-
Notifications
You must be signed in to change notification settings - Fork 0
/
run_gan.py
2479 lines (2016 loc) · 91.9 KB
/
run_gan.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
from __future__ import absolute_import
from email.mime import image
import os
import torch
import shutil
import imageio
import numpy as np
import pickle
import torch
from torch.utils.data import Dataset
from torchvision.transforms import Normalize
import torch.distributions as dist
import numpy as np
import cv2
from os.path import join
import glob
import deepdish as dd
from run_nerf import render_path
from run_nerf import config_parser as nerf_config_parser
from core.imutils import *
from core.pose_opt import load_poseopt_from_state_dict, pose_ckpt_to_pose_data
from core.load_data import generate_bullet_time, get_dataset
from core.raycasters import create_raycaster
from os.path import join
from core.utils.evaluation_helpers import txt_to_argstring
from core.utils.skeleton_utils import CMUSkeleton, get_smpl_l2ws_torch, nerf_c2w_to_extrinsic_torch, nerf_extrinsic_to_c2w_torch, smpl_rest_pose, get_smpl_l2ws, get_per_joint_coords
from core.utils.skeleton_utils import draw_skeletons_3d, rotate_x, rotate_y, axisang_to_rot, rot_to_axisang
from pytorch_msssim import SSIM
import torch.nn as nn
import time
from progress.bar import Bar
from collections import namedtuple
import numpy as np
import torch
import torch.nn as nn
import pytorch3d.transforms as torch3d
import os.path as path
import datetime
from tensorboardX import SummaryWriter
import copy
from scipy.spatial.transform import Rotation
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
from torch.utils.data import DataLoader,Dataset
from skimage.transform import rescale, resize, downscale_local_mean
from torchvision.models.detection import fasterrcnn_resnet50_fpn
from torchvision import transforms as T
import gc
gc.collect()
torch.cuda.empty_cache()
det_transform = T.Compose([T.ToTensor()])
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
import h5py
GT_PREFIXES = {
'h36m': 'data/h36m/',
'surreal': None,
'perfcap': 'data/',
'mixamo': 'data/mixamo',
}
def config_parser():
import configargparse
parser = configargparse.ArgumentParser()
parser.add_argument('--config', is_config_file=True,
help='config file path')
parser.add_argument('--batch_size', type=int, default=1024,
help='batch size')
# Training detail
parser.add_argument('--epochs', default=4, type=int, metavar='N', help='number of training epochs')
parser.add_argument('--decay_epoch', default=0, type=int, metavar='N', help='number of decay epochs')
# Learning rate
parser.add_argument('--lr_g', default=1.0e-4, type=float, metavar='LR', help='initial learning rate for augmentor/generator')
parser.add_argument('--lr_d', default=1.0e-4, type=float, metavar='LR', help='initial learning rate for discriminator')
parser.add_argument('--lr_p', default=1.0e-4, type=float, metavar='LR', help='initial learning rate for posenet')
parser.add_argument("--lr_spin", type=float, default=5e-5, help="Learning rate")
parser.add_argument('--no_max', dest='max_norm', action='store_false', help='if use max_norm clip on grad')
parser.add_argument('--df', default=2, type=int, help='update discriminator frequency')
parser.set_defaults(max_norm=True)
# nerf config
parser.add_argument('--nerf_args', type=str, required=True,
help='path to nerf configuration (args.txt in log)')
parser.add_argument('--ckptpath', type=str, required=True,
help='path to ckpt')
# render config
parser.add_argument('--render_res', nargs='+', type=int, default=[1000, 1000],
help='tuple of resolution in (H, W) for rendering')
parser.add_argument('--dataset', type=str, required=True,
help='dataset to render')
parser.add_argument('--entry', type=str, required=True,
help='entry in the dataset catalog to render')
parser.add_argument('--white_bkgd', action='store_true',
help='render with white background')
parser.add_argument('--render_type', type=str, default='retarget',
help='type of rendering to conduct')
parser.add_argument('--save_gt', action='store_true',
help='save gt frames')
parser.add_argument('--fps', type=int, default=14,
help='fps for video')
parser.add_argument('--rpi', type=int, default=20,
help='number of images to render per iteration')
parser.add_argument('--mesh_res', type=int, default=255,
help='resolution for marching cubes')
# kp-related
parser.add_argument('--render_refined', action='store_true',
help='render from refined poses')
parser.add_argument('--subject_idx', type=int, default=0,
help='which subject to render (for MINeRF)')
# frame-related
parser.add_argument('--selected_idxs', nargs='+', type=int, default=None,
help='hand-picked idxs for rendering')
parser.add_argument('--selected_framecode', type=int, default=None,
help='hand-picked framecode for rendering')
# saving
parser.add_argument('--outputdir', type=str, default='render_output/',
help='output directory')
parser.add_argument('--runname', type=str, required=True,
help='run name as an identifier ')
# evaluation
parser.add_argument('--eval', action='store_true',
help='to do evaluation at the end or not (only in bounding box)')
parser.add_argument('--no_save', action='store_true',
help='no image saving operation')
return parser
def load_nerf(args, nerf_args, skel_type=CMUSkeleton):
ckptpath = args.ckptpath
nerf_args.ft_path = args.ckptpath
# some info are unknown/not provided in nerf_args
# dig those out from state_dict directly
nerf_sdict = torch.load(ckptpath)
# get data_attrs used for training the models
data_attrs = get_dataset(nerf_args).get_meta()
if 'framecodes.codes.weight' in nerf_sdict['network_fn_state_dict']:
framecodes = nerf_sdict['network_fn_state_dict']['framecodes.codes.weight']
data_attrs['n_views'] = framecodes.shape[0]
# load poseopt_layer (if exist)
popt_layer = None
if nerf_args.opt_pose:
popt_layer = load_poseopt_from_state_dict(nerf_sdict)
nerf_args.finetune = True
render_kwargs_train, render_kwargs_test, _, grad_vars, _, _ = create_raycaster(nerf_args, data_attrs, device=device)
# freeze weights
for grad_var in grad_vars:
grad_var.requires_grad = False
render_kwargs_test['ray_caster'] = render_kwargs_train['ray_caster']
render_kwargs_test['ray_caster'].eval()
return render_kwargs_test, popt_layer
def load_render_data(args, bones,c2ws,nerf_args, poseopt_layer=None, opt_framecode=True):
# TODO: note that for models trained on SPIN data, they may not react well
catalog = init_catalog(args)[args.dataset][args.entry]
render_data = catalog.get(args.render_type, {})
data_h5 = catalog['data_h5']
# to real cameras (due to the extreme focal length they were trained on..)
# TODO: add loading with opt pose option
if poseopt_layer is not None:
rest_pose = poseopt_layer.get_rest_pose().cpu().numpy()[0]
# print("Load rest pose for poseopt!")
else:
try:
rest_pose = dd.io.load(data_h5, '/rest_pose')
# print("Load rest pose from h5!")
except:
rest_pose = smpl_rest_pose * dd.io.load(data_h5, '/ext_scale')
# print("Load smpl rest pose!")
if args.render_refined:
if 'refined' in catalog:
# print(f"loading refined poses from {catalog['refined']}")
#poseopt_layer = load_poseopt_from_state_dict(torch.load(catalog['refined']))
kps, bones = pose_ckpt_to_pose_data(catalog['refined'], legacy=True)[:2]
else:
with torch.no_grad():
bones = poseopt_layer.get_bones().cpu().numpy()
kps = poseopt_layer(np.arange(len(bones)))[0].cpu().numpy()
if render_data is not None:
render_data['refined'] = [kps, bones]
render_data['idx_map'] = catalog.get('idx_map', None)
else:
render_data = {'refined': [kps, bones],
'idx_map': catalog['idx_map']}
else:
render_data['idx_map'] = catalog.get('idx_map', None)
pose_keys = ['/kp3d', '/bones']
cam_keys = ['/c2ws', '/focals']
# Do partial load here!
# Need:
# 1. kps: for root location
# 2. bones: for bones
# 3. camera stuff: focals, c2ws, ... etc
_, focals = dd.io.load(data_h5, cam_keys)
_, H, W, _ = dd.io.load(data_h5, ['/img_shape'])[0]
# handel resolution
if args.render_res is not None:
assert len(args.render_res) == 2, "Image resolution should be in (H, W)"
H_r, W_r = args.render_res
# TODO: only check one side for now ...
scale = float(H_r) / float(H)
focals *= scale
H, W = H_r, W_r
# Load data based on type:
# bones = None
bg_imgs, bg_indices = None, None
if args.render_type in ['retarget', 'mesh']:
# print(f'Load data for retargeting!')
kps, skts, c2ws, cam_idxs, focals, bones = load_retarget(data_h5, bones,c2ws, focals,
rest_pose, pose_keys,
**render_data)
gt_paths, gt_mask_paths = None, None
is_gt_paths = True
subject_idxs = None
if nerf_args.nerf_type.startswith('minerf'):
subject_idxs = (np.ones((len(kps),)) * args.subject_idx).astype(np.int64)
ret_dict = {'kp': kps, 'skts': skts, 'render_poses': c2ws,
'cams': cam_idxs if opt_framecode else None,
'hwf': (H, W, focals),
'bones': bones,
'bg_imgs': bg_imgs,
'bg_indices': bg_indices,
'subject_idxs': subject_idxs}
return ret_dict
def init_catalog(args, n_bullet=10):
RenderCatalog = {
'h36m': None,
'surreal': None,
'perfcap': None,
'mixamo': None,
'3dhp': None,
}
def load_idxs(path):
if not os.path.exists(path):
# print(f'Index file {path} does not exist.')
return []
return np.load(path)
def set_dict(selected_idxs, **kwargs):
return {'selected_idxs': np.array(selected_idxs), **kwargs}
# H36M
s9_idx = np.arange(20)*20
h36m_s9 = {
'data_h5': 'data/h36m/S9_processed_h5py.h5',
'refined': 'neurips21_ckpt/trained/ours/h36m/s9_sub64_500k.tar',
'retarget': set_dict(s9_idx, length=5),
'bullet': set_dict(s9_idx, n_bullet=n_bullet, undo_rot=True,
center_cam=True),
'interpolate': set_dict(s9_idx, n_step=10, undo_rot=True,
center_cam=True),
'correction': set_dict(load_idxs('data/h36m/S9_top50_refined.npy')[:1], n_step=30),
'animate': set_dict([1000, 1059, 2400], n_step=10, center_cam=True, center_kps=True,
joints=np.array([17,19,21,23])),
'bubble': set_dict(s9_idx, n_step=30),
'poserot': set_dict(np.array([1000])),
'val': set_dict(load_idxs('data/h36m/S9_val_idxs.npy'), length=1, skip=1),
}
s11_idx = [213, 656, 904, 1559, 1815, 2200, 2611, 2700, 3110, 3440, 3605]
h36m_s11 = {
'data_h5': 'data/h36m/S11_processed_h5py.h5',
'refined': 'neurips21_ckpt/trained/ours/h36m/s11_sub64_500k.tar',
'retarget': set_dict(s11_idx, length=5),
'bullet': set_dict(s11_idx, n_bullet=n_bullet, undo_rot=True,
center_cam=True),
'interpolate': set_dict(s11_idx, n_step=10, undo_rot=True,
center_cam=True),
'correction': set_dict(load_idxs('data/h36m/S11_top50_refined.npy')[:1], n_step=30),
'animate': set_dict([2507, 700, 900], n_step=10, center_cam=True, center_kps=True,
joints=np.array([3,6,9,12,15,16,18])),
'bubble': set_dict(s11_idx, n_step=30),
'val': set_dict(load_idxs('data/h36m/S11_val_idxs.npy'), length=1, skip=1),
}
# SURREAL
easy_idx = [10, 70, 350, 420, 490, 910, 980, 1050]
surreal_val = {
'data_h5': 'data/surreal/surreal_val_h5py.h5',
'val': set_dict(load_idxs('data/surreal/surreal_val_idxs.npy'), length=1, skip=1),
'val2': set_dict(load_idxs('data/surreal/surreal_val_idxs.npy')[:300], length=1, skip=1),
}
surreal_easy = {
'data_h5': 'data/surreal/surreal_train_h5py.h5',
'retarget': set_dict(easy_idx, length=25, skip=2, center_kps=True),
'bullet': set_dict(easy_idx, n_bullet=n_bullet),
'bubble': set_dict(easy_idx, n_step=30),
}
hard_idx = [140, 210, 280, 490, 560, 630, 700, 770, 840, 910]
surreal_hard = {
'data_h5': 'data/surreal/surreal_train_h5py.h5',
'retarget': set_dict(hard_idx, length=60, skip=5, center_kps=True),
'bullet': set_dict([190, 210, 230, 490, 510, 530, 790, 810, 830, 910, 930, 950, 1090, 1110, 1130],
n_bullet=n_bullet, center_kps=True, center_cam=False),
'bubble': set_dict(hard_idx, n_step=30),
'val': set_dict(np.array([1200 * i + np.arange(420, 700)[::5] for i in range(0, 9, 2)]).reshape(-1), length=1, skip=1),
'mesh': set_dict([930], length=1, skip=1),
}
# PerfCap
weipeng_idx = [0, 50, 100, 150, 200, 250, 300, 350, 430, 480, 560,
600, 630, 660, 690, 720, 760, 810, 850, 900, 950, 1030,
1080, 1120]
perfcap_weipeng = {
'data_h5': 'data/MonoPerfCap/Weipeng_outdoor/Weipeng_outdoor_processed_h5py.h5',
'refined': 'neurips21_ckpt/trained/ours/perfcap/weipeng_tv_500k.tar',
'retarget': set_dict(weipeng_idx, length=30, skip=2),
'bullet': set_dict(weipeng_idx, n_bullet=n_bullet),
'interpolate': set_dict(weipeng_idx, n_step=10, undo_rot=True,
center_cam=True),
'bubble': set_dict(weipeng_idx, n_step=30),
'val': set_dict(np.arange(1151)[-230:], length=1, skip=1),
'animate': set_dict([300, 480, 700], n_step=10, center_cam=True, center_kps=True,
joints=np.array([1,4,7,10,17,19,21,23])),
}
nadia_idx = [0, 65, 100, 125, 230, 280, 410, 560, 600, 630, 730, 770,
830, 910, 1010, 1040, 1070, 1100, 1285, 1370, 1450, 1495,
1560, 1595]
perfcap_nadia = {
'data_h5': 'data/MonoPerfCap/Nadia_outdoor/Nadia_outdoor_processed_h5py.h5',
'refined': 'neurips21_ckpt/trained/ours/perfcap/nadia_tv_500k.tar',
'retarget': set_dict(nadia_idx, length=30, skip=2),
'bullet': set_dict(nadia_idx, n_bullet=n_bullet),
'interpolate': set_dict(nadia_idx, n_step=10, undo_rot=True,
center_cam=True, center_kps=True),
'bubble': set_dict(nadia_idx, n_step=30),
'animate': set_dict([280, 410, 1040], n_step=10, center_cam=True, center_kps=True,
joints=np.array([1,2,4,5,7,8,10,11])),
'val': set_dict(np.arange(1635)[-327:], length=1, skip=1),
}
# Mixamo
james_idx = [20, 78, 138, 118, 1149, 333, 3401, 2221, 4544]
mixamo_james = {
'data_h5': 'data/mixamo/James_processed_h5py.h5',
'idx_map': load_idxs('data/mixamo/James_selected.npy'),
'refined': 'neurips21_ckpt/trained/ours/mixamo/james_tv_500k.tar',
'retarget': set_dict(james_idx, length=30, skip=2),
'bullet': set_dict(james_idx, n_bullet=n_bullet, center_cam=True, center_kps=True),
'interpolate': set_dict(james_idx, n_step=10, undo_rot=True,
center_cam=True),
'bubble': set_dict(james_idx, n_step=30),
'animate': set_dict([3401, 1149, 4544], n_step=10, center_cam=True, center_kps=True,
joints=np.array([18,19,20,21,22,23])),
'mesh': set_dict([20, 78], length=1, undo_rot=False),
}
archer_idx = [158, 672, 374, 414, 1886, 2586, 2797, 4147, 4465]
mixamo_archer = {
'data_h5': 'data/mixamo/Archer_processed_h5py.h5',
'idx_map': load_idxs('data/mixamo/Archer_selected.npy'),
'refined': 'neurips21_ckpt/trained/ours/mixamo/archer_tv_500k.tar',
'retarget': set_dict(archer_idx, length=30, skip=2),
'bullet': set_dict(archer_idx, n_bullet=n_bullet, center_cam=True, center_kps=True),
'interpolate': set_dict(archer_idx, n_step=10, undo_rot=True,
center_cam=True),
'bubble': set_dict(archer_idx, n_step=30),
'animate': set_dict([1886, 2586, 4465], n_step=10, center_cam=True, center_kps=True,
joints=np.array([18,19,20,21,22,23])),
}
# NeuralBody
nb_subjects = ['315', '377', '386', '387', '390', '392', '393', '394']
# TODO: hard-coded: 6 views
nb_idxs = np.arange(len(np.concatenate([np.arange(1, 31), np.arange(400, 601)])) * 6)
nb_dict = lambda subject: {'data_h5': f'data/zju_mocap/{subject}_test_h5py.h5',
'val': set_dict(nb_idxs, length=1, skip=1)}
RenderCatalog['h36m'] = {
'S9': h36m_s9,
'S11': h36m_s11,
'gt_to_mask_map': ('imageSequence', 'Mask'),
}
RenderCatalog['surreal'] = {
'val': surreal_val,
'easy': surreal_easy,
'hard': surreal_hard,
}
RenderCatalog['perfcap'] = {
'weipeng': perfcap_weipeng,
'nadia': perfcap_nadia,
'gt_to_mask_map': ('images', 'masks'),
}
RenderCatalog['mixamo'] = {
'james': mixamo_james,
'archer': mixamo_archer,
}
RenderCatalog['neuralbody'] = {
f'{subject}': nb_dict(subject) for subject in nb_subjects
}
return RenderCatalog
def find_idxs_with_map(selected_idxs, idx_map):
if idx_map is None:
return selected_idxs
match_idxs = []
for sel in selected_idxs:
for i, m in enumerate(idx_map):
if m == sel:
match_idxs.append(i)
break
return np.array(match_idxs)
def load_retarget(pose_h5,bones, c2ws, focals, rest_pose, pose_keys,
selected_idxs, length, skip=1, refined=None,
center_kps=False, idx_map=None, is_surreal=False,
undo_rot=False, is_neuralbody=False):
bones=bones
cam_idxs=np.arange(bones.shape[0])
focals = focals[cam_idxs]
c2ws = c2ws[cam_idxs]
l2ws = np.array([get_smpl_l2ws(bone, rest_pose, 1.0) for bone in bones])
kps = l2ws[..., :3, -1]
skts = np.linalg.inv(l2ws)
return kps, skts, c2ws, cam_idxs, focals, bones
def to_tensors(data_dict):
tensor_dict = {}
for k in data_dict:
if isinstance(data_dict[k], np.ndarray):
if k == 'bg_indices' or k == 'subject_idxs':
tensor_dict[k] = torch.tensor(data_dict[k]).long()
else:
tensor_dict[k] = torch.tensor(data_dict[k]).float()
elif k == 'hwf' or k == 'cams' or k == 'bones':
tensor_dict[k] = data_dict[k]
elif data_dict[k] is None:
tensor_dict[k] = None
else:
raise NotImplementedError(f"{k}: only nparray and hwf are handled now!")
return tensor_dict
# self define tools
class Summary(object):
def __init__(self, directory):
self.directory = directory
self.epoch = 0
self.writer = None
self.phase = 0
self.train_iter_num = 0
self.train_realpose_iter_num = 0
self.train_fakepose_iter_num = 0
self.test_iter_num = 0
self.test_MPI3D_iter_num = 0
def create_summary(self):
self.writer = SummaryWriter(log_dir=os.path.join(self.directory))
return self.writer
def summary_train_iter_num_update(self):
self.train_iter_num = self.train_iter_num + 1
def summary_train_realpose_iter_num_update(self):
self.train_realpose_iter_num = self.train_realpose_iter_num + 1
def summary_train_fakepose_iter_num_update(self):
self.train_fakepose_iter_num = self.train_fakepose_iter_num + 1
def summary_test_iter_num_update(self):
self.test_iter_num = self.test_iter_num + 1
def summary_test_MPI3D_iter_num_update(self):
self.test_MPI3D_iter_num = self.test_MPI3D_iter_num + 1
def summary_epoch_update(self):
self.epoch = self.epoch + 1
def summary_phase_update(self):
self.phase = self.phase + 1
class Logger(object):
'''Save training process to log file with simple plot function.'''
def __init__(self, fpath, args, title=None, resume=False):
self.file = None
self.resume = resume
self.title = '' if title == None else title
if fpath is not None:
if resume:
self.file = open(fpath, 'r')
name = self.file.readline()
self.names = name.rstrip().split('\t')
self.numbers = {}
for _, name in enumerate(self.names):
self.numbers[name] = []
for numbers in self.file:
numbers = numbers.rstrip().split('\t')
for i in range(0, len(numbers)):
self.numbers[self.names[i]].append(numbers[i])
self.file.close()
self.file = open(fpath, 'a')
else:
self.file = open(fpath, 'w')
self.record_args(args)
def record_args(self, args):
self.file.write(str(args))
self.file.write('\n')
self.file.flush()
def set_names(self, names):
if self.resume:
pass
# initialize numbers as empty list
self.numbers = {}
self.names = names
for _, name in enumerate(self.names):
self.file.write(name)
self.file.write('\t')
self.numbers[name] = []
self.file.write('\n')
self.file.flush()
def append(self, numbers):
assert len(self.names) == len(numbers), 'Numbers do not match names'
for index, num in enumerate(numbers):
self.file.write("{0:.6f}".format(num))
self.file.write('\t')
self.numbers[self.names[index]].append(num)
self.file.write('\n')
self.file.flush()
def plot(self, names=None):
names = self.names if names == None else names
numbers = self.numbers
for _, name in enumerate(names):
x = np.arange(len(numbers[name]))
plt.plot(x, np.asarray(numbers[name]))
plt.legend([self.title + '(' + name + ')' for name in names])
plt.grid(True)
def close(self):
if self.file is not None:
self.file.close()
# To store data in a pool and sample from it when it is full
# Shrivastava et al’s strategy
class Sample_from_Pool(object):
def __init__(self, max_elements=4096):
self.max_elements = max_elements
self.cur_elements = 0
self.items = []
def __call__(self, in_items):
return_items = []
for in_item in in_items:
if self.cur_elements < self.max_elements:
self.items.append(in_item)
self.cur_elements = self.cur_elements + 1
return_items.append(in_item)
else:
if np.random.ranf() > 0.5:
idx = np.random.randint(0, self.max_elements)
tmp = copy.copy(self.items[idx])
self.items[idx] = in_item
return_items.append(tmp)
else:
return_items.append(in_item)
return return_items
class AverageMeter(object):
"""Computes and stores the average and current value"""
def __init__(self):
self.reset()
def reset(self):
self.val = 0
self.avg = 0
self.sum = 0
self.count = 0
def update(self, val, n=1):
self.val = val
self.sum += val * n
self.count += n
self.avg = self.sum / self.count
def lr_decay(optimizer, step, lr, decay_step, gamma):
lr = lr * gamma ** (step / decay_step)
for param_group in optimizer.param_groups:
param_group['lr'] = lr
return lr
def set_grad(nets, requires_grad=False):
for net in nets:
for param in net.parameters():
param.requires_grad = requires_grad
from torch.optim import lr_scheduler
def get_scheduler(optimizer, policy, nepoch_fix=None, nepoch=None, decay_step=None):
if policy == 'lambda':
def lambda_rule(epoch):
lr_l = 1.0 - max(0, epoch - nepoch_fix) / float(nepoch - nepoch_fix + 1)
return lr_l
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda_rule)
elif policy == 'step':
scheduler = lr_scheduler.StepLR(
optimizer, step_size=decay_step, gamma=0.1)
elif policy == 'plateau':
scheduler = lr_scheduler.ReduceLROnPlateau(
optimizer, mode='min', factor=0.2, threshold=0.01, patience=5)
else:
return NotImplementedError('learning rate policy [%s] is not implemented', policy)
return scheduler
def init_weights(m):
if isinstance(m, nn.Linear):
nn.init.kaiming_normal_(m.weight)
def swap_mat(mat):
# [right, -up, -forward]
# equivalent to right multiply by:
# [1, 0, 0, 0]
# [0,-1, 0, 0]
# [0, 0,-1, 0]
# [0, 0, 0, 1]
return np.concatenate([
mat[..., 0:1], -mat[..., 1:2], -mat[..., 2:3], mat[..., 3:]
], axis=-1)
def nerf_c2w_to_extrinsic(c2w):
return np.linalg.inv(swap_mat(c2w))
def coord_to_homogeneous(c):
assert c.shape[-1] == 3
if len(c.shape) == 2:
h = np.ones((c.shape[0], 1)).astype(c.dtype)
return np.concatenate([c, h], axis=1)
elif len(c.shape) == 1:
h = np.array([0, 0, 0, 1]).astype(c.dtype)
h[:3] = c
return h
else:
raise NotImplementedError(f"Input must be a 2-d or 1-d array, got {len(c.shape)}")
def focal_to_intrinsic_np(focal):
if isinstance(focal, float) or (len(focal.reshape(-1)) < 2):
focal_x = focal_y = focal
else:
focal_x, focal_y = focal
return np.array([[focal_x, 0, 0, 0],
[ 0, focal_y, 0, 0],
[ 0, 0, 1, 0]],
dtype=np.float32)
def world_to_cam(pts, extrinsic, H, W, focal, center=None):
if center is None:
offset_x = W * .5
offset_y = H * .5
else:
offset_x, offset_y = center
if pts.shape[-1] < 4:
pts = coord_to_homogeneous(pts)
intrinsic = focal_to_intrinsic_np(focal)
cam_pts = pts @ extrinsic.T @ intrinsic.T
cam_pts = cam_pts[..., :2] / cam_pts[..., 2:3]
cam_pts[cam_pts == np.inf] = 0.
cam_pts[..., 0] += offset_x
cam_pts[..., 1] += offset_y
return cam_pts
def skeleton3d_to_2d(kps, c2ws, H, W, focals, centers=None):
exts = np.array([nerf_c2w_to_extrinsic(c2w) for c2w in c2ws])
kp2ds = []
for i, (kp, ext) in enumerate(zip(kps, exts)):
f = focals[i] if not isinstance(focals, float) else focals
h = H if isinstance(H, int) else H[i]
w = W if isinstance(W, int) else W[i]
center = centers[i] if centers is not None else None
kp2d = world_to_cam(kp, ext, h, w, f, center)
kp2ds.append(kp2d)
return np.array(kp2ds)
def project_to_2d(kps, exts, H, W, focals, centers=None):
kps,exts,H,W,focals=torch.as_tensor(kps),torch.as_tensor(exts),torch.as_tensor(H),torch.as_tensor(W),torch.as_tensor(focals)
offset_x = W * .5
offset_y = H * .5
# exts=np.repeat(np.expand_dims(exts,axis=0),kps.shape[0],axis=0)
kps_homogeneous=torch.ones(kps.shape[0],kps.shape[1],4).to(device)
kps_homogeneous[:,:,:3]=kps
intrinsics=torch.zeros(kps.shape[0],3,4).to(device)
intrinsics[:,0,0]=focals[0]
intrinsics[:,1,1]=focals[1]
intrinsics[:,2,2]=1
intrinsics=intrinsics.float()
kps_homogeneous=kps_homogeneous.float()
exts=exts.float()
[email protected](1,2)
[email protected](1,2)
kp2ds=kp2ds[...,:2]/kp2ds[...,2:3]
kp2ds[kp2ds == np.inf] = 0.
kp2ds[..., 0] += offset_x
kp2ds[..., 1] += offset_y
return kp2ds, kp3ds[:,:,:3]
def normalize_screen_coordinates(X, w, h):
assert X.shape[-1] == 2
# Normalize so that [0, w] is mapped to [-1, 1], while preserving the aspect ratio
return X / w * 2 - torch.tensor([1, h / w]).to(device)
class Linear(nn.Module):
def __init__(self, linear_size):
super(Linear, self).__init__()
self.l_size = linear_size
self.relu = nn.LeakyReLU(inplace=True)
self.w1 = nn.Linear(self.l_size, self.l_size)
self.batch_norm1 = nn.BatchNorm1d(self.l_size)
self.w2 = nn.Linear(self.l_size, self.l_size)
self.batch_norm2 = nn.BatchNorm1d(self.l_size)
def forward(self, x):
y = self.w1(x)
y = self.batch_norm1(y)
y = self.relu(y)
y = self.w2(y)
y = self.batch_norm2(y)
y = self.relu(y)
return y
class PoseGenerator(nn.Module):
def __init__(self, args, input_size=24 * 3):
super(PoseGenerator, self).__init__()
self.BAprocess = BAGenerator(input_size=24 * 3,noise_channle=32)
self.RTprocess = RTGenerator(input_size=24 * 3) #target
def forward(self, inputs_3d):
'''
input: 3D pose
:param inputs_3d: nx16x3, with hip root
:return: nx16x3
'''
pose_ba = self.BAprocess(inputs_3d) # diff may be used for div loss
R,T,pose_rt = self.RTprocess(inputs_3d) # rt=(r,t) used for debug
return {'pose_ba': pose_ba,
'ba_diff': None,
'pose_bl': None,
'blr': None,
'pose_rt': pose_rt,
'R': R,
'T': T}
class BAGenerator(nn.Module):
def __init__(self, input_size, noise_channle=72, linear_size=256, num_stage=2, p_dropout=0.5):
super(BAGenerator, self).__init__()
self.linear_size = linear_size
self.p_dropout = p_dropout
self.num_stage = num_stage
self.noise_channle = noise_channle
# 3d joints
self.input_size = input_size # 24 * 3
# process input to linear size
self.w1 = nn.Linear(self.noise_channle, self.linear_size)
self.batch_norm1 = nn.BatchNorm1d(self.linear_size)
self.linear_stages = []
for l in range(num_stage):
self.linear_stages.append(Linear(self.linear_size))
self.linear_stages = nn.ModuleList(self.linear_stages)
# post processing
self.w2 = nn.Linear(self.linear_size,24*4) #*2+(self.input_size-3)//3
self.relu = nn.LeakyReLU(inplace=True)
def Sampler(self,dim0,device):
# Define the parameters for the Gaussian modes
num_modes = 5
means = [torch.ones(self.noise_channle)*-10, torch.ones(self.noise_channle)*-5, torch.zeros(self.noise_channle), torch.ones(self.noise_channle)*5, torch.ones(self.noise_channle)*10]
covariances = torch.eye(self.noise_channle) #torch.tensor([[[0.5, 0.2], [0.2, 0.5]]]*num_modes)
# Create a list of Gaussian distributions with different means and covariances
gaussians = [dist.MultivariateNormal(means[i], covariance_matrix=covariances,device=device) for i in range(num_modes)]
# Sample from each Gaussian distribution
samples_per_mode = dim0
samples = torch.cat([gaussian.sample([samples_per_mode]) for gaussian in gaussians], dim=0)
idxs=torch.randperm(samples.shape[0])
samples=samples[idxs]
return samples[:dim0]
def forward(self, inputs_3d):
'''
:param inputs_3d: nx16x3.
:return: nx16x3
'''
# pre-processing
noise = torch.randn(inputs_3d.shape[0], self.noise_channle, device=inputs_3d.device) #*2-1
# noise=self.Sampler(inputs_3d.shape[0],device=inputs_3d.device)
# noise = noise / noise.norm(dim=1, keepdim=True)
y = self.w1(noise)
y = self.batch_norm1(y)
y = self.relu(y)
# linear layers
for i in range(self.num_stage):
y = self.linear_stages[i](y)
y = self.w2(y)
y = y.view(inputs_3d.size(0), -1, 4)
y_axis=y[:,:,:3]
y_axis = y_axis/torch.linalg.norm(y_axis,dim=-1,keepdim=True)
y_theta =y[:,:,3:4]
y_theta=y_theta
out = y_axis*y_theta
out[:,0]*=3.14*2
return out
class RTGenerator(nn.Module):
def __init__(self, input_size, noise_channle=72, linear_size=256, num_stage=2, p_dropout=0.5):
super(RTGenerator, self).__init__()
'''
:param input_size: n x 16 x 3
:param output_size: R T 3 3 -> get new pose for pose 3d projection.
'''
self.linear_size = linear_size
self.p_dropout = p_dropout
self.num_stage = num_stage
self.noise_channle = noise_channle
# 3d joints
self.input_size = input_size # 16 * 3
# process input to linear size -> for R
self.w1_R = nn.Linear(self.noise_channle, self.linear_size)
self.batch_norm_R = nn.BatchNorm1d(self.linear_size)
self.linear_stages_R = []
for l in range(num_stage):
self.linear_stages_R.append(Linear(self.linear_size))
self.linear_stages_R = nn.ModuleList(self.linear_stages_R)
# process input to linear size -> for T
self.w1_T = nn.Linear(self.noise_channle, self.linear_size)
self.batch_norm_T = nn.BatchNorm1d(self.linear_size)
self.linear_stages_T = []
for l in range(num_stage):
self.linear_stages_T.append(Linear(self.linear_size))
self.linear_stages_T = nn.ModuleList(self.linear_stages_T)
# post processing
self.w2_R = nn.Linear(self.linear_size, 7)
self.w2_T = nn.Linear(self.linear_size, 3)
self.relu = nn.LeakyReLU(inplace=True)
def forward(self,inputs_3d):
'''
:param inputs_3d: nx16x3
:return: nx16x3
'''
# caculate R
noise = torch.randn(inputs_3d.shape[0], self.noise_channle, device=inputs_3d.device)
r = self.w1_R(noise) #torch.cat((x, noise), dim=1)
r = self.batch_norm_R(r)
r = self.relu(r)
for i in range(self.num_stage):
r = self.linear_stages_R[i](r)
# r = self.w2_R(r)
r_mean=r[:,:3]
r_std=r[:,3:6]*r[:,3:6]
r_axis = torch.normal(mean=r_mean,std=r_std)
r_axis = r_axis/torch.linalg.norm(r_axis,dim=-1,keepdim=True)
r_axis = r_axis*r[:,6:7]
rM=torch3d.axis_angle_to_matrix(r_axis) #axis_angle
# caculate T
noise = torch.randn(inputs_3d.shape[0], self.noise_channle, device=inputs_3d.device)
t = self.w1_T(noise) #torch.cat((x, noise), dim=1)
t = self.batch_norm_T(t)
t = self.relu(t)
for i in range(self.num_stage):
t = self.linear_stages_T[i](t)
t = self.w2_T(t)
t[:, 2] = t[:, 2].clone() * t[:, 2].clone()
t = t.view(inputs_3d.shape[0],1, 3) # Nx1x3 translation t
inputs_3d=inputs_3d-inputs_3d[:,:1,:]
inputs_3d=inputs_3d.permute(0,2,1).contiguous()
out=torch.matmul(rM,inputs_3d)
out=out.permute(0,2,1).contiguous()
out=out+t
return rM, t[:,0] , out # return r t for debug
class Disc_Joint_Path(nn.Module):
def __init__(self, num_joints=1, channel=500, channel_mid=1000):
super(Disc_Joint_Path, self).__init__()
# KCS path
self.layer_1 = nn.Linear(num_joints*3, channel)
self.layer_2 = nn.Linear(channel, channel)
self.layer_3 = nn.Linear(channel, channel)
self.layer_last = nn.Linear(channel, channel_mid)
self.layer_pred = nn.Linear(channel_mid, 1)
self.relu = nn.LeakyReLU()
self.relu = nn.LeakyReLU()
def forward(self, x):
# KCS path
x=self.relu(self.layer_1(x))
x=self.relu(self.layer_2(x))
x=self.relu(self.layer_3(x))
x=self.relu(self.layer_last(x))