-
Notifications
You must be signed in to change notification settings - Fork 0
/
exp_runner.py
1606 lines (1249 loc) · 77 KB
/
exp_runner.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 cv2 as cv
import numpy as np
import os, logging, argparse, trimesh, copy
import torch
import torch.nn.functional as F
from torch.utils.tensorboard import SummaryWriter
from models.dataset import Dataset
from models.fields import SDFNetwork, RenderingNetwork, SingleVarianceNetwork, NeRF, Cascaded_SDFNetwork
from models.renderer import NeuSRenderer, extract_fields
from models.nerf_renderer import NeRFRenderer
from models.loss import NeuSLoss
import models.patch_match_cuda as PatchMatch
from shutil import copyfile
from tqdm import tqdm
from icecream import ic
from datetime import datetime
from pyhocon import ConfigFactory, HOCONConverter
import utils.utils_io as IOUtils
import utils.utils_geometry as GeoUtils
import utils.utils_image as ImageUtils
import utils.utils_training as TrainingUtils
import clip
from PIL import Image
from tqdm import tqdm
import open3d as o3d
class Runner:
def __init__(self, conf_path, conf, scene_name = '', mode='train', model_type='', is_continue=False, checkpoint_id = -1):
# Initial setting: Genreal
self.device = torch.device('cuda')
self.conf_path = conf_path
self.conf = conf
self.dataset_type = self.conf['general.dataset_type']
self.scan_name = self.conf['general.scan_name']
if len(scene_name)>0:
self.scan_name = scene_name
print(f"**********************Scan name: {self.scan_name}\n********************** \n\n\n")
self.exp_name = self.conf['general.exp_name']
# parse cmd args
self.is_continue = is_continue
self.checkpoint_id = checkpoint_id
self.mode = mode
self.model_type = self.conf['general.model_type']
if model_type != '': # overwrite
self.model_type = model_type
self.model_list = []
self.writer = None
self.curr_img_idx = -1
# novel view training (nvt)
self.nvt_begin = self.conf['train'].get('nvt_begin', default=-1)
self.nvt_freq = self.conf['train'].get('nvt_freq', default=-1)
self.b_use_nvt = self.nvt_begin >= 0 and self.nvt_freq >= 0
self.parse_parameters()
self.build_dataset()
self.build_model()
if self.is_continue:
self.load_pretrained_model()
# recording
if self.mode[:5] == 'train':
self.file_backup()
with open(f'{self.base_exp_dir}/recording/config_modified.conf', 'w') as configfile:
configfile.write(HOCONConverter.to_hocon(self.conf))
def load_pretrained_model(self):
# Load checkpoint
latest_model_name = None
model_list_raw = os.listdir(os.path.join(self.base_exp_dir, 'checkpoints'))
model_list = []
for model_name in model_list_raw:
if 'pretrained_sdf' in model_name:
continue
if model_name[-3:] == 'pth' and int(model_name[5:-4]) <= self.end_iter:
model_list.append(model_name)
model_list.sort()
if self.checkpoint_id == -1:
latest_model_name = model_list[-1]
else:
if f'ckpt_{self.checkpoint_id:06d}.pth' in model_list:
latest_model_name = f'ckpt_{self.checkpoint_id:06d}.pth'
else:
logging.error(f"path_ckpt is not exist. (f'ckpt_{self.checkpoint_id:06d}.pth')")
exit()
if latest_model_name is not None:
logging.info('Find checkpoint: {}'.format(latest_model_name))
self.load_checkpoint(latest_model_name)
def parse_parameters(self):
# patch-match
self.thres_robust_ncc = self.conf['dataset.patchmatch_thres_ncc_robust']
logging.info(f'Robust ncc threshold: {self.thres_robust_ncc}')
self.patchmatch_start = self.conf['dataset.patchmatch_start']
self.patchmatch_mode = self.conf['dataset.patchmatch_mode']
self.nvs = self.conf.get_bool('train.nvs', default=False)
self.save_normamap_npz = self.conf['train.save_normamap_npz']
self.base_exp_dir = os.path.join(self.conf['general.exp_dir'], self.dataset_type, self.model_type, self.scan_name, str(self.exp_name))
os.makedirs(self.base_exp_dir, exist_ok=True)
logging.info(f'Exp dir: {self.base_exp_dir}')
self.iter_step = 0
self.radius_norm = 1.0
# trainning parameters
self.end_iter = self.conf.get_int('train.end_iter')
self.batch_size = self.conf.get_int('train.batch_size')
self.validate_resolution_level = self.conf.get_int('train.validate_resolution_level')
self.learning_rate = self.conf.get_float('train.learning_rate')
self.learning_rate_milestone = self.conf.get_list('train.learning_rate_milestone')
self.learning_rate_factor = self.conf.get_float('train.learning_rate_factor')
self.warm_up_end = self.conf.get_float('train.warm_up_end', default=0.0)
self.learning_rate_alpha = self.conf.get_float('train.learning_rate_alpha')
logging.info(f'Warm up end: {self.warm_up_end}; learning_rate_alpha: {self.learning_rate_alpha}')
self.use_white_bkgd = self.conf.get_bool('train.use_white_bkgd')
self.anneal_start = self.conf.get_float('train.anneal_start', default=0.0)
self.anneal_end = self.conf.get_float('train.anneal_end', default=0.0)
self.n_samples = self.conf.get_int('model.neus_renderer.n_samples')
# validate parameters
for attr in ['save_freq','report_freq', 'val_image_freq', 'val_mesh_freq', 'val_fields_freq',
'freq_valid_points', 'freq_valid_weights',
'freq_save_confidence' ]:
setattr(self, attr, self.conf.get_int(f'train.{attr}'))
def build_dataset(self):
# dataset and directories
if self.dataset_type == 'indoor' or self.dataset_type == 'indoor-paper':
self.sample_range_indoor = self.conf['dataset']['sample_range_indoor']
# self.conf['dataset']['use_normal'] = True if self.conf['model.loss.normal_weight'] > 0 else False
self.conf['dataset']['use_normal'] = True
self.use_indoor_data = True
logging.info(f"Ray sample range: {self.sample_range_indoor}")
elif self.dataset_type == 'dtu':
self.use_indoor_data = False
self.conf['model.sdf_network']['reverse_geoinit'] = False
self.conf['dataset']['use_normal'] = True if self.conf['model.loss.normal_weight'] > 0 else False
self.scan_id = int(self.scan_name[4:])
logging.info(f"DTU scan ID: {self.scan_id}")
else:
raise NotImplementedError
logging.info(f"Use normal: {self.conf['dataset']['use_normal']}")
self.conf['dataset']['use_planes'] = True if self.conf['model.loss.normal_consistency_weight'] > 0 else False
self.conf['dataset']['use_plane_offset_loss'] = True if self.conf['model.loss.plane_offset_weight'] > 0 else False
self.conf['dataset']['data_dir'] = os.path.join(self.conf['general.data_dir'] , self.dataset_type, self.scan_name)
if self.b_use_nvt:
self.conf['dataset'].update({
'b_use_nvt': True,
'nvt_resolution_level': 3
})
self.dataset = Dataset(self.conf['dataset'])
def build_model(self):
# Networks
params_to_train = []
if self.model_type == 'neus':
self.nerf_outside = NeRF(**self.conf['model.tiny_nerf']).to(self.device)
if 'cascaded_network' in self.conf['model'].keys():
self.sdf_network_fine = Cascaded_SDFNetwork(**self.conf['model.cascaded_network']).to(self.device)
else:
self.sdf_network_fine = SDFNetwork(**self.conf['model.sdf_network']).to(self.device)
self.variance_network_fine = SingleVarianceNetwork(**self.conf['model.variance_network']).to(self.device)
self.color_network_fine = RenderingNetwork(**self.conf['model.rendering_network']).to(self.device)
params_to_train += list(self.nerf_outside.parameters())
params_to_train += list(self.sdf_network_fine.parameters())
params_to_train += list(self.variance_network_fine.parameters())
params_to_train += list(self.color_network_fine.parameters())
self.renderer = NeuSRenderer(self.nerf_outside,
self.sdf_network_fine,
self.variance_network_fine,
self.color_network_fine,
**self.conf['model.neus_renderer'])
elif self.model_type == 'nerf': # NeRF
self.nerf_coarse = NeRF(**self.conf['model.nerf']).to(self.device)
self.nerf_fine = NeRF(**self.conf['model.nerf']).to(self.device)
self.nerf_outside = NeRF(**self.conf['model.tiny_nerf']).to(self.device)
params_to_train += list(self.nerf_coarse.parameters())
params_to_train += list(self.nerf_fine.parameters())
params_to_train += list(self.nerf_outside.parameters())
self.renderer = NeRFRenderer(self.nerf_coarse,
self.nerf_fine,
self.nerf_outside,
**self.conf['model.nerf_renderer'])
else:
NotImplementedError
self.optimizer = torch.optim.Adam(params_to_train, lr=self.learning_rate)
self.loss_neus = NeuSLoss(self.conf['model.loss'])
def train(self):
if self.model_type == 'neus':
self.train_neus()
elif self.model_type == 'nerf':
self.train_nerf()
else:
NotImplementedError
def get_near_far(self, rays_o, rays_d, image_perm = None, iter_i= None, pixels_x = None, pixels_y= None):
log_vox = {}
log_vox.clear()
batch_size = len(rays_o)
if self.dataset_type == 'dtu':
near, far = self.dataset.near_far_from_sphere(rays_o, rays_d)
elif self.dataset_type == 'indoor' or self.dataset_type == 'indoor-paper':
near, far = torch.zeros(batch_size, 1), self.sample_range_indoor * torch.ones(batch_size, 1)
else:
NotImplementedError
logging.debug(f"[{self.iter_step}] Sample range: max, {torch.max(far -near)}; min, {torch.min(far -near)}")
return near, far, log_vox
def get_near_far_nvt(self, rays_o, rays_d, corners_3D_list):
'''
- corners_3D_list: [8,3,1]
'''
# import pdb; pdb.set_trace()
batch_size = len(rays_o)
z_min = torch.min(corners_3D_list[:,2,:]) * 0.9
z_max = torch.max(corners_3D_list[:,2,:]) * 1.1
near = torch.ones(batch_size, 1) * z_min
far = torch.ones(batch_size, 1) * z_max
return near, far
def get_model_input_nvt(self, image_perm, iter_i, resolution_level=3):
input_model = {}
idx_img = image_perm[self.iter_step % self.dataset.n_images]
self.curr_img_idx = idx_img
with torch.no_grad():
rays_o, rays_d, corners_3D_list = self.dataset.gen_rays_at_novel_pose(idx_img, resolution_level=resolution_level)
if rays_o is None:
return None
## render the image at novel view.
H, W, _ = rays_o.shape
rays_o = rays_o.reshape(-1, 3).split(self.batch_size)
rays_d = rays_d.reshape(-1, 3).split(self.batch_size)
imgs_render = {}
imgs_render.update({
'color_fine': [],
'normal': [],
'color_peak': [],
'normal_peak': []
})
with torch.cuda.amp.autocast(enabled=True):
# for rays_o_batch, rays_d_batch in tqdm(zip(rays_o, rays_d)):
for rays_o_batch, rays_d_batch in zip(rays_o, rays_d):
near, far = self.get_near_far_nvt(rays_o_batch, rays_d_batch, corners_3D_list)
background_rgb = torch.ones([1, 3]) if self.use_white_bkgd else None
## change the number of sample points.
original_n_samples = self.renderer.n_samples
original_n_importance = self.renderer.n_importance
self.renderer.n_samples = 32
self.renderer.n_importance = 0
render_out, _ = self.renderer.render(rays_o_batch, rays_d_batch, near, far, alpha_inter_ratio=self.get_alpha_inter_ratio(), background_rgb=background_rgb)
## recover the sampling setting.
self.renderer.n_samples = original_n_samples
self.renderer.n_importance = original_n_importance
# import pdb; pdb.set_trace()
feasible = lambda key: ((key in render_out) and (render_out[key] is not None))
for key in imgs_render:
if feasible(key):
# imgs_render[key].append(render_out[key].detach().cpu().numpy())
imgs_render[key].append(render_out[key])
# import pdb; pdb.set_trace()
# (2) reshape rendered images
for key in imgs_render:
if len(imgs_render[key]) > 0:
# imgs_render[key] = np.concatenate(imgs_render[key], axis=0)
imgs_render[key] = torch.cat(imgs_render[key], dim=0)
if imgs_render[key].shape[1] == 3: # for color and normal
# imgs_render[key] = imgs_render[key].reshape([H, W, 3])
imgs_render[key] = torch.reshape(imgs_render[key], (H, W, 3))
elif imgs_render[key].shape[1] == 1:
# imgs_render[key] = imgs_render[key].reshape([H, W])
imgs_render[key] = torch.reshape(imgs_render[key], (H, W))
# (3) normalize rendered image
for key in imgs_render:
img_temp = imgs_render[key]
if key in ['normal', 'normal_peak']:
img_temp = (((img_temp + 1) * 0.5).clip(0,1) * 255)
if key in ['color_fine', 'color_peak']:
img_temp *= 255
imgs_render[key] = img_temp
# ## save imgs_render['normal_peak'] and imgs_render['color_peak'] to disk
# # import pdb; pdb.set_trace()
# cv.imwrite(f'./normal_peak_{iter_i}_{idx_img}.png', imgs_render['normal_peak'].cpu().numpy().astype(np.uint8))
# cv.imwrite(f'./color_peak_{iter_i}_{idx_img}.png', imgs_render['color_peak'].cpu().numpy().astype(np.uint8))
if imgs_render['color_fine'][...,1].max() > 0.2:
mask = imgs_render['color_fine'][:,:,1] > 0.2
if mask.sum() < 200:
return None
else:
return None
# import pdb; pdb.set_trace()
novel_normal_map_feature = self.encode_img_with_clip(imgs_render['normal'])
novel_color_img_feature = self.encode_img_with_clip(imgs_render['color_fine'])
input_model.update({
'reference_color_feature': self.dataset.reference_color_img_feature,
'reference_normal_feature': self.dataset.reference_normal_map_feature,
'novel_color_feature': novel_color_img_feature,
'novel_normal_feature': novel_normal_map_feature,
'text_feature': self.dataset.text_feature,
})
return input_model
def encode_img_with_clip(self, img):
# import pdb; pdb.set_trace()
if img[...,1].max() > 0.2:
mask = img[:,:,1] > 0.2
## [min_x, min_y, max_x, max_y]
try:
img_bbox = torch.tensor([torch.min(torch.where(mask)[1]), torch.min(torch.where(mask)[0]), torch.max(torch.where(mask)[1]), torch.max(torch.where(mask)[0])]).cpu().numpy()
except:
import pdb; pdb.set_trace()
else:
img_bbox = np.array([0, 0, img.shape[1], img.shape[0]])
if img_bbox[0] == img_bbox[2] or img_bbox[1] == img_bbox[3]:
img_bbox = np.array([0, 0, img.shape[1], img.shape[0]])
square_img_bbox = self.dataset.get_square_bbox(img_bbox, (img.shape[1], img.shape[0]), scale = 1.2)
cropped_img = img[square_img_bbox[1]:square_img_bbox[3], square_img_bbox[0]:square_img_bbox[2], :]
preprocessed_img = self.dataset.clip_preprocess_torch_api(cropped_img)
clip_feature = self.dataset.clip_model.encode_image(preprocessed_img)
return clip_feature
def get_model_input(self, image_perm, iter_i):
input_model = {}
idx_img = image_perm[self.iter_step % self.dataset.n_images]
self.curr_img_idx = idx_img
data, pixels_x, pixels_y, normal_sample, planes_sample, subplanes_sample, normal_mask, depth_sample = self.dataset.random_get_rays_at(idx_img, self.batch_size)
rays_o, rays_d, true_rgb, true_mask = data[:, :3], data[:, 3: 6], data[:, 6: 9], data[:, 9: 10]
true_mask = (true_mask > 0.5).float()
mask = true_mask
if self.conf['model.loss.normal_weight'] > 0:
input_model['normals_gt'] = normal_sample
input_model['normal_mask'] = normal_mask
if self.conf['model.loss.depth_weight'] > 0:
input_model['depth_sample'] = depth_sample
else:
input_model['depth_sample'] = depth_sample
if self.conf['model.loss.normal_consistency_weight'] > 0:
input_model['planes_gt'] = planes_sample
if self.conf['model.loss.plane_offset_weight'] > 0:
input_model['subplanes_gt'] = subplanes_sample
near, far, logs_input = self.get_near_far(rays_o, rays_d, image_perm, iter_i, pixels_x, pixels_y)
background_rgb = None
if self.use_white_bkgd:
background_rgb = torch.ones([1, 3])
if self.conf['model.loss.mask_weight'] > 0.0:
mask = (mask > 0.5).float()
else:
mask = torch.ones_like(mask)
mask_sum = mask.sum() + 1e-5
pixels_uv = torch.stack([pixels_x, pixels_y], dim=-1)
pixels_vu = torch.stack([pixels_y, pixels_x], dim=-1)
input_model.update({
'rays_o': rays_o,
'rays_d': rays_d,
'near': near,
'far': far,
'mask': mask,
'mask_sum': mask_sum,
'true_rgb': true_rgb,
'background_rgb': background_rgb,
'pixels_x': pixels_x, # u
'pixels_y': pixels_y, # v,
'pixels_uv': pixels_uv,
'pixels_vu': pixels_vu
})
return input_model, logs_input
def train_neus(self):
self.writer = SummaryWriter(log_dir=os.path.join(self.base_exp_dir, 'logs'))
self.update_learning_rate()
self.update_iter_step()
res_step = self.end_iter - self.iter_step
if self.dataset.cache_all_data:
self.dataset.shuffle()
# self.validate_mesh() # save mesh at iter 0
logs_summary = {}
image_perm = torch.randperm(self.dataset.n_images)
for iter_i in tqdm(range(res_step)):
logs_summary.clear()
## Normal training.
input_model, logs_input = self.get_model_input(image_perm, iter_i)
logs_summary.update(logs_input)
render_out, logs_render = self.renderer.render(input_model['rays_o'], input_model['rays_d'],
input_model['near'], input_model['far'],
background_rgb=input_model['background_rgb'],
alpha_inter_ratio=self.get_alpha_inter_ratio())
logs_summary.update(logs_render)
patchmatch_out, logs_patchmatch = self.patch_match(input_model, render_out)
logs_summary.update(logs_patchmatch)
# if self.b_use_nvt and self.iter_step > self.nvt_begin:
# ## mask weight = 0.0
# self.loss_neus.mask_weight = 0.1
loss, logs_loss, mask_keep_gt_normal = self.loss_neus(input_model, render_out, self.sdf_network_fine, patchmatch_out)
logs_summary.update(logs_loss)
self.optimizer.zero_grad()
# import pdb; pdb.set_trace()
loss.backward()
self.optimizer.step()
self.iter_step += 1
logs_val = self.validate(input_model, logs_loss, render_out)
logs_summary.update(logs_val)
logs_summary.update({'Log/lr': self.optimizer.param_groups[0]['lr']})
self.write_summary(logs_summary)
self.update_learning_rate()
self.update_iter_step()
self.accumulate_rendered_results(input_model, render_out, patchmatch_out,
b_accum_render_difference = False,
b_accum_ncc = False,
b_accum_normal_pts = False)
if self.b_use_nvt and self.iter_step > self.nvt_begin and self.iter_step % self.nvt_freq == 0:
# import pdb; pdb.set_trace()
input_model = self.get_model_input_nvt(image_perm, iter_i, resolution_level=4)
if input_model is None:
continue
normal_feature_loss = 1 - F.cosine_similarity(input_model['novel_normal_feature'], input_model['reference_normal_feature'], dim=1)
color_feature_loss = 1 - F.cosine_similarity(input_model['novel_color_feature'], input_model['reference_color_feature'], dim=1)
text_feature_loss = 2 - F.cosine_similarity(input_model['novel_color_feature'], input_model['text_feature'], dim=1) - F.cosine_similarity(input_model['novel_normal_feature'], input_model['text_feature'], dim=1)
loss = color_feature_loss + normal_feature_loss + text_feature_loss
logs_summary.update({
'color_feature_loss': color_feature_loss.item(),
'normal_feature_loss': normal_feature_loss.item(),
'text_feature_loss': text_feature_loss.item(),
})
self.optimizer.zero_grad()
# import pdb; pdb.set_trace()
loss.backward()
self.optimizer.step()
self.iter_step += 1
logs_summary.update({'Log/lr': self.optimizer.param_groups[0]['lr']})
self.write_summary(logs_summary)
self.update_learning_rate()
self.update_iter_step()
if self.iter_step % self.dataset.n_images == 0:
image_perm = torch.randperm(self.dataset.n_images)
if self.iter_step % 1000 == 0:
torch.cuda.empty_cache()
logging.info(f'Done. [{self.base_exp_dir}]')
def calculate_ncc_samples(self, input_model, render_out,
use_peak_value = True,
use_normal_prior = False):
pixels_coords_vu =input_model['pixels_vu']
if use_peak_value:
pts_render = render_out['point_peak']
normals_render = render_out['normal_peak']
else:
raise NotImplementedError
if use_normal_prior:
normals_render = input_model['normals_gt']
pts_all = pts_render[:, None, :]
normals_all = normals_render[:, None, :]
coords_all = pixels_coords_vu[:, None, :]
with torch.no_grad():
scores_ncc_all, diff_patch_all, mask_valid_all = self.dataset.score_pixels_ncc(self.curr_img_idx, pts_all.reshape(-1, 3),
normals_all.reshape(-1, 3),
coords_all.reshape(-1, 2))
num_valid = mask_valid_all.sum()
scores_ncc_all = scores_ncc_all.reshape(self.batch_size, -1)
mask_valid_all = mask_valid_all.reshape(self.batch_size, -1)
return scores_ncc_all, diff_patch_all, mask_valid_all
def patch_match(self, input_model, render_out):
patchmatch_out, logs_summary = None, {}
if self.iter_step > self.patchmatch_start:
# ensure initialization of confidence_map, depth_map and points_map
if self.dataset.confidence_accum is None:
self.initialize_accumulated_results(mode=self.conf['dataset.mode_init_accum'],
iter_step_npz=self.conf['dataset.init_accum_step_npz'],
resolution_level=self.conf['dataset.init_accum_reso_level'])
if self.patchmatch_mode == 'use_geocheck':
scores_ncc, diffs_patch, mask_valid = self.calculate_ncc_samples(input_model, render_out)
# (1) mask of pixels with high confidence
mask_high_confidence_curr = scores_ncc < self.thres_robust_ncc
# (2) check previous ncc confidence
pixels_u, pixels_v = input_model['pixels_x'].cpu().numpy(), input_model['pixels_y'].cpu().numpy()
ncc_prev = self.dataset.confidence_accum[self.curr_img_idx][pixels_v, pixels_u]
mask_high_confidence_prev = ncc_prev < self.thres_robust_ncc
# (3) remove normals with large difference between rendered normal and gt normal
# large difference means large inconsistency of normal between differenct views
normal_render = render_out['normal_peak']
angles_diff = TrainingUtils.get_angles(normal_render, input_model['normals_gt'])
MAX_ANGLE_DIFF = 30.0 / 180.0 * 3.1415926
mask_small_normal_diffence = angles_diff < MAX_ANGLE_DIFF
# (4) merge masks
mask_use_gt = mask_high_confidence_curr & torch.from_numpy(mask_high_confidence_prev[:,None]).cuda() & mask_small_normal_diffence
# (5) update confidence, discard the normals
mask_not_use_gt = (mask_use_gt==False)
scores_ncc_all2 = copy.deepcopy(scores_ncc)
scores_ncc_all2[mask_not_use_gt] = 1.0
self.dataset.confidence_accum[self.curr_img_idx][pixels_v, pixels_u] = scores_ncc_all2.squeeze().cpu().numpy()
idx_scores_min = mask_use_gt.long() # 1 use gt, 0 not
patchmatch_out = {
'patchmatch_mode': self.patchmatch_mode,
'scores_ncc_all': scores_ncc,
'idx_scores_min': idx_scores_min,
}
logs_summary.update({
'Log/pixels_use_volume_rendering_only': (idx_scores_min==0).sum(),
'Log/pixels_use_prior': (idx_scores_min==1).sum(),
})
else:
raise NotImplementedError
return patchmatch_out, logs_summary
def accumulate_rendered_results(self, input_model, render_out, patchmatch_out,
b_accum_render_difference = False,
b_accum_ncc = False,
b_accum_normal_pts = False):
'''Cache rendererd depth, normal and confidence (if avaliable) in the training process
Args:
scores: (N,)
'''
pixels_u, pixels_v = input_model['pixels_x'], input_model['pixels_y']
pixels_u_np, pixels_v_np = pixels_u.cpu().numpy(), pixels_v.cpu().numpy()
if b_accum_render_difference:
color_gt = self.dataset.images_denoise_np[self.curr_img_idx.item()][pixels_v_np, pixels_u_np]
color_render = render_out['color_fine'].detach().cpu.numpy()
diff = np.abs(color_render - color_gt).sum(axis=-1)
self.dataset.render_difference_accum[self.curr_img_idx.item()][pixels_v_np, pixels_u_np] = diff
if b_accum_ncc:
scores_vr = patchmatch_out['scores_samples']
self.dataset.confidence_accum[self.curr_img_idx.item()][pixels_v_np, pixels_u_np] = scores_vr.cpu().numpy()
if b_accum_normal_pts:
point_peak = render_out['point_peak']
self.dataset.points_accum[self.curr_img_idx][pixels_v, pixels_u] = point_peak.detach().cpu()
normal_peak = render_out['normal_peak']
self.dataset.normals_accum[self.curr_img_idx][pixels_v, pixels_u] = normal_peak.detach().cpu()
b_accum_all_data = False
if b_accum_all_data:
self.dataset.depths_accum[self.curr_img_idx][pixels_v, pixels_u] = render_out['depth_peak'].squeeze()
self.dataset.colors_accum[self.curr_img_idx][pixels_v, pixels_u] = render_out['color_peak']
def write_summary(self, logs_summary):
for key in logs_summary:
self.writer.add_scalar(key, logs_summary[key], self.iter_step )
def update_iter_step(self):
self.sdf_network_fine.iter_step = self.iter_step
self.sdf_network_fine.end_iter = self.end_iter
self.loss_neus.iter_step = self.iter_step
self.loss_neus.iter_end = self.end_iter
self.dataset.iter_step = self.iter_step
def get_alpha_inter_ratio(self):
if self.anneal_end == 0.0:
return 1.0
elif self.iter_step < self.anneal_start:
return 0.0
else:
return np.min([1.0, (self.iter_step - self.anneal_start) / (self.anneal_end - self.anneal_start)])
def get_alpha_inter_ratio_decrease(self):
anneal_end = 5e4
anneal_start = 0
if anneal_end == 0.0:
return 0.0
elif self.iter_step < anneal_start:
return 1.0
else:
return 1.0 - np.min([1.0, (self.iter_step - anneal_start) / (anneal_end - anneal_start)])
def update_learning_rate(self):
if self.iter_step < self.warm_up_end:
learning_factor = self.iter_step / self.warm_up_end
else:
alpha = self.learning_rate_alpha
progress = (self.iter_step - self.warm_up_end) / (self.end_iter - self.warm_up_end)
learning_factor = (np.cos(np.pi * progress) + 1.0) * 0.5 * (1 - alpha) + alpha
for g in self.optimizer.param_groups:
g['lr'] = self.learning_rate * learning_factor
def train_nerf(self):
self.writer = SummaryWriter(log_dir=os.path.join(self.base_exp_dir, 'logs'))
self.update_learning_rate()
res_step = self.end_iter - self.iter_step
image_perm = self.get_image_perm()
for iter_i in tqdm(range(res_step)):
data, _, _, normal_sample, planes_sample, subplanes_sample, mask_normal_certain_sample, depth_sample = self.dataset.random_get_rays_at(image_perm[self.iter_step % len(image_perm)], self.batch_size)
rays_o, rays_d, true_rgb, mask = data[:, :3], data[:, 3: 6], data[:, 6: 9], data[:, 9: 10]
if self.dataset_type == 'dtu':
near, far = self.dataset.near_far_from_sphere(rays_o, rays_d)
elif self.dataset_type == 'indoor' or self.dataset_type == 'indoor-paper':
near, far = torch.zeros(len(rays_o), 1), self.sample_range_indoor * torch.ones(len(rays_o), 1)
else:
NotImplementedError
background_rgb = None
if self.use_white_bkgd:
background_rgb = torch.ones([1, 3])
if self.conf['model.loss.mask_weight'] > 0.0:
mask = (mask > 0.5).float()
else:
mask = torch.ones_like(mask)
mask_sum = mask.sum() + 1e-5
render_out = self.renderer.render(rays_o, rays_d, near, far,
background_rgb=background_rgb)
color_coarse = render_out['color_coarse']
color_fine = render_out['color_fine']
weight_sum = render_out['weight_sum']
# Color loss
color_coarse_error = (color_coarse - true_rgb) * mask
color_coarse_loss = F.mse_loss(color_coarse_error, torch.zeros_like(color_coarse_error), reduction='sum') / mask_sum
color_fine_error = (color_fine - true_rgb) * mask
color_fine_loss = F.mse_loss(color_fine_error, torch.zeros_like(color_fine_error), reduction='sum') / mask_sum
psnr = 20.0 * torch.log10(1.0 / (((color_fine - true_rgb)**2 * mask).sum() / (mask_sum * 3.0)).sqrt())
# Mask loss, optional
mask_loss = F.binary_cross_entropy(weight_sum.clip(1e-3, 1.0 - 1e-3), mask)
loss = color_coarse_loss + color_fine_loss
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
self.iter_step += 1
self.writer.add_scalar('Loss/loss', loss, self.iter_step)
self.writer.add_scalar('Loss/color_loss', color_fine_loss, self.iter_step)
self.writer.add_scalar('Loss/color_coarse_loss', color_coarse_loss, self.iter_step)
self.writer.add_scalar('Log/psnr', psnr, self.iter_step)
self.writer.add_scalar('Log/lr', self.optimizer.param_groups[0]['lr'], self.iter_step)
if self.iter_step % self.report_freq == 0:
print(self.base_exp_dir)
logging.info('iter:{:8>d} loss = {} lr={}'.format(self.iter_step, loss, self.optimizer.param_groups[0]['lr']))
if self.iter_step % self.save_freq == 0:
self.save_checkpoint()
if self.iter_step % self.val_image_freq == 0:
self.validate_image_nerf()
if self.iter_step % self.val_mesh_freq == 0:
self.validate_mesh_nerf()
# if self.iter_step % self.freq_eval_mesh == 0:
# self.validate_mesh_nerf(world_space=False, resolution=512, threshold=25)
self.update_learning_rate()
self.dataset.iter_step = self.iter_step
if self.iter_step % len(image_perm) == 0:
image_perm = self.get_image_perm()
def get_image_perm(self):
if not self.nvs:
return torch.randperm(self.dataset.n_images)
else:
lis = [i for i in range(self.dataset.n_images) if not i in [8, 13, 16, 21, 26, 31, 34, 56]]
lis = torch.tensor(lis, dtype=torch.long)
return lis[torch.randperm(len(lis))]
def file_backup(self):
# copy python file
dir_lis = self.conf['general.recording']
os.makedirs(os.path.join(self.base_exp_dir, 'recording'), exist_ok=True)
for dir_name in dir_lis:
cur_dir = os.path.join(self.base_exp_dir, 'recording', dir_name)
os.makedirs(cur_dir, exist_ok=True)
files = os.listdir(dir_name)
for f_name in files:
if f_name[-3:] == '.py':
copyfile(os.path.join(dir_name, f_name), os.path.join(cur_dir, f_name))
# copy configs
copyfile(self.conf_path, os.path.join(self.base_exp_dir, 'recording', 'config.conf'))
def load_checkpoint(self, checkpoint_name):
checkpoint = torch.load(os.path.join(self.base_exp_dir, 'checkpoints', checkpoint_name), map_location=self.device)
if self.model_type == 'neus':
self.nerf_outside.load_state_dict(checkpoint['nerf'])
self.sdf_network_fine.load_state_dict(checkpoint['sdf_network_fine'])
self.variance_network_fine.load_state_dict(checkpoint['variance_network_fine'])
self.color_network_fine.load_state_dict(checkpoint['color_network_fine'])
self.optimizer.load_state_dict(checkpoint['optimizer'])
self.iter_step = checkpoint['iter_step']
elif self.model_type == 'nerf':
self.nerf_coarse.load_state_dict(checkpoint['nerf_coarse'])
self.nerf_fine.load_state_dict(checkpoint['nerf_fine'])
self.nerf_outside.load_state_dict(checkpoint['nerf_outside'])
self.optimizer.load_state_dict(checkpoint['optimizer'])
self.iter_step = checkpoint['iter_step']
else:
NotImplementedError
def save_checkpoint(self):
checkpoint = None
if self.model_type == 'neus':
checkpoint = {
'nerf': self.nerf_outside.state_dict(),
'sdf_network_fine': self.sdf_network_fine.state_dict(),
'variance_network_fine': self.variance_network_fine.state_dict(),
'color_network_fine': self.color_network_fine.state_dict(),
'optimizer': self.optimizer.state_dict(),
'iter_step': self.iter_step,
}
elif self.model_type == 'nerf':
checkpoint = {
'nerf_coarse': self.nerf_coarse.state_dict(),
'nerf_fine': self.nerf_fine.state_dict(),
'nerf_outside': self.nerf_outside.state_dict(),
'optimizer': self.optimizer.state_dict(),
'iter_step': self.iter_step,
}
else:
NotImplementedError
os.makedirs(os.path.join(self.base_exp_dir, 'checkpoints'), exist_ok=True)
torch.save(checkpoint, os.path.join(self.base_exp_dir, 'checkpoints', 'ckpt_{:0>6d}.pth'.format(self.iter_step)))
def validate(self, input_model, loss_out, render_out):
mask, rays_o, rays_d, near, far = input_model['mask'], input_model['rays_o'], input_model['rays_d'], \
input_model['near'], input_model['far']
mask_sum = mask.sum() + 1e-5
loss, psnr = loss_out['Loss/loss'], loss_out['Log/psnr']
log_val = {}
if self.iter_step % self.report_freq == 0:
print('\n', self.base_exp_dir)
logging.info('iter:{:8>d} loss={:.03f} lr={:.06f} var={:.04f}'.format(self.iter_step, loss, self.optimizer.param_groups[0]['lr'], render_out['variance'].mean()))
ic((render_out['weight_sum'] * mask).sum() / mask_sum)
ic(self.get_alpha_inter_ratio())
ic(psnr)
if self.iter_step % self.save_freq == 0:
self.save_checkpoint()
if self.iter_step % self.val_mesh_freq == 0 or self.iter_step == 1:
self.validate_mesh()
if self.iter_step % self.val_image_freq == 0:
self.validate_image(save_normalmap_npz=self.save_normamap_npz)
if self.dataset.b_use_nvt:
self.validate_image(save_normalmap_npz=self.save_normamap_npz, b_novel_pose=True, validate_confidence=False)
if self.iter_step % self.val_fields_freq == 0:
self.validate_fields()
if self.iter_step % self.freq_valid_points == 0:
self.validate_points(rays_o, rays_d, near, far)
if self.iter_step % self.freq_save_confidence == 0:
self.save_accumulated_results()
self.save_accumulated_results(idx=self.dataset.n_images//2)
return log_val
def validate_image(self, idx=-1, resolution_level=-1,
save_normalmap_npz = False,
save_peak_value = False,
validate_confidence = True,
save_image_render = False,
b_novel_pose = False,
b_save_mask = False):
# validate image
ic(self.iter_step, idx)
logging.info(f'Validate begin: idx {idx}...')
if idx < 0:
idx = np.random.randint(self.dataset.n_images)
if resolution_level < 0:
resolution_level = self.validate_resolution_level
imgs_render = {}
for key in ['color_fine', 'confidence', 'normal', 'depth', 'variance_surface', 'confidence_mask', 'weight_sum']:
imgs_render[key] = []
if save_peak_value:
imgs_render.update({
'color_peak': [],
'normal_peak': [],
'depth_peak':[]
})
pts_peak_all = []
# (1) render images
if b_novel_pose:
rays_o, rays_d, _ = self.dataset.gen_rays_at_novel_pose(idx, resolution_level=resolution_level)
if rays_o is None:
return
else:
rays_o, rays_d = self.dataset.gen_rays_at(idx, resolution_level=resolution_level)
H, W, _ = rays_o.shape
rays_o = rays_o.reshape(-1, 3).split(self.batch_size)
rays_d = rays_d.reshape(-1, 3).split(self.batch_size)
idx_pixels_vu = torch.tensor([[i, j] for i in range(0, H) for j in range(0, W)]).split(self.batch_size)
for rays_o_batch, rays_d_batch, idx_pixels_vu_batch in zip(rays_o, rays_d, idx_pixels_vu):
near, far, _ = self.get_near_far(rays_o = rays_o_batch, rays_d = rays_d_batch)
background_rgb = torch.ones([1, 3]) if self.use_white_bkgd else None
render_out, _ = self.renderer.render(rays_o_batch, rays_d_batch, near, far, alpha_inter_ratio=self.get_alpha_inter_ratio(), background_rgb=background_rgb)
feasible = lambda key: ((key in render_out) and (render_out[key] is not None))
for key in imgs_render:
if feasible(key):
imgs_render[key].append(render_out[key].detach().cpu().numpy())
if validate_confidence:
pts_peak = rays_o_batch + rays_d_batch * render_out['depth_peak']
scores_all_mean, diff_patch_all, mask_valid_all = self.dataset.score_pixels_ncc(idx, pts_peak, render_out['normal_peak'], idx_pixels_vu_batch, reso_level=resolution_level)
imgs_render['confidence'].append(scores_all_mean.detach().cpu().numpy()[:,None])
imgs_render['confidence_mask'].append(mask_valid_all.detach().cpu().numpy()[:,None])
if save_peak_value:
pts_peak_all.append(pts_peak.detach().cpu().numpy())
del render_out
# (2) reshape rendered images
for key in imgs_render:
if len(imgs_render[key]) > 0:
imgs_render[key] = np.concatenate(imgs_render[key], axis=0)
if imgs_render[key].shape[1] == 3: # for color and normal
imgs_render[key] = imgs_render[key].reshape([H, W, 3])
elif imgs_render[key].shape[1] == 1:
imgs_render[key] = imgs_render[key].reshape([H, W])
# confidence map
if save_normalmap_npz:
# For each view, save point(.npz), depthmap(.png) point cloud(.ply), normalmap(.png), normal(.npz)
# rendered depth in volume rednering and projected depth of points in world are different
shape_depthmap = imgs_render['depth'].shape[:2]
pts_world = torch.vstack(rays_o).cpu().numpy() + torch.vstack(rays_d).cpu().numpy() * imgs_render['depth'].reshape(-1,1)
os.makedirs(os.path.join(self.base_exp_dir, 'depth'), exist_ok=True)
GeoUtils.save_points(os.path.join(self.base_exp_dir, 'depth', f'{self.iter_step:0>8d}_{idx}_reso{resolution_level}.ply'),
pts_world.reshape(-1,3),
colors = imgs_render['color_fine'].squeeze().reshape(-1,3),
normals= imgs_render['normal'].squeeze().reshape(-1,3),
BRG2RGB=True)
# save peak depth and normal
if save_peak_value:
pts_world_peak = torch.vstack(rays_o).cpu().numpy() + torch.vstack(rays_d).cpu().numpy() * imgs_render['depth_peak'].reshape(-1,1)
os.makedirs(os.path.join(self.base_exp_dir, 'depth'), exist_ok=True)
GeoUtils.save_points(os.path.join(self.base_exp_dir, 'depth', f'{self.iter_step:0>8d}_{idx}_reso{resolution_level}_peak.ply'),
pts_world_peak.reshape(-1,3),
colors = imgs_render['color_peak'].squeeze().reshape(-1,3),
normals= imgs_render['normal_peak'].squeeze().reshape(-1,3),
BRG2RGB=True)
os.makedirs(os.path.join(self.base_exp_dir, 'normal_peak'), exist_ok=True)
np.savez(os.path.join(self.base_exp_dir, 'normal_peak', f'{self.iter_step:08d}_{self.dataset.vec_stem_files[idx]}_reso{resolution_level}.npz'),
imgs_render['normal_peak'].squeeze())
os.makedirs(os.path.join(self.base_exp_dir, 'normal_render'), exist_ok=True)
np.savez(os.path.join(self.base_exp_dir, 'normal_render', f'{self.iter_step:08d}_{self.dataset.vec_stem_files[idx]}_reso{resolution_level}.npz'),
imgs_render['normal'].squeeze())
pts_world2 = pts_world_peak.reshape([H, W, 3])
np.savez(os.path.join(self.base_exp_dir, 'depth', f'{self.iter_step:08d}_{idx}_reso{resolution_level}_peak.npz'),
pts_world2)