-
Notifications
You must be signed in to change notification settings - Fork 6
/
engine.py
1347 lines (1142 loc) · 68 KB
/
engine.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 torch
from easydict import EasyDict as edict
import os
from torch.utils.data import DataLoader
import time
import tqdm
import imageio
from collections import OrderedDict
import numpy as np
import re
import math
import torch.nn.functional as F
from misc.utils import log, compute_image_diff, compute_depth_diff
from misc.metrics import EvalTools
from misc.train_helpers import summarize_loss, summarize_metrics_list
from datasets import datas_dict
from models.murf import MuRF
from misc import utils
from loss import SSIM
import lpips
from misc.depth_viz import viz_depth_tensor
from datasets.ibrnet_mix.create_training_dataset import create_training_dataset
from datasets.transforms import RandomCrop
class Engine():
def __init__(self, opts):
super().__init__()
self.opts = opts
self.n_src_views = opts.n_src_views
self.epoch_start = 0
self.iter_start = 0
os.makedirs(opts.output_path, exist_ok=True)
def L1_loss(self, pred, label=0):
loss = (pred.contiguous() - label).abs()
return loss.mean()
def load_dataset(self, splits):
# load training data
log.info(f"loading datasets...")
for split in splits:
if getattr(self.opts, f'data_{split}', None):
if split == 'test':
data_opts_list = [
v for _, v in self.opts.data_test.items()]
self.test_loaders = []
else:
data_opts_list = [getattr(self.opts, f'data_{split}')]
for data_opts in data_opts_list:
if data_opts is None:
continue
scene_list = getattr(data_opts, "scene_list", None)
test_views_method = getattr(
data_opts, "test_views_method", "nearest")
if split == 'train' and getattr(self.opts, 'mix_data_train', False):
assert self.opts.dist is True
cur_dataset, train_sampler = create_training_dataset(root_dir=data_opts.root_dir,
n_views=self.n_src_views,
distributed=self.opts.dist,
train_dataset=getattr(
self.opts, 'mix_datasets', "llff+spaces+ibrnet_collected+realestate+google_scanned"),
dataset_weights=getattr(self.opts, 'dataset_weights', [
0.3, 0.15, 0.35, 0.15, 0.05]),
num_replicas=torch.cuda.device_count(),
rank=self.opts.local_rank,
mixall=getattr(
self.opts, 'mixall', False),
no_random_view=getattr(
self.opts, 'no_random_view', False),
dataset_replicas=getattr(
self.opts, 'dataset_replicas', None),
realestate_full_set=getattr(
self.opts, 'realestate_full_set', False),
realestate_frame_dir=getattr(
self.opts, 'realestate_frame_dir', None),
mixall_random_view=getattr(
self.opts, 'mixall_random_view', False),
realestate_use_all_scenes=getattr(
self.opts, 'realestate_use_all_scenes', False),
)
self.train_sampler = train_sampler
# only support batch size 1 currently due to different image resolutions in the mixed datasets
assert self.opts.batch_size == 1
else:
cur_dataset = datas_dict[data_opts.dataset_name](data_opts.root_dir,
split=split,
pose_root=getattr(
data_opts, 'pose_dir', None),
n_views=self.n_src_views, img_wh=data_opts.img_wh, max_len=data_opts.max_len,
scene_list=scene_list,
test_views_method=test_views_method,
random_crop=(split == 'train' and getattr(
self.opts, 'random_crop', False)),
crop_height=getattr(
self.opts, 'crop_height', None),
crop_width=getattr(
self.opts, 'crop_width', None),
max_crop_height=getattr(
self.opts, 'max_crop_height', None),
max_crop_width=getattr(
self.opts, 'max_crop_width', None),
min_crop_height=getattr(
self.opts, 'min_crop_height', None),
min_crop_width=getattr(
self.opts, 'min_crop_width', None),
min_views=getattr(
self.opts, 'min_views', None),
max_views=getattr(
self.opts, 'max_views', None),
render_scan_id=getattr(
data_opts, 'render_scan_id', None),
random_resize=(split == 'train' and getattr(
self.opts, 'random_resize', False)),
test_scan_name=getattr(
self.opts, 'test_scan_name', None),
fixed_test_set=getattr(
self.opts, 'fixed_realestate_test_set', False),
frame_distance=getattr(
self.opts, 'realestate_frame_distance', 128),
view_selection_stride=getattr(
self.opts, 'view_selection_stride', 1),
img_scale_factor=getattr(
self.opts, 'llff_img_scale_factor', 4),
test_view_stride=getattr(
self.opts, 'dtu_test_view_stride', None),
window_size=getattr(
self.opts, 'realestate_window_size', 128),
continuous_view=getattr(
self.opts, 'dtu_continuous_view', False),
test_scene_name=getattr(
self.opts, 'test_scene_name', None),
downsample_factor=getattr(
self.opts, 'mipnerf360_downsample_factor', 8),
fixed_target_frame=getattr(
self.opts, 'realestate_fixed_target_frame', False),
)
if split == 'train' and self.opts.dist:
train_sampler = torch.utils.data.distributed.DistributedSampler(
cur_dataset,
num_replicas=torch.cuda.device_count(),
rank=self.opts.local_rank)
self.train_sampler = train_sampler
else:
train_sampler = None
if split == 'train':
shuffle = False if self.opts.dist else True
else:
shuffle = False
cur_loader = DataLoader(cur_dataset, shuffle=shuffle, num_workers=data_opts.num_workers,
batch_size=self.opts.batch_size if split == 'train' else 1, pin_memory=True,
sampler=train_sampler,
)
if split == 'test':
self.test_loaders.append(cur_loader)
else:
setattr(self, f"{split}_loader", cur_loader)
log.info(
f" * loaded {split} set of {data_opts.dataset_name}")
def build_networks(self):
log.info("building networks...")
self.model = MuRF(
self.opts).to(self.opts.device)
log.info(self.model)
if self.opts.encoder.pretrain_weight and (not self.opts.load) and (not self.opts.resume):
utils.load_gmflow_checkpoint(self.model.feat_enc, self.opts.encoder.pretrain_weight, self.opts.device,
gmflow_n_blocks=self.opts.encoder.num_transformer_layers,
no_strict_load=getattr(
self.opts, 'no_strict_load', False),
)
log.info(
f"loaded gmflow pretrained weight for encoder from {self.opts.encoder.pretrain_weight}.")
if self.opts.dist:
self.model_ddp = torch.nn.parallel.DistributedDataParallel(
self.model.to(self.opts.device),
device_ids=[self.opts.local_rank],
output_device=self.opts.local_rank,
find_unused_parameters=False,
)
self.model = self.model_ddp.module
# number of parameters
num_params = sum([p.numel()
for p in self.model.parameters() if p.requires_grad])
print('Nubmer of parameters: %d' % num_params)
# load lpips network for loss
if hasattr(self.opts, 'loss_weight') and getattr(self.opts.loss_weight, 'lpips', False):
self.lpips_loss_func = lpips.LPIPS(net='vgg').to(self.opts.device)
for param in self.lpips_loss_func.parameters():
param.requires_grad = False
# ssim loss
if hasattr(self.opts, 'loss_weight') and getattr(self.opts.loss_weight, 'ssim', False):
self.ssim_loss_func = SSIM(patch_size=getattr(
self.opts, 'ssim_patch_size', 3)).to(self.opts.device)
def setup_optimizer(self):
log.info("setting up optimizers...")
# load trainable params
optim_params = [dict(params=self.model.feat_enc.parameters(), lr=self.opts.optim.lr_enc),
dict(params=self.model.nerf_dec.parameters(), lr=self.opts.optim.lr_dec)]
lr_lists = [self.opts.optim.lr_enc, self.opts.optim.lr_dec]
# fine nerf
if hasattr(self.model, 'nerf_dec_fine'):
if getattr(self.opts, 'freeze_coarse_nerf', False):
for params in self.model.feat_enc.parameters():
params.requires_grad = False
for params in self.model.nerf_dec.parameters():
params.requires_grad = False
optim_params = [
dict(params=self.model.nerf_dec_fine.parameters(), lr=self.opts.optim.lr_dec)]
lr_lists = [self.opts.optim.lr_dec]
trainable_params = sum(
[p.numel() for p in self.model.nerf_dec_fine.parameters() if p.requires_grad])
# fine encoder
optim_params.append(
dict(params=self.model.feat_enc_fine.parameters(),
lr=self.opts.optim.lr_enc)
)
lr_lists.append(self.opts.optim.lr_enc)
trainable_params += sum(
[p.numel() for p in self.model.feat_enc_fine.parameters() if p.requires_grad])
print('Trainable parameters: %d' % trainable_params)
else:
optim_params.append(
dict(params=self.model.nerf_dec_fine.parameters(), lr=self.opts.optim.lr_dec))
lr_lists.append(self.opts.optim.lr_dec)
# set up optimizer
optim_type = self.opts.optim.algo.type
optim_kwargs = {k: v for k,
v in self.opts.optim.algo.items() if k != "type"}
self.optim = getattr(torch.optim, optim_type)(
optim_params, **optim_kwargs)
info = f" * {optim_type} optimizer (" + ', '.join(
[f'{k}={v}' for k, v in optim_kwargs.items()]) + ')'
log.info(info)
# set up scheduler if needed
self.sched_type = None
if self.opts.optim.sched:
sched_type = self.opts.optim.sched.type
sched_kwargs = {k: v for k,
v in self.opts.optim.sched.items() if k != "type"}
info = f" * {sched_type} scheduler"
if sched_type == 'OneCycleLR': # set additional param accordingly
assert hasattr(
self, 'train_loader'), "Must initialize the training data, to calculate total steps for OneCycleLR"
steps_per_epoch = len(self.train_loader)
if self.opts.batch_size > 1:
sched_kwargs.update(dict(
total_steps=self.opts.max_epoch * steps_per_epoch + 50,
max_lr=lr_lists,
))
else:
sched_kwargs.update(dict(
epochs=self.opts.max_epoch, steps_per_epoch=steps_per_epoch, max_lr=lr_lists,
))
self.sched_type = sched_type
self.sched = getattr(torch.optim.lr_scheduler, sched_type)(
self.optim, **sched_kwargs)
if getattr(self.opts, 'resume_train_iter', False):
# ugly but works
for _ in range(self.opts.resume_train_iter):
self.optim.step()
self.sched.step()
info = info + \
' (' + ', '.join([f'{k}={v}' for k,
v in sched_kwargs.items()]) + ')'
log.info(info)
def restore_checkpoint(self):
epoch_start, iter_start = 0, 0
if self.opts.resume:
log.info("resuming from previous checkpoint...")
ckpt_path = os.path.join(
self.opts.output_path, 'models', 'latest.pth')
if not os.path.isfile(ckpt_path):
log.warn(f"can NOT find previous checkpoints at {ckpt_path}")
log.warn("start training from scratch.")
else:
optims_scheds = {x: getattr(self, x) for x in [
'optim', 'sched'] if hasattr(self, x)}
epoch_start, iter_start = utils.restore_checkpoint(self.model, ckpt_path=ckpt_path,
device=self.opts.device, log=log, resume=True,
optims_scheds=optims_scheds,
no_strict_load=getattr(
self.opts, 'no_strict_load', False),
)
elif self.opts.load is not None:
log.info("loading weights from checkpoint {}...".format(self.opts.load))
epoch_start, iter_start = utils.restore_checkpoint(
self.model, ckpt_path=self.opts.load, device=self.opts.device, log=log,
no_strict_load=getattr(self.opts, 'no_strict_load', False),
)
else:
log.info("initializing weights from scratch...")
self.epoch_start = epoch_start or 0
self.iter_start = iter_start or 0
if getattr(self.opts, 'coarse_init_fine', False):
assert self.opts.load is not None
# training fine nerf with pretrained coarse nerf as init
log.info('initializing fine encoder with pretrained coarse encoder')
self.model.feat_enc_fine.load_state_dict(
self.model.feat_enc.state_dict(), strict=True)
if getattr(self.opts, 'load_coarse', False):
log.info('load coarse pretrained model')
checkpoint = torch.load(self.opts.load_coarse)['model']
# only partially load the coarse model
self.model.load_state_dict(checkpoint, strict=False)
if getattr(self.opts, 'resume_train_iter', False):
assert getattr(self.opts, 'resume_train_epoch', False)
self.epoch_start = self.opts.resume_train_epoch
self.iter_start = self.opts.resume_train_iter
def setup_visualizer(self):
log.info("setting up visualizers...")
if self.opts.tb and self.opts.local_rank == 0:
from torch.utils import tensorboard
self.tb = tensorboard.SummaryWriter(
log_dir=self.opts.output_path, flush_secs=10)
def train_model(self):
# before training
log.title("TRAINING START")
self.send_results("TRAINING START", log_msg=False)
self.timer = edict(start=time.time(), it_mean=None)
self.it = self.iter_start
self.val_it = math.ceil(self.opts.freq.val_it * len(self.train_loader)
) if self.opts.freq.val_it > 0 else self.opts.freq.val_it
self.ckpt_it = math.ceil(self.opts.freq.ckpt_it * len(self.train_loader)
) if self.opts.freq.ckpt_it > 0 else self.opts.freq.ckpt_it
# training
# resume at the end of one epoch or middle
if self.epoch_start == 0:
epoch_start = 0
else:
if self.epoch_start * len(self.train_loader) > self.iter_start:
epoch_start = self.epoch_start - 1
else:
epoch_start = self.epoch_start
for self.ep in range(epoch_start, self.opts.max_epoch):
if self.opts.dist:
self.train_sampler.set_epoch(self.ep)
self.train_epoch()
if getattr(self.opts, 'stop_epoch', False):
if self.ep >= self.opts.stop_epoch - 1:
break
# after training
if self.opts.local_rank == 0 and self.opts.tb:
self.tb.flush()
self.tb.close()
log.title("TRAINING DONE")
self.send_results("TRAINING DONE", reset_status=True, log_msg=False)
def train_epoch(self):
# before train epoch
if self.opts.dist:
self.model_ddp.train()
else:
self.model.train()
# train epoch
tqdm_bar = tqdm.tqdm(
self.train_loader, desc="training epoch {}".format(self.ep + 1), leave=False)
for batch_idx, batch in enumerate(tqdm_bar):
# train iteration
if self.opts.resume and self.ep * len(self.train_loader) + batch_idx < self.iter_start:
continue
if getattr(self.opts, 'mix_data_train', False) or hasattr(self.opts, 'max_crop_height') or hasattr(self.opts, 'max_crop_width'):
# random crop due to different image resolutions
# print(batch['images'].shape) # [B, V, 3, H, W], B=1
# random crop
ori_h, ori_w = batch['images'].shape[3:]
assert hasattr(self.opts, 'max_crop_height')
assert hasattr(self.opts, 'max_crop_width')
max_crop_height = getattr(self.opts, 'max_crop_height')
max_crop_width = getattr(self.opts, 'max_crop_width')
# nearest size to be divisable by 16 or 32
resize_factor = getattr(self.opts, 'resize_factor', 16)
near_h = ori_h // resize_factor * resize_factor
near_w = ori_w // resize_factor * resize_factor
landscape = ori_w > ori_h
if landscape:
crop_h = min(near_h, max_crop_height)
crop_w = min(near_w, max_crop_width)
else:
crop_h = min(near_h, max_crop_width)
crop_w = min(near_w, max_crop_height)
batch = RandomCrop(crop_h, crop_w, with_batch_dim=True)(batch)
var = edict(batch)
var = utils.move_to_device(var, self.opts.device)
loss = self.train_iteration(var)
tqdm_bar.set_postfix(it=self.it, loss="{:.3f}".format(loss.all))
if self.sched_type == 'OneCycleLR':
self.sched.step()
# after train epoch
if batch_idx == len(self.train_loader) - 1:
lr_dict = self.get_cur_lrates()
log.loss_train(self.opts, self.ep + 1,
lr_dict, loss.all, self.timer)
if self.sched_type is not None and self.sched_type != 'OneCycleLR':
self.sched.step()
if self.opts.freq.test_ep > 0 and (self.ep + 1) % self.opts.freq.test_ep == 0:
if getattr(self.opts, 'test_rank0', False):
if self.opts.local_rank == 0:
self.test_model(ep=self.ep + 1, send_log=True, save_images=False,
with_depth_metric=getattr(self.opts, 'with_depth_metric', False))
else:
self.test_model(ep=self.ep + 1, send_log=True, save_images=False,
with_depth_metric=getattr(self.opts, 'with_depth_metric', False))
if self.opts.freq.ckpt_ep > 0 and (self.ep + 1) % self.opts.freq.ckpt_ep == 0 and self.opts.local_rank == 0:
self.save_checkpoint(ep=self.ep + 1, it=self.it, backup_ckpt=True)
def train_iteration(self, var):
# before train iteration
self.timer.it_start = time.time()
# train iteration
self.optim.zero_grad()
if self.opts.dist:
var_pred = self.model_ddp(var, mode="train")
else:
var_pred = self.model(var, mode="train")
loss = self.compute_loss(var_pred, var, mode="train")
loss = summarize_loss(loss, self.opts.loss_weight)
if isinstance(loss.all, float):
print('float loss, skip backward')
return loss
if torch.isnan(loss.all):
print('nan loss, skip backward')
return loss
loss.all.backward()
if self.opts.optim.clip_enc is not None:
torch.nn.utils.clip_grad_norm_(
self.model.feat_enc.parameters(), self.opts.optim.clip_enc)
self.optim.step()
# after train iteration
self.it += 1
self.timer.it_end = time.time()
utils.update_timer(self.opts, self.timer, self.ep,
len(self.train_loader))
if self.opts.freq.scalar > 0 and self.it % self.opts.freq.scalar == 0:
cur_lrates = self.get_cur_lrates()
self.log_scalars(loss, self.opts.loss_weight,
lrates=cur_lrates, step=self.it, split="train")
if self.ckpt_it > 0 and self.it % self.ckpt_it == 0 and self.opts.local_rank == 0:
self.save_checkpoint(ep=self.ep + 1, it=self.it,
backup_ckpt=getattr(
self.opts, 'backup_ckpt', False),
backup_latest_ckpt=getattr(self.opts.freq, 'latest_ckpt_ep', False) and (
(self.ep + 1) % self.opts.freq.latest_ckpt_ep == 0)
)
if getattr(self.opts.freq, 'latest_ckpt_iter', False) and self.opts.local_rank == 0:
self.save_checkpoint(ep=self.ep + 1, it=self.it,
backup_ckpt=False,
backup_latest_ckpt_iter=self.it % self.opts.freq.latest_ckpt_iter == 0,
)
if self.val_it > 0 and self.it % self.val_it == 0 and hasattr(self, 'val_loader') and self.opts.local_rank == 0:
self.validate_model(iter=self.it,
with_depth_metric=getattr(self.opts, 'with_depth_metric', False))
# log training images to get some sense of the training data
if getattr(self.opts, 'log_train_imgs', False) and self.opts.freq.scalar > 0 and self.it % self.opts.freq.scalar == 0 and self.opts.local_rank == 0:
# log pred and gt images
batch_size = var['images'].shape[0]
img_hw = var['img_wh'][0].cpu().numpy().tolist()[::-1]
pred_rgb = var_pred.rgb.reshape(batch_size, *img_hw, -1)
pred_rgb_tb = pred_rgb.permute(0, 3, 1, 2) # [B, 3, H, W]
target_gt_tb = var.images[:, -1] # [B, 3, H, W]
img_pred_gt = torch.cat((pred_rgb_tb, target_gt_tb), dim=-1)
self.tb.add_image('0train/img_pred_gt',
img_pred_gt[0], self.it) # [3, H, W*2]
# log input views
b, v, _, h, w = var['images'].size()
input_rgbs = var['images'].permute(0, 2, 3, 4, 1).reshape(
b, 3, h, w, v).to(pred_rgb_tb.device) # [B, 3, H, W*(V+1)]
img_list = [rgb[:, :, :, :, 0]
for rgb in input_rgbs.chunk(v, dim=-1)]
# [B, 3, H, W*(V-1)], source image
img_input = torch.cat(img_list[:-1], dim=-1)
self.tb.add_image('0train/img_input',
img_input[0], self.it) # [3, H, W*(V-1)]
# log depth
if getattr(self.opts, 'log_train_depth', False):
pred_depth = var_pred['depth'].reshape(
batch_size, *img_hw).detach() # [B, H, W]
# vis depth (inverse depth)
pred_depth_viz = viz_depth_tensor(
1. / pred_depth[0].cpu()).float() / 255. # [3, H, W]
self.tb.add_image('0train/depth_pred', pred_depth_viz, self.it)
# check gpu
if self.it == 5 and self.opts.local_rank == 0:
os.system('nvidia-smi')
return loss
def compute_loss(self, pred, src, mode=None):
self.n_src_views = src.images.size(1) - 1
loss = edict()
batch_size, n_views, n_chnl = src.images.shape[:3]
assert n_views == (
self.n_src_views + 1), "Make sure the last views are provided as the GT target view"
# (b, h*w, 3)
target_gt = src.images[:, -
1].reshape(batch_size, n_chnl, -1).permute(0, 2, 1)
if getattr(self.opts.nerf, f"rand_rays_{mode}", False) and mode in ["train", "test-optim"] and not getattr(self.opts, 'radiance_downsample_factor', False):
target_gt = target_gt[:, pred.ray_idx]
# compute image losses
if self.opts.loss_weight.render is not None:
loss.render = self.L1_loss(pred.rgb, target_gt)
# ssim loss
if getattr(self.opts.loss_weight, 'ssim', False):
tmp_img_gt = target_gt
tmp_img_pred = pred.rgb
img_hw = src['img_wh'][0].cpu().numpy().tolist()[::-1]
curr_h, curr_w = img_hw
assert target_gt.size(1) == curr_h * curr_w
assert pred.rgb.size(1) == curr_h * curr_w
tmp_img_gt = tmp_img_gt.view(batch_size, curr_h, curr_w, 3).permute(
0, 3, 1, 2) # [B, 3, H, W]
tmp_img_pred = tmp_img_pred.view(
batch_size, curr_h, curr_w, 3).permute(0, 3, 1, 2)
loss.ssim = self.ssim_loss_func(tmp_img_gt, tmp_img_pred).mean()
# lpips loss
if getattr(self.opts.loss_weight, 'lpips', False):
tmp_img_gt = target_gt
tmp_img_pred = pred.rgb
img_hw = src['img_wh'][0].cpu().numpy().tolist()[::-1]
curr_h, curr_w = img_hw
assert target_gt.size(1) == curr_h * curr_w
assert pred.rgb.size(1) == curr_h * curr_w
tmp_img_gt = tmp_img_gt.view(batch_size, curr_h, curr_w, 3).permute(
0, 3, 1, 2) # [B, 3, H, W]
tmp_img_pred = tmp_img_pred.view(
batch_size, curr_h, curr_w, 3).permute(0, 3, 1, 2)
# images must be in [-1, 1] for computing lpips loss
tmp_img_gt = tmp_img_gt * 2 - 1
tmp_img_pred = tmp_img_pred * 2 - 1
# assert tmp_img_gt.min() >= -1 and tmp_img_gt.max() <= 1 and tmp_img_pred.min() >= -1 and tmp_img_pred.max() <= 1
loss.lpips = self.lpips_loss_func(tmp_img_gt, tmp_img_pred).mean()
return loss
@torch.no_grad()
def log_scalars(self, loss=None, loss_weight=None, metric=None, lrates=None, step=0, split="train"):
if loss is not None:
for key, value in loss.items():
if key == "all":
continue
if loss_weight[key] is not None and self.opts.local_rank == 0:
self.tb.add_scalar(
"{0}/loss_{1}".format(split, key), value, step)
if metric is not None and self.opts.local_rank == 0:
for key, value in metric.items():
mean_value = np.array(value).mean()
self.tb.add_scalar(
"{0}/{1}".format(split, key), mean_value, step)
if lrates is not None and self.opts.local_rank == 0:
for key, value in lrates.items():
self.tb.add_scalar("{0}/{1}".format('lrate', key), value, step)
@torch.no_grad()
def get_cur_lrates(self):
if self.opts.optim.sched:
if getattr(self.opts, 'freeze_encoder', False) or getattr(self.opts, 'freeze_coarse_nerf', False):
lr_enc = 0.
lr_dec = self.sched.get_last_lr()[0]
else:
lr_enc, lr_dec = self.sched.get_last_lr()[:2]
else:
lr_enc = self.opts.optim.lr_enc
lr_dec = self.opts.optim.lr_dec
lr_dict = dict(enc=lr_enc, dec=lr_dec)
if self.opts.nerf.fine_sampling:
lr_dict['dec_fine'] = lr_dec
return lr_dict
def save_checkpoint(self, ep=0, it=0, backup_ckpt=True, backup_latest_ckpt=False, backup_latest_ckpt_iter=False):
save_train_info = True
checkpoint = dict(model=self.model.state_dict())
if save_train_info:
train_info = dict(optim=self.optim.state_dict())
if self.sched_type is not None:
train_info.update(dict(sched=self.sched.state_dict()))
checkpoint.update(train_info)
utils.save_checkpoint(self.opts.output_path,
checkpoint, ep=ep, it=it,
backup_ckpt=backup_ckpt,
backup_latest_ckpt=backup_latest_ckpt,
backup_latest_ckpt_iter=backup_latest_ckpt_iter,
)
def send_results(self, msg, reset_status=False, log_msg=True):
if log_msg:
log.metric_test(re.sub('<[^<]+?>', '', msg.split('\n')[-1]))
@torch.no_grad()
def validate_model(self, iter=None,
with_depth_metric=False,
):
assert hasattr(self, 'val_loader'), "please load validation dataset."
self.model.eval()
data_outdir = os.path.join(self.opts.output_path, 'validation')
os.makedirs(data_outdir, exist_ok=True)
eval_tools = EvalTools(device=self.opts.device)
metrics_dict = {k: [] for k in eval_tools.support_metrics}
if not with_depth_metric:
del metrics_dict['depth_abs']
# 5 validation samples
tqdm_loader = tqdm.tqdm(
self.val_loader, desc="validating", leave=False)
for batch_idx, batch in enumerate(tqdm_loader):
if iter == 0 and batch_idx > 0:
break
var = edict(batch)
batch_size = var.images.shape[0]
var = utils.move_to_device(var, self.opts.device)
# nearest size to be divisable by 16 or 32
resize_factor = getattr(self.opts, 'resize_factor', 16)
if var.images.shape[3] % resize_factor != 0 or var.images.shape[4] % resize_factor != 0:
near_inference_size = [var.images.shape[3] // resize_factor *
resize_factor, var.images.shape[4] // resize_factor * resize_factor]
# print(near_inference_size)
else:
near_inference_size = None
if near_inference_size is not None:
inference_size = near_inference_size
ori_size = var.images.shape[3:]
scale_factor_y = inference_size[0] / ori_size[0]
scale_factor_x = inference_size[1] / ori_size[1]
batch_size = var.images.shape[0]
num_views = var.images.shape[1]
ori_images = var.images.clone()
tmp_imgs = var.images.view(-1, 3, *ori_size) # [B*V, 3, H, W]
tmp_imgs = F.interpolate(
tmp_imgs, size=inference_size, mode='bilinear', align_corners=True)
var.images = tmp_imgs.view(
batch_size, num_views, 3, *inference_size)
# update intrinsics
intrinsic = var.intrinsics.clone() # [B, V, 3, 3]
intrinsic[:, :, :1] = intrinsic[:, :, :1] * scale_factor_x
intrinsic[:, :, 1:2] = intrinsic[:, :, 1:2] * scale_factor_y
var.intrinsics = intrinsic
# update size
ori_wh = var.img_wh.clone()
var.img_wh[:, 0] = inference_size[1]
var.img_wh[:, 1] = inference_size[0]
var = self.model(var, mode="val")
# resize back
if near_inference_size is not None:
# resize rgb
var.images = ori_images
var.img_wh = ori_wh
tmp_rgb = var.rgb.reshape(
batch_size, *inference_size, 3).permute(0, 3, 1, 2) # [B, 3, H, W]
tmp_rgb = F.interpolate(
tmp_rgb, size=ori_size, mode='bilinear', align_corners=True)
var.rgb = tmp_rgb.view(
batch_size, 3, -1).permute(0, 2, 1) # [B, H*W, 3]
# resize depth
tmp_depth = var.depth.reshape(
batch_size, *inference_size, 1).permute(0, 3, 1, 2) # [B, 1, H, W]
tmp_depth = F.interpolate(
tmp_depth, size=ori_size, mode='bilinear', align_corners=True)
var.depth = tmp_depth.view(batch_size, -1) # [B, H*W]
batch['img_wh'] = ori_wh.cpu()
# save image
img_hw = batch['img_wh'][0].numpy().tolist()[::-1]
pred_rgb = var['rgb'].reshape(
batch_size, *img_hw, -1) # [B, H, W, 3]
for batch_idx, cur_rgb in enumerate(pred_rgb):
pred_rgb_nb = cur_rgb.detach().cpu().numpy()
# h,w,3
gt_rgb_nb = var.images[batch_idx, -
1].permute(1, 2, 0).detach().cpu().numpy()
if 'dtu' == self.val_loader.dataset.get_name():
assert 'depth' in batch, "Must provide 'depth' of target view for validation"
depth = batch['depth'][batch_idx].detach().cpu().numpy()
image_mask = depth == 0
elif 'mask' in batch: # regnerf evaluation
# 0 is background
image_mask = batch['mask'][batch_idx] == 0
else:
image_mask = None
eval_tools.set_inputs(pred_rgb_nb, gt_rgb_nb, image_mask,
pred_depth=var['depth'].reshape(batch_size, *img_hw)[batch_idx].detach(
).cpu().numpy() if with_depth_metric and 'depth' in var else None,
gt_depth=var['depth_gt'].reshape(batch_size, *img_hw)[batch_idx].detach(
).cpu().numpy() if with_depth_metric and 'depth_gt' in var else None,
)
cur_metrics = eval_tools.get_metrics()
for k, v in cur_metrics.items():
metrics_dict[k].append(v)
self.log_scalars(metric=metrics_dict, step=iter, split="val")
if self.opts.dist:
self.model_ddp.train()
else:
self.model.train()
@torch.no_grad()
def test_model(self, ep=None, send_log=True, save_images=False, leave_tqdm=False,
save_depth=True,
save_gt_depth=True,
save_depth_np=False,
save_gt_depth_np=False,
with_depth_metric=False,
test_on_val_set=False,
):
if test_on_val_set:
assert hasattr(
self, 'val_loader'), "Must load the val data for testing."
test_loaders = [self.val_loader]
else:
assert hasattr(
self, 'test_loaders'), "Must load the test data for testing."
test_loaders = self.test_loaders
test_outroot = os.path.join(self.opts.output_path, 'test')
os.makedirs(test_outroot, exist_ok=True)
eval_tools = EvalTools(device=self.opts.device)
metrics_dict = {}
metrics_list_dict = {}
self.model.eval()
for data_loader in test_loaders:
dataname = data_loader.dataset.get_name()
metrics_dict[dataname] = OrderedDict()
metrics_list_dict[dataname] = []
data_outdir = os.path.join(test_outroot, dataname)
if getattr(self.opts, 'save_name_suffix', None):
data_outdir = data_outdir + '_' + self.opts.save_name_suffix
os.makedirs(data_outdir, exist_ok=True)
# tensorboard summary images
num_summary_images = 8
sample_interval = len(data_loader) // num_summary_images if len(
data_loader) % num_summary_images == 0 else len(data_loader) // num_summary_images + 1
count = 0
tqdm_desc = f"testing {dataname}" if ep is None else f"testing {dataname} [epoch {ep}]"
for batch_idx, batch in enumerate(tqdm.tqdm(data_loader, desc=tqdm_desc, leave=leave_tqdm)):
if hasattr(self, "it") and self.it == 0 and batch_idx > 0:
break
var = edict(batch)
var = utils.move_to_device(var, self.opts.device)
# nearest size to be divisable by 16 or 32
resize_factor = getattr(self.opts, 'resize_factor', 16)
if var.images.shape[3] % resize_factor != 0 or var.images.shape[4] % resize_factor != 0:
near_inference_size = [var.images.shape[3] // resize_factor *
resize_factor, var.images.shape[4] // resize_factor * resize_factor]
else:
near_inference_size = None
# resize then inference
if getattr(self.opts, 'inference_size', False) or near_inference_size is not None:
if getattr(self.opts, 'llff_inference_size', False) and dataname == 'llff':
inference_size = self.opts.llff_inference_size
elif getattr(self.opts, 'blender_inference_size', False) and dataname == 'blender':
inference_size = self.opts.blender_inference_size
else:
inference_size = self.opts.inference_size if getattr(
self.opts, 'inference_size', False) else near_inference_size
ori_size = var.images.shape[3:]
scale_factor_y = inference_size[0] / ori_size[0]
scale_factor_x = inference_size[1] / ori_size[1]
batch_size = var.images.shape[0]
num_views = var.images.shape[1]
ori_images = var.images.clone()
# [B*V, 3, H, W]
tmp_imgs = var.images.view(-1, 3, *ori_size)
tmp_imgs = F.interpolate(
tmp_imgs, size=inference_size, mode='bilinear', align_corners=True)
var.images = tmp_imgs.view(
batch_size, num_views, 3, *inference_size)
# update intrinsics
intrinsic = var.intrinsics.clone() # [B, V, 3, 3]
intrinsic[:, :, :1] = intrinsic[:, :, :1] * scale_factor_x
intrinsic[:, :, 1:2] = intrinsic[:,
:, 1:2] * scale_factor_y
var.intrinsics = intrinsic
# update size
ori_wh = var.img_wh.clone()
var.img_wh[:, 0] = inference_size[1]
var.img_wh[:, 1] = inference_size[0]
var = self.model(var, mode="test")
if var['rgb'].isnan().any():
print('pred nan')
print(var['rgb'].isnan().sum())
var['rgb'] = torch.nan_to_num(var['rgb'], 0.)
if var['images'].isnan().any():
print('ori image nan')
# resize back
if not getattr(self.opts, 'no_resize_back', False) and (getattr(self.opts, 'inference_size', False) or near_inference_size is not None):
# resize rgb
var.images = ori_images
var.img_wh = ori_wh
tmp_rgb = var.rgb.reshape(
batch_size, *inference_size, 3).permute(0, 3, 1, 2) # [B, 3, H, W]
tmp_rgb = F.interpolate(
tmp_rgb, size=ori_size, mode='bilinear', align_corners=True)
var.rgb = tmp_rgb.view(
batch_size, 3, -1).permute(0, 2, 1) # [B, H*W, 3]
# resize depth
tmp_depth = var.depth.reshape(
batch_size, *inference_size, 1).permute(0, 3, 1, 2) # [B, 1, H, W]
tmp_depth = F.interpolate(
tmp_depth, size=ori_size, mode='bilinear', align_corners=True)
var.depth = tmp_depth.view(batch_size, -1) # [B, H*W]
batch['img_wh'] = ori_wh.cpu()
if getattr(self.opts, 'no_resize_back', False):
batch['img_wh'][0, 0] = self.opts.inference_size[1]
batch['img_wh'][0, 1] = self.opts.inference_size[0]
# save image
batch_size = var['images'].shape[0]
img_hw = batch['img_wh'][0].numpy().tolist()[::-1]
pred_rgb = var['rgb'].reshape(batch_size, *img_hw, -1)
# tensorboard summary
if hasattr(self, 'tb') and batch_idx % sample_interval == 0:
pred_rgb_tb = pred_rgb.permute(0, 3, 1, 2) # [B, 3, H, W]
b, _, h, w = pred_rgb_tb.shape
v = batch['images'].size(1)
input_rgbs = batch['images'].permute(0, 2, 3, 4, 1).reshape(
b, 3, h, w, v).to(pred_rgb_tb.device) # [B, 3, H, W*(V+1)]
img_list = [rgb[:, :, :, :, 0]
for rgb in input_rgbs.chunk(v, dim=-1)]
# [B, 3, H, W*(V-1)], source image
img_input = torch.cat(img_list[:-1], dim=-1)
# [B, 3, H, W*2], pred and gt
img_pred_gt = torch.cat(
[pred_rgb_tb, img_list[-1]], dim=-1)
img_error = compute_image_diff(
pred_rgb_tb, batch['images'][:, -1].to(pred_rgb_tb.device)) # [B, H, W]
img_pred_gt_error = torch.cat((img_pred_gt, img_error.unsqueeze(
1).repeat(1, 3, 1, 1)), dim=-1) # [B, 3, H, W*3]
self.tb.add_image('%s_%d/0_img_pred_gt_error' % (
dataname, batch_idx // sample_interval), img_pred_gt_error[0], ep) # [3, H, W*3]
self.tb.add_image('%s_%d/1_img_input' % (dataname, batch_idx //
sample_interval), img_input[0], ep) # [3, H, W*(V+2)]
# summary depth if have
if 'depth' in batch:
pred_depth = var['depth'].reshape(
batch_size, *img_hw) # [B, H, W]
gt_depth = batch['depth'].to(
pred_depth.device) # [B, H, W]
# vis depth
pred_depth_viz = viz_depth_tensor(
pred_depth[0].cpu()) # [3, H, W]
gt_depth_viz = viz_depth_tensor(gt_depth[0].cpu())
depth_pred_gt = torch.cat(
(pred_depth_viz, gt_depth_viz), dim=-1) # [3, H, W*2]
# vis depth error
depth_error = compute_depth_diff(pred_depth, gt_depth, valid_mask=gt_depth > 0.,
min_depth=batch['near_fars'][0, -1,
0], max_depth=batch['near_fars'][0, -1, 1],
).cpu()
depth_pred_gt_error = torch.cat((depth_pred_gt.float(
) / 255., depth_error.unsqueeze(1).repeat(1, 3, 1, 1)[0]), dim=-1)
self.tb.add_image('%s_%d/2_depth_pred_gt_error' % (
dataname, batch_idx // sample_interval), depth_pred_gt_error, ep)
# log depth prediction
if getattr(self.opts, 'log_test_depth', False):
pred_depth = var['depth'].reshape(
batch_size, *img_hw).detach() # [B, H, W]
# vis depth (inverse depth)
pred_depth_viz = viz_depth_tensor(
1. / pred_depth[0].cpu()).float() / 255. # [3, H, W]
self.tb.add_image(
'%s_%d/2_depth_pred' % (dataname, batch_idx // sample_interval), pred_depth_viz, ep)
if save_images and save_depth and 'depth' in var:
pred_depth = var['depth'].reshape(
batch_size, *img_hw) # [B, H, W]