rays_d_tr, viewdirs_tr, imsz = dvgo.get_training_rays(
rgb_tr=rgb_tr_ori,
train_poses=poses[i_train],
HW=HW[i_train], Ks=Ks[i_train], ndc=cfg.data.ndc, inverse_y=cfg.data.inverse_y,
flip_x=cfg.data.flip_x, flip_y=cfg.data.flip_y)
if args.distill_active:
f_tr = f_tr_ori
index_generator = dvgo.batch_indices_generator(len(rgb_tr), cfg_train.N_rand)
batch_index_sampler = lambda: next(index_generator)
if args.distill_active:
return rgb_tr, f_tr, rays_o_tr, rays_d_tr, viewdirs_tr, imsz, batch_index_sampler
else:
return rgb_tr, rays_o_tr, rays_d_tr, viewdirs_tr, imsz, batch_index_sampler
if args.distill_active:
rgb_tr, f_tr, rays_o_tr, rays_d_tr, viewdirs_tr, imsz, batch_index_sampler = gather_training_rays()
else:
rgb_tr, rays_o_tr, rays_d_tr, viewdirs_tr, imsz, batch_index_sampler = gather_training_rays()
# view-count-based learning rate
if cfg_train.pervoxel_lr:
def per_voxel_init():
f_cnt = None
if args.distill_active:
cnt, f_cnt = model.voxel_count_views(
rays_o_tr=rays_o_tr, rays_d_tr=rays_d_tr, imsz=imsz, near=near, far=far,
stepsize=cfg_model.stepsize, downrate=cfg_train.pervoxel_lr_downrate,
irregular_shape=data_dict['irregular_shape'], distill_active=True)
else:
cnt = model.voxel_count_views(
rays_o_tr=rays_o_tr, rays_d_tr=rays_d_tr, imsz=imsz, near=near, far=far,
stepsize=cfg_model.stepsize, downrate=cfg_train.pervoxel_lr_downrate,
irregular_shape=data_dict['irregular_shape'])
optimizer.set_pervoxel_lr(cnt, f_cnt)
model.mask_cache.mask[cnt.squeeze() <= 2] = False
per_voxel_init()
if cfg_train.maskout_lt_nviews > 0:
model.update_occupancy_cache_lt_nviews(
rays_o_tr, rays_d_tr, imsz, render_kwargs, cfg_train.maskout_lt_nviews)
# GOGO
torch.cuda.empty_cache()
psnr_lst = []
time0 = time.time()
global_step = -1
for global_step in trange(1+start, 1+args.stop_at):
# renew occupancy grid
if model.mask_cache is not None and (global_step + 500) % 1000 == 0:
model.update_occupancy_cache()
# progress scaling checkpoint
if global_step in cfg_train.pg_scale:
n_rest_scales = len(cfg_train.pg_scale)-cfg_train.pg_scale.index(global_step)-1
cur_voxels = int(cfg_model.num_voxels / (2**n_rest_scales))
f_cur_voxels = int(cfg_model.f_num_voxels / (2**n_rest_scales))
if isinstance(model, (dvgo.DirectVoxGO, dcvgo.DirectContractedVoxGO)):
model.scale_volume_grid(cur_voxels, f_cur_voxels)
elif isinstance(model, dmpigo.DirectMPIGO):
model.scale_volume_grid(cur_voxels, model.mpi_depth)
else:
raise NotImplementedError
optimizer = utils.create_optimizer_or_freeze_model(model, cfg_train, global_step=0)
#optimizer = torch.optim.Adam(model.parameters())
model.act_shift -= cfg_train.decay_after_scale
torch.cuda.empty_cache()
# random sample rays
if cfg_train.ray_sampler in ['flatten', 'in_maskcache']:
sel_i = batch_index_sampler()
target = rgb_tr[sel_i]
if args.distill_active: f_target = f_tr[sel_i]
rays_o = rays_o_tr[sel_i]
rays_d = rays_d_tr[sel_i]
viewdirs = viewdirs_tr[sel_i]
elif cfg_train.ray_sampler == 'random':
sel_b = torch.randint(rgb_tr.shape[0], [cfg_train.N_rand])
sel_r = torch.randint(rgb_tr.shape[1], [cfg_train.N_rand])
sel_c = torch.randint(rgb_tr.shape[2], [cfg_train.N_rand])
target = rgb_tr[sel_b, sel_r, sel_c]
if args.distill_active: f_target = f_tr[sel_b, sel_r, sel_c]
rays_o = rays_o_tr[sel_b, sel_r, sel_c]
rays_d = rays_d_tr[sel_b, sel_r, sel_c]
viewdirs = viewdirs_tr[sel_b, sel_r, sel_c]
else:
raise NotImplementedError
if cfg.data.load2gpu_on_the_fly:
target = target.to(device)
if args.distill_active: f_target = f_target.to(device)
rays_o = rays_o.to(device)
rays_d = rays_d.to(device)
viewdirs = viewdirs.to(device)
# volume rendering
render_result = model(
rays_o, rays_d, viewdirs,
global_step=global_step, is_train=True,
distill_active=args.distill_active,
**render_kwargs)
# gradient descent step
optimizer.zero_grad(set_to_none=True)
if args.only_distill_loss:
f_loss = F.mse_loss(render_result['f_marched'], f_target)
loss = f_loss
psnr = utils.mse2psnr(loss.detach())
elif args.weighted_distill_loss:
f_loss = cfg_train.weight_main * F.mse_loss(render_result['f_marched'], f_target)
loss = cfg_train.weight_main * F.mse_loss(render_result['rgb_marched'], target)
psnr = utils.mse2psnr(loss.detach())
loss += 0.001 * f_loss
else:
loss = cfg_train.weight_main * F.mse_loss(render_result['rgb_marched'], target)
psnr = utils.mse2psnr(loss.detach())
#if args.distill_active:
# loss += cfg_train.weight_main * 0.001 * F.mse_loss(render_result['f_marched'], f_target)
if cfg_train.weight_entropy_last > 0:
pout = render_result['alphainv_last'].clamp(1e-6, 1-1e-6)
entropy_last_loss = -(pout*torch.log(pout) + (1-pout)*torch.log(1-pout)).mean()
loss += cfg_train.weight_entropy_last * entropy_last_loss
if cfg_train.weight_nearclip > 0:
near_thres = data_dict['near_clip'] / model.scene_radius[0].item()
near_mask = (render_result['t'] < near_thres)
density = render_result['raw_density'][near_mask]
if len(density):
nearclip_loss = (density - density.detach()).sum()
loss += cfg_train.weight_nearclip * nearclip_loss
if cfg_train.weight_distortion > 0:
n_max = render_result['n_max']
s = render_result['s']
w = render_result['weights']
ray_id = render_result['ray_id']
loss_distortion = flatten_eff_distloss(w, s, 1/n_max, ray_id)
loss += cfg_train.weight_distortion * loss_distortion
if cfg_train.weight_rgbper > 0:
rgbper = (render_result['raw_rgb'] - target[render_result['ray_id']]).pow(2).sum(-1)
rgbper_loss = (rgbper * render_result['weights'].detach()).sum() / len(rays_o)
loss += cfg_train.weight_rgbper * rgbper_loss
loss.backward()
if global_step<cfg_train.tv_before and global_step>cfg_train.tv_after and global_step%cfg_train.tv_every==0:
if not args.freeze_density:
if cfg_train.weight_tv_density>0:
model.density_total_variation_add_grad(
cfg_train.weight_tv_density/len(rays_o), global_step<cfg_train.tv_dense_before)
if not args.freeze_rgb:
if cfg_train.weight_tv_k0>0:
model.k0_total_variation_add_grad(
cfg_train.weight_tv_k0/len(rays_o), global_step<cfg_train.tv_dense_before)
if args.distill_active and (args.only_distill_loss or args.weighted_distill_loss):
if cfg_train.weight_tv_k0>0:
model.f_k0_total_variation_add_grad(
cfg_train.weight_tv_k0/len(rays_o), global_step<cfg_train.tv_dense_before)
optimizer.step()
psnr_lst.append(psnr.item())
# update lr
decay_steps = cfg_train.lrate_decay * 1000
decay_factor = 0.1 ** (1/decay_steps)
for i_opt_g, param_group in enumerate(optimizer.param_groups):
param_group['lr'] = param_group['lr'] * decay_factor
# check log & save
if global_step%args.i_print==0:
eps_time = time.time() - time0
eps_time_str = f'{eps_time//3600:02.0f}:{eps_time//60%60:02.0f}:{eps_time%60:02.0f}'
tqdm.write(f'scene_rep_reconstruction ({stage}): iter {global_step:6d} / '
f'Loss: {loss.item():.9f} / PSNR: {np.mean(psnr_lst):5.2f} / '
f'FeatPSNR: {utils.mse2psnr(f_loss.detach()) if args.distill_active else 0:5.2f} / '
f'Eps: {eps_time_str}')
psnr_lst = []
if global_step%args.i_weights==0:
path = os.path.join(cfg.basedir, cfg.expname, f'{stage}_{global_step:06d}.tar')
torch.save({
'global_step': global_step,
'model_kwargs': model.get_kwargs(),
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
}, path)
torch.save({
'global_step': global_step,
'model_kwargs': model.get_kwargs(),
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
}, last_ckpt_path)
print(f'scene_rep_reconstruction ({stage}): saved checkpoints at', path)
print(f'scene_rep_reconstruction ({stage}): saved checkpoints at', last_ckpt_path)
if global_step != -1:
torch.save({
'global_step': global_step,
'model_kwargs': model.get_kwargs(),
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
}, last_ckpt_path)
print(f'scene_rep_reconstruction ({stage}): saved checkpoints at', last_ckpt_path)
def train(args, cfg, data_dict):
# init
print('train: start')
eps_time = time.time()
os.makedirs(os.path.join(cfg.basedir, cfg.expname), exist_ok=True)
with open(os.path.join(cfg.basedir, cfg.expname, 'args.txt'), 'w') as file:
for arg in sorted(vars(args)):
attr = getattr(args, arg)
file.write('{} = {}\n'.format(arg, attr))
cfg.dump(os.path.join(cfg.basedir, cfg.expname, 'config.py'))
# coarse geometry searching (only works for inward bounded scenes)
eps_coarse = time.time()
xyz_min_coarse, xyz_max_coarse = compute_bbox_by_cam_frustrm(args=args, cfg=cfg, **data_dict)
if cfg.coarse_train.N_iters > 0:
scene_rep_reconstruction(
args=args, cfg=cfg,
cfg_model=cfg.coarse_model_and_render, cfg_train=cfg.coarse_train,
xyz_min=xyz_min_coarse, xyz_max=xyz_max_coarse,
data_dict=data_dict, stage='coarse')
eps_coarse = time.time() - eps_coarse
eps_time_str = f'{eps_coarse//3600:02.0f}:{eps_coarse//60%60:02.0f}:{eps_coarse%60:02.0f}'
print('train: coarse geometry searching in', eps_time_str)
coarse_ckpt_path = os.path.join(cfg.basedir, cfg.expname, f'coarse_last.tar')
else:
print('train: skip coarse geometry searching')
coarse_ckpt_path = None
# fine detail reconstruction
eps_fine = time.time()
if cfg.coarse_train.N_iters == 0:
xyz_min_fine, xyz_max_fine = xyz_min_coarse.clone(), xyz_max_coarse.clone()
else:
xyz_min_fine, xyz_max_fine = compute_bbox_by_coarse_geo(
model_class=dvgo.DirectVoxGO, model_path=coarse_ckpt_path,
thres=cfg.fine_model_and_render.bbox_thres)
scene_rep_reconstruction(
args=args, cfg=cfg,
cfg_model=cfg.fine_model_and_render, cfg_train=cfg.fine_train,
xyz_min=xyz_min_fine, xyz_max=xyz_max_fine,
data_dict=data_dict, stage='fine',
coarse_ckpt_path=coarse_ckpt_path)
eps_fine = time.time() - eps_fine
eps_time_str = f'{eps_fine//3600:02.0f}:{eps_fine//60%60:02.0f}:{eps_fine%60:02.0f}'
print('train: fine detail reconstruction in', eps_time_str)
eps_time = time.time() - eps_time
eps_time_str = f'{eps_time//3600:02.0f}:{eps_time//60%60:02.0f}:{eps_time%60:02.0f}'
print('train: finish (eps time', eps_time_str, ')')
if __name__=='__main__':
# load setup
parser = config_parser()
args = parser.parse_args()
cfg = mmcv.Config.fromfile(args.config)
# init enviroment
if torch.cuda.is_available():
torch.set_default_tensor_type('torch.cuda.FloatTensor')
device = torch.device('cuda')
else:
device = torch.device('cpu')
seed_everything(args)
# load images / poses / camera settings / data split
data_dict = load_everything(args=args, cfg=cfg)
# export scene bbox and camera poses in 3d for debugging and visualization
if args.export_bbox_and_cams_only:
print('Export bbox and cameras...')
xyz_min, xyz_max = compute_bbox_by_cam_frustrm(args=args, cfg=cfg, **data_dict)
poses, HW, Ks, i_train = data_dict['poses'], data_dict['HW'], data_dict['Ks'], data_dict['i_train']
near, far = data_dict['near'], data_dict['far']
if data_dict['near_clip'] is not None:
near = data_dict['near_clip']
cam_lst = []
for c2w, (H, W), K in zip(poses[i_train], HW[i_train], Ks[i_train]):
rays_o, rays_d, viewdirs = dvgo.get_rays_of_a_view(
H, W, K, c2w, cfg.data.ndc, inverse_y=cfg.data.inverse_y,
flip_x=cfg.data.flip_x, flip_y=cfg.data.flip_y,)
cam_o = rays_o[0,0].cpu().numpy()
cam_d = rays_d[[0,0,-1,-1],[0,-1,0,-1]].cpu().numpy()
cam_lst.append(np.array([cam_o, *(cam_o+cam_d*max(near, far*0.05))]))
np.savez_compressed(args.export_bbox_and_cams_only,
xyz_min=xyz_min.cpu().numpy(), xyz_max=xyz_max.cpu().numpy(),
cam_lst=np.array(cam_lst))
print('done')
sys.exit()
if args.export_coarse_only:
print('Export coarse visualization...')
with torch.no_grad():
ckpt_path = os.path.join(cfg.basedir, cfg.expname, 'coarse_last.tar')
model = utils.load_model(dvgo.DirectVoxGO, ckpt_path).to(device)
alpha = model.activate_density(model.density.get_dense_grid()).squeeze().cpu().numpy()
rgb = torch.sigmoid(model.k0.get_dense_grid()).squeeze().permute(1,2,3,0).cpu().numpy()
np.savez_compressed(args.export_coarse_only, alpha=alpha, rgb=rgb)
print('done')
sys.exit()
# train
if not args.render_only:
train(args, cfg, data_dict)
# load model for rendring
if args.render_test or args.render_train or args.render_video or args.render_segment:
if args.ft_path:
ckpt_path = args.ft_path
else:
ckpt_path = os.path.join(cfg.basedir, cfg.expname, 'fine_last.tar')
ckpt_name = ckpt_path.split('/')[-1][:-4]
if cfg.data.ndc:
#model_class = dmpigo.DirectMPIGO
model_class = dvgo.DirectVoxGO
elif cfg.data.unbounded_inward:
model_class = dcvgo.DirectContractedVoxGO
else:
model_class = dvgo.DirectVoxGO
model, optimizer, start = load_existed_model(args, cfg, cfg.fine_train, ckpt_path, device)
#model = utils.load_model(model_class, ckpt_path).to(device)
stepsize = cfg.fine_model_and_render.stepsize
render_viewpoints_kwargs = {
'model': model,
'ndc': cfg.data.ndc,
'render_kwargs': {
'near': data_dict['near'],
'far': data_dict['far'],
'bg': 1 if cfg.data.white_bkgd else 0,
'stepsize': stepsize,
'inverse_y': cfg.data.inverse_y,
'flip_x': cfg.data.flip_x,
'flip_y': cfg.data.flip_y,
'render_depth': True,
},
}
# render trainset and eval
if args.segment:
with torch.no_grad():
render_viewpoints_kwargs_bkp = deepcopy(render_viewpoints_kwargs)
model = model.cpu()
fv = torch.load(args.fv)
faiss_kmeans = kmeans(fv)
print("Reconstructing feature grid.")
start_time = time.time()
model.f_k0 = model.f_k0.cuda()
# maintain two versions of feature grid
# one structured for faster querying of kmeans
# one structured for region growing
fg = get_dense_grid_batch_processing(model.f_k0).cpu()
fg_kmeans = fg.clone()
fg = fg.cuda()
xyz = fg_kmeans.shape[2:]
fg_kmeans = fg_kmeans.squeeze(0).permute(1, 2, 3, 0) # x, y, z, 64
fg_kmeans = fg_kmeans.reshape(-1, 64)
fg_kmeans = fg_kmeans.contiguous()
model.f_k0 = model.f_k0.cpu()
# fg = model.f_k0.get_dense_grid().cpu() # 1, 64, x, y, z
print("Reconstructing feature grid.", time.time() - start_time)
torch.cuda.empty_cache()
if args.segment or args.render_test or args.render_train or args.render_video:
for it in range(1 + int(args.segment) * 1000):
if args.segment:
thresh = float(input("Enter new threshold value:"))
render_viewpoints_kwargs = deepcopy(render_viewpoints_kwargs_bkp)
mask = query_kmeans(faiss_kmeans, fg_kmeans, thresh, xyz)
torch.cuda.empty_cache()
# mask, fg = hybrid_kmeans(model, thresh, fv)
#mask, fg = hybrid_nnfm(model, thresh, fv)
#mask, fg = hybrid_average(model, thresh, fv)
bf = bool(int(input("Do you want to apply bilateral filter? ")))
if bf:
sigma_d = float(input("Enter new sigma_d value:"))
sigma_f = float(input("Enter new sigma_f value:"))
number = int(input("How many times do you want to apply bilateral filter?: "))
for _ in range(number):
# mask = bilateral_filter_mask(mask, fg, sigma_d, sigma_f)
# mask = region_grower_mask(mask, fg, sigma_d, sigma_f)
mask = dev_region_grower_mask(mask, fg, sigma_d, sigma_f)
torch.cuda.empty_cache()
#mask = 1.0 - mask
# del fg
render_viewpoints_kwargs['model'].density.grid = torch.nn.Parameter(render_viewpoints_kwargs['model'].density.grid.cpu() * mask.cpu())
if hasattr(render_viewpoints_kwargs['model'], 'segmentation_mask'):
render_viewpoints_kwargs['model'].segmentation_mask = torch.nn.Parameter(mask.cpu(), requires_grad=False)
del mask
#render_viewpoints_kwargs['model'] = hybrid_nnfm(model, thresh, fv).cuda()
render_viewpoints_kwargs['model'] = render_viewpoints_kwargs['model'].cuda()
if args.render_train:
#breakpoint()
testsavedir = os.path.join(cfg.basedir, cfg.expname, f'render_train_{ckpt_name}')
os.makedirs(testsavedir, exist_ok=True)
print('All results are dumped into', testsavedir)
rgbs, depths, bgmaps, _ = render_viewpoints(
render_poses=data_dict['poses'][data_dict['i_train']],
HW=data_dict['HW'][data_dict['i_train']],
Ks=data_dict['Ks'][data_dict['i_train']],
gt_imgs=[data_dict['images'][i].cpu().numpy() for i in data_dict['i_train']],
cfg=cfg,savedir=testsavedir, dump_images=args.dump_images,
eval_ssim=args.eval_ssim, eval_lpips_alex=args.eval_lpips_alex, eval_lpips_vgg=args.eval_lpips_vgg,
distill_active=args.distill_active,
**render_viewpoints_kwargs)
imageio.mimwrite(os.path.join(testsavedir, 'video.rgb.mp4'), utils.to8b(rgbs), fps=30, quality=8)
imageio.mimwrite(os.path.join(testsavedir, 'video.depth.mp4'), utils.to8b(1 - depths / np.max(depths)), fps=30, quality=8)
# render segment
if args.render_segment:
testsavedir = os.path.join(cfg.basedir, cfg.expname, f'render_segment_{ckpt_name}')
os.makedirs(testsavedir, exist_ok=True)
print('All results are dumped into', testsavedir)
stepsize = (len(data_dict['i_test']) // 3) + 1
rgbs, depths, bgmaps, _ = render_viewpoints(
render_poses=data_dict['poses'][data_dict['i_test'][::stepsize]],
HW=data_dict['HW'][data_dict['i_test'][::stepsize]],
Ks=data_dict['Ks'][data_dict['i_test'][::stepsize]],
cfg=cfg, gt_imgs=[data_dict['images'][i].cpu().numpy() for i in data_dict['i_test'][::3]],
savedir=testsavedir, dump_images=args.dump_images,
eval_ssim=args.eval_ssim, eval_lpips_alex=args.eval_lpips_alex, eval_lpips_vgg=args.eval_lpips_vgg,
distill_active=args.distill_active,
**render_viewpoints_kwargs)
imageio.mimwrite(os.path.join(testsavedir, 'video.rgb.mp4'), utils.to8b(rgbs), fps=30, quality=8)
imageio.mimwrite(os.path.join(testsavedir, 'video.depth.mp4'), utils.to8b(1 - depths / np.max(depths)), fps=30, quality=8)
# render testset and eval
if args.render_test:
testsavedir = os.path.join(cfg.basedir, cfg.expname, f'render_test_{ckpt_name}')
os.makedirs(testsavedir, exist_ok=True)
print('All results are dumped into', testsavedir)
rgbs, depths, bgmaps, _ = render_viewpoints(
render_poses=data_dict['poses'][data_dict['i_test']],
HW=data_dict['HW'][data_dict['i_test']],
Ks=data_dict['Ks'][data_dict['i_test']],
cfg=cfg, gt_imgs=[data_dict['images'][i].cpu().numpy() for i in data_dict['i_test']],
savedir=testsavedir, dump_images=args.dump_images,
eval_ssim=args.eval_ssim, eval_lpips_alex=args.eval_lpips_alex, eval_lpips_vgg=args.eval_lpips_vgg,
distill_active=args.distill_active,
**render_viewpoints_kwargs)
imageio.mimwrite(os.path.join(testsavedir, 'video.rgb.mp4'), utils.to8b(rgbs), fps=30, quality=8)
imageio.mimwrite(os.path.join(testsavedir, 'video.depth.mp4'), utils.to8b(1 - depths / np.max(depths)), fps=30, quality=8)
# render video
if args.render_video:
testsavedir = os.path.join(cfg.basedir, cfg.expname, f'render_video_{ckpt_name}')
os.makedirs(testsavedir, exist_ok=True)
print('All results are dumped into', testsavedir)
rgbs, depths, bgmaps, _ = render_viewpoints(
render_poses=data_dict['render_poses'],
HW=data_dict['HW'][data_dict['i_test']][[0]].repeat(len(data_dict['render_poses']), 0),
Ks=data_dict['Ks'][data_dict['i_test']][[0]].repeat(len(data_dict['render_poses']), 0),
cfg=cfg,
render_factor=args.render_video_factor,
render_video_flipy=args.render_video_flipy,
render_video_rot90=args.render_video_rot90,
savedir=testsavedir, dump_images=args.dump_images,
**render_viewpoints_kwargs)
imageio.mimwrite(os.path.join(testsavedir, 'video.rgb.mp4'), utils.to8b(rgbs), fps=30, quality=8)
import matplotlib.pyplot as plt
depths_vis = depths * (1-bgmaps) + bgmaps
dmin, dmax = np.percentile(depths_vis[bgmaps < 0.1], q=[5, 95])
depth_vis = plt.get_cmap('rainbow')(1 - np.clip((depths_vis - dmin) / (dmax - dmin), 0, 1)).squeeze()[..., :3]
imageio.mimwrite(os.path.join(testsavedir, 'video.depth.mp4'), utils.to8b(depth_vis), fps=30, quality=8)
save_path = str(input("Do you want to save the checkpoint? "))
if save_path:
torch.save({
'global_step': start,
'model_kwargs': render_viewpoints_kwargs['model'].get_kwargs(),
'model_state_dict': render_viewpoints_kwargs['model'].state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
}, os.path.join(cfg.basedir, cfg.expname, save_path))
print('Done')
@torch.no_grad()
def do_setup(args):
cfg = mmcv.Config.fromfile(args.config)
# init enviroment
if torch.cuda.is_available():
torch.set_default_tensor_type('torch.cuda.FloatTensor')
device = torch.device('cuda')
else:
device = torch.device('cpu')
seed_everything(args)
# load images / poses / camera settings / data split
data_dict = load_everything(args=args, cfg=cfg)
return cfg, data_dict, device
@torch.no_grad()
def load_model(args, cfg, data_dict, device):
if args.ft_path:
ckpt_path = args.ft_path
else:
ckpt_path = os.path.join(cfg.basedir, cfg.expname, 'fine_last.tar')
ckpt_name = ckpt_path.split('/')[-1][:-4]
if cfg.data.ndc:
#model_class = dmpigo.DirectMPIGO
model_class = dvgo.DirectVoxGO
elif cfg.data.unbounded_inward:
model_class = dcvgo.DirectContractedVoxGO
else:
model_class = dvgo.DirectVoxGO
model, optimizer, start = load_existed_model(args, cfg, cfg.fine_train, ckpt_path, device)