/
datasets.py
1162 lines (1040 loc) · 58.6 KB
/
datasets.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
import os
from glob import glob
import itertools
from collections import OrderedDict
import numpy as np
import json
import openmesh as om
import time
from collections import defaultdict
import pickle
# from sklearn.neighbors import NearestNeighbors
import joblib
import pymesh
from multiprocessing.pool import ThreadPool
import pytorch_points.utils.pc_utils as pc_utils
from pytorch_points.network.geo_operations import normalize_point_batch_to_sphere, compute_face_normals_and_areas, furthest_point_sample
from pytorch_points.network.operations import scatter_add
from pytorch_points.utils.pytorch_utils import check_values, AverageValueMeter
from pytorch_points.utils.geometry_utils import Mesh, get_edge_points
from pytorch_points.misc import logger
from losses import LabeledChamferDistance
from common import find_files, renderMeshes, call_proc, read_trimesh
labeled_chamfer_distance = LabeledChamferDistance(beta=1.0, gamma=1, delta=0)
def _numpy_chamfer(P1, P2):
"""
used for sklearn NearestNeighbor P1, P2 (1, 3xP)
"""
P1 = torch.from_numpy(P1).view(-1, 3).unsqueeze(0).cuda().float()
P2 = torch.from_numpy(P2).view(-1, 3).unsqueeze(0).cuda().float()
size_1 = P1.size(1)
orig_size_tmp_1 = (P1.view(-1)[P1.view(-1) == 100]).nelement()
orig_size_1 = orig_size_tmp_1 / 3
assert orig_size_1 * 3 == orig_size_tmp_1
orig_size_1 = size_1 - orig_size_1
size_2 = P2.size(1)
orig_size_tmp_2 = (P2.view(-1)[P2.view(-1) == 100]).nelement()
orig_size_2 = orig_size_tmp_2 / 3
assert orig_size_2 * 3 == orig_size_tmp_2
orig_size_2 = size_2 - orig_size_2
dist, _, _ = labeled_chamfer_distance(P1, P2)
dist = dist.cpu().numpy()
return dist
class FileListDataset(torch.utils.data.Dataset):
def __init__(self, opt, normalization_function=lambda x: pc_utils.normalize_to_box(x)[0]):
super().__init__()
self.mesh_data = True
self.source_files = opt.source_model
self.target_files = opt.target_model
pref = os.path.commonpath([os.path.dirname(f) for f in self.source_files])
self.source_names = [os.path.relpath(f, pref) for f in self.source_files]
pref = os.path.commonpath([os.path.dirname(f) for f in self.target_files])
self.target_names = [os.path.relpath(f, pref) for f in self.target_files]
assert(len(self.source_files) == len(self.target_files))
self.opt = opt
self.normalize = normalization_function or (lambda x: pc_utils.normalize_to_box(x)[0])
def __len__(self):
return len(self.target_files)
def __getitem__(self,idx):
source_file = self.source_files[idx]
target_file = self.target_files[idx]
# TODO handle point inputs
if source_file[-4:] in (".pts"):
source_mesh = pc_utils.load(source_file)
source_face = None
else:
if self.opt.is_poly:
source_mesh = om.read_polymesh(source_file)
source_mesh = source_mesh.points().copy()
source_face = None
else:
source_mesh, source_face = read_trimesh(source_file,clean=False)
source_face = torch.from_numpy(source_face).to(torch.int64)
source_mesh = torch.from_numpy(source_mesh[:,:3].astype(np.float32))
source_mesh = self.normalize(source_mesh)
if target_file[-4:] in (".pts"):
target_mesh = pc_utils.load(target_file)
target_face = None
else:
if self.opt.is_poly:
target_mesh = om.read_polymesh(target_file)
target_mesh = target_mesh.points().copy()
target_face = None
else:
target_mesh, target_face = read_trimesh(target_file,clean=False)
target_face = torch.from_numpy(target_face).to(torch.int64)
target_mesh = torch.from_numpy(target_mesh[:,:3].astype(np.float32))
target_mesh = self.normalize(target_mesh)
# sample source point
if source_mesh.shape[0] != self.opt.num_point:
source_shape = furthest_point_sample(source_mesh.unsqueeze(0).cuda(), self.opt.num_point, NCHW=False)[1]
source_shape = source_shape.squeeze(0)
else:
source_shape = source_mesh.clone()
if target_mesh.shape[0] != self.opt.num_point:
target_shape = furthest_point_sample(target_mesh.unsqueeze(0).cuda(), self.opt.num_point, NCHW=False)[1]
target_shape = target_shape.squeeze(0)
else:
target_shape = target_mesh.clone()
source_fn = os.path.splitext(self.source_names[idx])[0].replace("/","_")
target_fn = os.path.splitext(self.target_names[idx])[0].replace("/","_")
if self.opt.is_poly:
source_mesh = source_file
target_mesh = target_file
return {"source_shape": source_shape, "source_mesh": source_mesh, "source_file": source_fn,
"target_shape": target_shape, "target_mesh": target_mesh, "target_file": target_fn,}
return {"source_shape": source_shape, "source_face": source_face, "source_mesh": source_mesh, "source_file": source_fn,
"target_shape": target_shape, "target_face": target_face, "target_mesh": target_mesh, "target_file": target_fn,}
@staticmethod
def uncollate(batch_data):
source_shape = batch_data["source_shape"][:,:,:3].detach().cuda().contiguous()
target_shape = batch_data["target_shape"][:,:,:3].detach().cuda().contiguous()
if "source_mesh" not in batch_data:
source_mesh = None
elif isinstance(batch_data["source_mesh"], torch.Tensor):
source_mesh = batch_data["source_mesh"].cuda().detach()
else:
source_mesh = batch_data["source_mesh"]
if "target_mesh" not in batch_data:
target_mesh = None
elif isinstance(batch_data["target_mesh"], torch.Tensor):
target_mesh = batch_data["target_mesh"].cuda().detach()
else:
target_mesh = batch_data["target_mesh"]
source_face = batch_data["source_face"].cuda().detach() if "source_face" in batch_data else None
target_face = batch_data["target_face"].cuda().detach() if "target_face" in batch_data else None
return {
"source_shape": source_shape, "source_file": batch_data["source_file"],
"target_shape": target_shape, "target_file": batch_data["target_file"],
"source_mesh": source_mesh, "source_face": source_face, "target_mesh": target_mesh, "target_face": target_face
}
class CoSegDataset(torch.utils.data.Dataset):
def __init__(self, root_dir="/home/mnt/points/data/Coseg_Wang/Coseg_Wang/", use_init_cage=False, cat="Vase300", phase="train", max=-1):
super().__init__()
self.mesh_data = False
self.use_init_cage = use_init_cage
self.root_dir = root_dir
self.cat = cat
self.label_dir = os.path.join(self.root_dir,"{}GT".format(self.cat))
self.sym_plane = ("yz", "xz")
if phase=="train":
# self.training_source = glob(os.path.join(root_dir, "training", "registrations", "tr_reg_*.ply"))
self.training_source = glob(os.path.join(root_dir, "{}Points".format(cat), "*.pts"))
self.training_target = self.training_source[::-1]
assert(len(self.training_source) > 0)
assert(len(self.training_target) > 0)
if max > 0:
self.training_source = self.training_source[:max]
self.training_target = self.training_target[:max]
if use_init_cage:
# self.cage_dir = os.path.join(root_dir, "training", "simplified_registrations")
RuntimeWarning("Coseg dataset doesn't support initial cage")
self.use_init_cage = False
# self.training_target = self.training_source[:]
else:
test_list = os.path.join(root_dir, "{}Shapes".format(cat), "test.txt")
self.testing_source, self.testing_target = np.loadtxt(os.path.join(root_dir, "{}Shapes".format(cat), "test.txt"), dtype=str, unpack=True)
self.testing_source = self.testing_source[:max]
self.testing_target = self.testing_target[:max]
self.testing_dir = os.path.join(root_dir, "{}Shapes".format(cat))
self.get_testing_source = lambda x: os.path.join(root_dir, "{}Points".format(cat), "{}.pts".format(self.testing_source[x]))
self.get_testing_target = lambda x: os.path.join(root_dir, "{}Points".format(cat), "{}.pts".format(self.testing_target[x]))
# self.get_source_mesh = lambda x: os.path.join(root_dir, "{}Shapes".format(cat), "{}.obj".format(self.testing_source[x]))
# self.get_target_mesh = lambda x: os.path.join(root_dir, "{}Shapes".format(cat), "{}.obj".format(self.testing_target[x]))
self.get_source_mesh = lambda x: os.path.join(root_dir, "{}Shapes".format(cat), "{}.off".format(self.testing_source[x]))
self.get_target_mesh = lambda x: os.path.join(root_dir, "{}Shapes".format(cat), "{}.off".format(self.testing_target[x]))
self.isTrain = phase == "train"
self.get_default_label_path = lambda fname: os.path.join(self.label_dir,"{}.seg".format(fname))
self.rot_mat = torch.from_numpy(pc_utils.get_3D_rot_matrix(0, -np.pi/2).astype(np.float32))
@staticmethod
def normalize(shape):
shape = pc_utils.normalize_to_box(shape)[0]
return shape
def __getitem__(self, idx):
if self.isTrain:
TARGET_IDX = idx
INPUT_IDX = np.random.randint(0, len(self.training_source))
source_file = self.training_source[INPUT_IDX]
target_file = self.training_target[TARGET_IDX]
else:
source_file = self.get_testing_source(idx)
target_file = self.get_testing_target(idx)
# source
# point coordinates (Nx3), normals (Nx3) and labels (Nx1)
try:
pnl = np.loadtxt(source_file, dtype=np.float32, converters={6: lambda x: np.float32(x[1:-1])})
V = pnl[:,:3]
# seems that the normals are inverted
V_n = -pnl[:,3:6]
V_l = pnl[:,6:]
except IndexError:
pnl = np.loadtxt(source_file, dtype=np.float32)
V = pnl[:,:3]
V_n = pnl[:,3:6]
V_l = np.zeros_like(pnl[:,:1])
V, center, scale = pc_utils.normalize_to_box(V)
V = torch.from_numpy(V)
V_n = torch.from_numpy(V_n)
V_l = torch.from_numpy(V_l)
# target
try:
pnl = np.loadtxt(target_file, dtype=np.float32, converters={6: lambda x: np.float32(x[1:-1])})
V_target = pnl[:,:3]
# seems that the normals are inverted
V_n_target = -pnl[:,3:6]
V_l_target = pnl[:,6:]
except IndexError:
pnl = np.loadtxt(target_file, dtype=np.float32)
V_target = pnl[:,:3]
V_n_target = pnl[:,3:6]
V_l_target = np.zeros_like(pnl[:,:1])
V_target, _, _ = pc_utils.normalize_to_box(V_target)
V_target = torch.from_numpy(V_target)
V_n_target = torch.from_numpy(V_n_target)
V_l_target = torch.from_numpy(V_l_target)
# V = V.matmul(self.rot_mat)
# V_n = V_n.matmul(self.rot_mat)
# V_target = V_target.matmul(self.rot_mat)
# V_n_target = V_n_target.matmul(self.rot_mat)
V = torch.cat([V, V_n], dim=-1)
V_target = torch.cat([V_target, V_n_target], dim=-1)
# source_label_file = self.get_testing_source_label(idx)
# target_label_file = self.get_testing_target_label(idx)
if self.isTrain:
return {"source_shape": V, "source_label": V_l, "source_file": os.path.basename(source_file)[:-4],
"target_shape": V_target, "target_label": V_l_target, "target_file": os.path.basename(target_file)[:-4]}
V_mesh, F_mesh = read_trimesh(self.get_source_mesh(idx))
V_mesh = V_mesh[:,:3]
F_mesh = F_mesh[:,:3]
# V_mesh = (V_mesh - center)/scale
V_mesh, _, _ = pc_utils.normalize_to_box(V_mesh)
V_mesh = torch.from_numpy(V_mesh).to(dtype=torch.float)
F_mesh = torch.from_numpy(F_mesh).to(dtype=torch.int64)
assert(F_mesh.max() < V_mesh.shape[0])
V_mesh_target, F_mesh_target = read_trimesh(self.get_target_mesh(idx))
V_mesh_target = V_mesh_target[:,:3]
F_mesh_target = F_mesh_target[:,:3]
# V_mesh_target = (V_mesh_target - center)/scale
V_mesh_target, _, _ = pc_utils.normalize_to_box(V_mesh_target)
V_mesh_target = torch.from_numpy(V_mesh_target).to(dtype=torch.float)
F_mesh_target = torch.from_numpy(F_mesh_target).to(dtype=torch.int64)
assert(F_mesh_target.max() < V_mesh_target.shape[0])
# V_mesh_target = V_mesh_target.matmul(self.rot_mat)
# V_mesh = V_mesh.matmul(self.rot_mat)
return {"source_shape": V, "source_label": V_l, "source_file": os.path.basename(source_file)[:-4],
"target_shape": V_target, "target_label": V_l_target, "target_file": os.path.basename(target_file)[:-4],
"source_mesh": V_mesh, "source_face": F_mesh,
"target_mesh": V_mesh_target, "target_face": F_mesh_target }
@staticmethod
def uncollate(batch_data):
source_normals = batch_data["source_shape"][:,:,3:6].detach().cuda().contiguous()
target_normals = batch_data["target_shape"][:,:,3:6].detach().cuda().contiguous()
source_shape = batch_data["source_shape"][:,:,:3].detach().cuda().contiguous()
target_shape = batch_data["target_shape"][:,:,:3].detach().cuda().contiguous()
source_mesh = batch_data["source_mesh"].cuda().detach() if "source_mesh" in batch_data else None
source_face = batch_data["source_face"].cuda().detach() if "source_face" in batch_data else None
target_mesh = batch_data["target_mesh"].cuda().detach() if "target_mesh" in batch_data else None
target_face = batch_data["target_face"].cuda().detach() if "target_face" in batch_data else None
return {
"source_shape": source_shape, "source_normals": source_normals, "source_label": batch_data["source_label"].detach().cuda(), "source_file": batch_data["source_file"],
"target_shape": target_shape, "target_normals": target_normals, "target_label": batch_data["target_label"].detach().cuda(), "target_file": batch_data["target_file"],
"source_mesh": source_mesh, "source_face": source_face, "target_mesh": target_mesh, "target_face": target_face
}
def __len__(self):
if self.isTrain:
return len(self.training_source)
else:
return len(self.testing_source)
@staticmethod
def render_result(shape_dir, **kwargs):
"""render shapes inside a directory with thea"""
forward = (-.5,-.5,0)
pos = (5,5,0)
up = (0,0,1)
renderMeshes(shape_dir, forward=forward, pos=pos, up=up, **kwargs)
def _unwrap_self(arg, **kwarg):
return arg[0]._getitem(*(arg[1:]), **kwarg)
class ShapeNetSeg(torch.utils.data.Dataset):
def __init__(self, root_dir="/home/mnt/points/data/ShapeNet/PartSeg_v0/shapenetcore_partanno_segmentation_benchmark_v0_normal",
mesh_dir="/home/mnt/points/data/ShapeNet/ShapeNetCore.v1.remesh",
num_samples=-1, normalization="BoundingBox", knn=False,
num_neighbors=20,
class_choice=None, phase="train", normal=False, npoints=2500, sample=True,
data_augmentation_Z_rotation=False, data_augmentation_Z_rotation_range=360,
data_augmentation_3D_rotation=False, random_translation=False, anisotropic_scaling=False,
shuffle=False, use_fixed_pairs=False, use_preprocessed=True, isV2=False, **kwargs):
self.mesh_data = False
self.use_init_cage = False
self.use_preprocessed = use_preprocessed
self.isV2 = isV2
self.v2_rot_mat = torch.from_numpy(pc_utils.get_3D_rot_matrix(1, np.pi/2).astype(np.float32))
self.dataset_string_args = str(phase) + "_" + \
str(class_choice) + \
"_" + str(num_samples) + \
"_" + str(2500)
self.mesh_dir = mesh_dir
self.path_dataset = os.path.join("./data/processed_shapenetseg", self.dataset_string_args)
os.makedirs(os.path.dirname(self.path_dataset), exist_ok=True)
self.shuffle = shuffle
self.num_samples = num_samples
self.anisotropic_scaling = anisotropic_scaling
self.fine_tune = (phase == 'fine_tune_test')
if phase == 'fine_tune_test':
phase = 'ALLDATA'
self.knn = knn
self.num_neighbors = num_neighbors
self.normalization = normalization
self.random_translation = random_translation
self.data_augmentation_Z_rotation = data_augmentation_Z_rotation
self.data_augmentation_Z_rotation_range = data_augmentation_Z_rotation_range # range in degree of random rotation
self.data_augmentation_3D_rotation = data_augmentation_3D_rotation
self.sym_plane = ("xy", )
self.npoints = npoints
self.sample = sample
self.normal = normal
self.phase = phase
self.root = root_dir
self.datapath = [] # List to store all path of dataset files
self.catfile = os.path.join(self.root, 'synsetoffset2category.txt')
self.namecat2numbercat = {}
self.numbercat2namecat = {}
self.numsamples_by_cat = {}
self.class_choice = class_choice
self.meta = {}
self.data = []
self.use_fixed_pairs = use_fixed_pairs
# if self.shuffle and self.use_fixed_pairs:
# logger.warn("{}: shuffle and use_fixed_pairs is exclusive, shuffle to False".format(__name__))
# if not os.path.exists(self.root):
# print("Downloading Shapetnet for segmentation...")
# os.system('chmod +x ./data/download_dataset_shapenet.sh')
# os.system('./data/download_dataset_shapenet.sh')
# ----------------------------------------------------------#
## Create dictionaries with keys : name of class, value : name of folder and vice-versa
with open(self.catfile, 'r') as f:
for line in f:
ls = line.strip().split()
self.numsamples_by_cat[ls[0]] = 0
self.namecat2numbercat[ls[0]] = ls[1]
self.numbercat2namecat[ls[1]] = ls[0]
self.meta[ls[0]] = [] # List to store all path of files per category
# ----------------------------------------------------------#
# add all paths to the same list, keep track of sizes
for cat in self.meta.keys():
for file in self.meta[cat]:
self.datapath.append(file)
# fixed pairs? create or load a dict of source_target_pairs storing the cat/file_name: cat/file_name
if self.use_fixed_pairs and self.phase in ("val", "test", "svr_test"):
if os.path.exists(self.path_dataset + "_pairs.txt"):
source_target_file = np.loadtxt(self.path_dataset + "_pairs.txt", dtype=str)
self.source_target_pairs = source_target_file
self.datas = [(None, self.numbercat2namecat[source.split("/")[0]], os.path.join(self.root, source)) for source, _ in self.source_target_pairs]
self.datas += [(None, self.numbercat2namecat[source.split("/")[0]], os.path.join(self.root, source)) for _, source in self.source_target_pairs]
self.datapath = self.source_target_pairs
if not use_fixed_pairs or len(self.source_target_pairs)==0:
if self.phase == "train":
with open(os.path.join(os.path.join(self.root, "train_test_split"), 'shuffled_train_file_list.json')) as f:
file_list = json.load(f)
if self.phase == "val":
with open(os.path.join(os.path.join(self.root, "train_test_split"), 'shuffled_val_file_list.json')) as f:
file_list = json.load(f)
if "test" in self.phase :
with open(os.path.join(os.path.join(self.root, "train_test_split"), 'shuffled_test_file_list.json')) as f:
file_list = json.load(f)
for file in file_list:
# Typical example : shape_data/03001627/355fa0f35b61fdd7aa74a6b5ee13e775 so remove 'shape_data/' and add '.txt'
file_path = os.path.join(self.root, file[11:]) + ".txt"
number_category = file[11:19] + ""
if (class_choice is None) or (self.numbercat2namecat[number_category] in class_choice):
self.meta[self.numbercat2namecat[number_category]].append(
(file_path, self.numbercat2namecat[number_category]))
self.numsamples_by_cat[self.numbercat2namecat[number_category]] = self.numsamples_by_cat[
self.numbercat2namecat[
number_category]] + 1
# ----------------------------------------------------------#
# add all paths to the same list, keep track of sizes
for cat in self.meta.keys():
for file in self.meta[cat]:
self.datapath.append(file)
if self.use_preprocessed:
self.preprocess()
else:
self.datas = [(None, file[1], file[0]) for file in self.datapath]
if self.knn:
start = time.time()
logger.info(None,
"Computing nearest neighbors graph... (can take some time if it's not already precomputed.)")
self.compute_nearest_neighbors_graph()
logger.success('Done!')
end = time.time()
logger.info("Ellapsed time : " + '"%.2f' % (end - start))
else:
self.source_target_pairs = []
for i in range(len(self.datapath)):
_, cat, filename = self.datas[i]
filename = os.path.basename(filename)
if self.knn:
j = self.indices[i][np.random.randint(self.num_neighbors)]
else:
j = np.random.randint(len(self.datas))
_, cat2, filename2 = self.datas[j]
filename2 = os.path.basename(filename2)
self.source_target_pairs = (self.namecat2numbercat[cat]+"/"+filename, self.namecat2numbercat[cat2]+"/"+filename2)
with open(self.path_dataset + "_pairs.txt","w") as f:
for key, value in self.source_target_pairs:
f.write("{} {}\n".format(key, value))
# Shuffle_list
self.shuffle_list = [i for i in range(len(self))]
if self.shuffle:
np.random.shuffle(self.shuffle_list)
@staticmethod
def _uniformize_sizes(knn):
"""
input : knn :
-- list of tensors of size(npoints x 3)
N_point can be different for each sample
It is used to compute the KNN in the dataloader.
Since all pointcloud have a point (100, 100, 100) it doesn't change their chamfer distance and doesn't affect the nearest neighbors.
:param knn: list of tensors of size(npoints x 3)
:return: Return a tensor of size B, max_npoints, 3. The missing values are filled with 100.
"""
sizes = [data.size(0) for data in knn]
scale = [data.max() for data in knn]
scale = max(scale)
max_sizes = max(sizes) + 1
knn_new_list = []
for data in knn:
data = data.squeeze()
new_tensor = torch.zeros((max_sizes, 3)) + 100
new_tensor[:data.size(0)] = data[:, :3]
knn_new_list.append(new_tensor.unsqueeze(0).float())
return torch.cat(knn_new_list, 0)
def compute_nearest_neighbors_graph(self):
if not os.path.exists(self.path_dataset + "_knn_indices.npy"):
knn = [data[0] for data in self.datas]
knn = self._uniformize_sizes(knn)
self.num_neighbors = min(self.num_neighbors, len(self.datas))
nbrs = NearestNeighbors(n_neighbors=self.num_neighbors, algorithm='ball_tree', metric=_numpy_chamfer,
n_jobs=4).fit(knn.view(knn.size(0), -1).numpy())
distances, indices = nbrs.kneighbors(knn.view(knn.size(0), -1).numpy())
self.indices = indices
self.distances = distances
np.save(self.path_dataset + "_knn_indices.npy", self.indices)
np.save(self.path_dataset + "_knn_distances.npy", self.distances)
else:
self.indices = np.load(self.path_dataset + "_knn_indices.npy")
self.distances = np.load(self.path_dataset + "_knn_distances.npy")
def _getitem(self, index):
"""return points (N,3), category name string, path string"""
file = self.datapath[index]
try:
points = pc_utils.load(file[0])
except Exception as e:
logger.warn("Could not read {}".format(file[0]))
return None
points = torch.from_numpy(points).float()
# Normalization is done before resampling !
points[:, :3] = self.normalize(points[:, :3], self.isV2)
return points, file[1], file[0]
def preprocess(self):
start = time.time()
if os.path.exists(self.path_dataset + ".pkl"):
print("Reload dataset : ", self.path_dataset)
with open(self.path_dataset + ".pkl", "rb") as fp:
self.datas = pickle.load(fp)
else:
logger.info("preprocess dataset...")
class BatchCompletionCallBack(object):
completed = defaultdict(int)
def __init__(se, time, index, parallel):
se.index = index
se.parallel = parallel
def __call__(se, index):
BatchCompletionCallBack.completed[se.parallel] += 1
if BatchCompletionCallBack.completed[se.parallel] % 100 == 0:
end = time.time()
etl = (end - start) * (
len(self.datapath) / float(BatchCompletionCallBack.completed[se.parallel])) - (
end - start)
print('\r' + "Progress : %f %% " %
float(BatchCompletionCallBack.completed[
se.parallel] * 100 / len(self.datapath)) + "ETL %d seconds" % int(etl), end='')
if se.parallel._original_iterator is not None:
se.parallel.dispatch_next()
joblib.parallel.BatchCompletionCallBack = BatchCompletionCallBack
self.datas = joblib.Parallel(n_jobs=-1, backend="multiprocessing")(
joblib.delayed(_unwrap_self)(i) for i in zip([self] * len(self.datapath), range(len(self.datapath))))
self.datas = list(filter(None, self.datas))
with open(self.path_dataset + ".pkl", "wb") as fp: # Pickling
pickle.dump(self.datas, fp)
logger.info("dataset : " + str(len(self.datas)))
end = time.time()
logger.info("Ellapsed time : " + '"%.2f' % (end - start))
def generate_parts_by_cat(self):
parts_by_cat = {}
for cat in self.meta.keys():
labels = set([])
for file in self.meta[cat]:
points = np.loadtxt(file[0])
labels = labels | set(points[:, 6])
print(cat, labels)
parts_by_cat[cat] = list(labels)
return parts_by_cat
@staticmethod
def normalize(x, isV2=False):
v2_rot_mat = pc_utils.get_3D_rot_matrix(1, np.pi/2).astype(np.float32)
pc = pc_utils.normalize_to_box(x)[0]
# place on the "ground"
ground = pc[:,1].min()
pc[:,1] -= (ground+1.0)
if isV2:
x[:,:3] = x[:,:3].dot(v2_rot_mat)
if x.shape[1] > 3:
x[:,3:] = x[:,3:].dot(v2_rot_mat)
return pc
def getAnItem(self, index):
"""
returns:
points : float (N,6) positions and normals
cat : string category
file_path : string
"""
# ----------------------------------------------------------#
points, cat, file_path = self.datas[index]
file_name = os.path.splitext(os.path.basename(file_path))[0]
if points is None or not self.use_preprocessed:
found = glob(os.path.splitext(file_path)[0]+".*")
if len(found) == 0:
# use mesh
mesh_path = os.path.join(self.mesh_dir, self.namecat2numbercat[cat], file_name, "model.obj")
mesh = pymesh.load_mesh(mesh_path)
mesh = pymesh.split_long_edges(mesh, 0.02)[0]
points = mesh.vertices
else:
points = pc_utils.load(found[0])
points = self.normalize(points[:,:3])
points = torch.from_numpy(points).float()
points = points.clone()
# Resample
if self.sample:
choice = np.random.choice(points.size(0), self.npoints, replace=True)
points = points[choice, :]
rot_matrix = pc_utils.uniform_rotation_axis_matrix(axis=1, range_rot=self.data_augmentation_Z_rotation_range)
if self.data_augmentation_Z_rotation:
# Uniform random Rotation of axis Y
points, rot_matrix = pc_utils.uniform_rotation_axis(points, axis=1, normals=self.normal,
range_rot=self.data_augmentation_Z_rotation_range)
if self.anisotropic_scaling:
# Data augmentation : anisotropic scaling
points[:, :3] = pc_utils.anisotropic_scaling(points[:, :3]).contiguous()
points[:, :3] = self.normalize(points[:, :3])
if self.data_augmentation_3D_rotation:
# Uniform random 3D rotation of the sphere.
points, rot_matrix = pc_utils.uniform_rotation_sphere(points, normals=self.normal)
# Remark : if input of data_augmentation is normalized to unit ball and centered, the output rotated is as well.
if self.random_translation:
points = pc_utils.add_random_translation(points, scale=0.03)
# also load mesh during test
if self.phase == "test" and self.mesh_dir is not None:
mesh_path = os.path.join(self.mesh_dir, self.namecat2numbercat[cat], file_name, "model.obj")
V_mesh, F_mesh = read_trimesh(mesh_path)
V_mesh = V_mesh[:,:3]
F_mesh = F_mesh[:,:3]
# V_mesh = (V_mesh - center)/scale
V_mesh = self.normalize(V_mesh)
V_mesh = torch.from_numpy(V_mesh).to(dtype=torch.float)
F_mesh = torch.from_numpy(F_mesh).to(dtype=torch.int64)
return points, V_mesh, F_mesh, cat, file_name
else:
return points, cat, file_name
# ----------------------------------------------------------#
def __getitem__(self, index):
if self.use_fixed_pairs:
# _, cat, filename = self.datas[index]
# target = self.source_target_pairs[self.namecat2numbercat[cat]+'/'+os.path.basename(filename)]
# index_2 = [i for i, tupl in enumerate(self.datas) if ((self.namecat2numbercat[tupl[1]]+'/'+os.path.basename(tupl[2])) == target)]
# assert(len(index_2) == 1)
# index_2 = index_2[0]
index_2 = index + len(self.source_target_pairs)//2
else:
index = self.shuffle_list[index]
if self.knn:
index_2 = self.indices[index][np.random.randint(self.num_neighbors)]
else:
index_2 = np.random.randint(self.__len__())
source_data = self.getAnItem(index)
target_data = self.getAnItem(index_2)
if len(source_data)==5 and len(target_data) == 5:
return {"source_shape": source_data[0], "source_cat": source_data[3], "source_file": source_data[4], "source_mesh": source_data[1], "source_face": source_data[2],
"target_shape": target_data[0], "target_cat": target_data[3], "target_file": target_data[4], "target_mesh": target_data[1], "target_face": target_data[2]}
else:
return {"source_shape": source_data[0], "source_cat": source_data[1], "source_file": source_data[2],
"target_shape": target_data[0], "target_cat": target_data[1], "target_file": target_data[2]}
@staticmethod
def uncollate(batch_data):
source_shape, target_shape = batch_data["source_shape"], batch_data["target_shape"]
source_normals = source_shape[:,:,3:6].detach().contiguous().cuda()
target_normals = target_shape[:,:,3:6].detach().contiguous().cuda()
source_label = source_shape[:,:,-1].detach().contiguous().cuda()
target_label = target_shape[:,:,-1].detach().contiguous().cuda()
source_shape = source_shape[:,:,:3].detach().contiguous().cuda()
target_shape = target_shape[:,:,:3].detach().contiguous().cuda()
source_shape = source_shape.detach().cuda()
target_shape = target_shape.detach().cuda()
source_mesh = None if "source_mesh" not in batch_data else batch_data["source_mesh"].detach().cuda()
target_mesh = None if "target_mesh" not in batch_data else batch_data["target_mesh"].detach().cuda()
source_face = None if "source_face" not in batch_data else batch_data["source_face"].detach().cuda()
target_face = None if "target_face" not in batch_data else batch_data["target_face"].detach().cuda()
return {"source_shape": source_shape, "source_normals": source_normals, "source_label": source_label, "source_file": batch_data["source_file"],
"target_shape": target_shape, "target_normals": target_normals, "target_label": target_label, "target_file": batch_data["target_file"],
"source_mesh": source_mesh, "source_face": source_face, "target_mesh": target_mesh, "target_face": target_face}
def __len__(self):
if self.use_fixed_pairs and hasattr(self, "source_target_pairs"):
if self.num_samples > 0:
return min(self.num_samples, len(self.source_target_pairs))
return len(self.source_target_pairs)
if self.num_samples > 0:
return min(len(self.datas), self.num_samples)
else:
return len(self.datas)
@staticmethod
def render_result(shape_dir, **kwargs):
"""render shapes inside a directory with thea"""
forward = (-1,-0.4,-1)
pos = (4,1.4,4)
up = (-0.1,0.2,-0.1)
# if "color" not in kwargs:
# kwargs["color"] = "eebad6a9"
renderMeshes(shape_dir, up=up, forward=forward, pos=pos, **kwargs)
class ShapeNetV2(ShapeNetSeg):
def __init__(self, root_dir="/home/mnt/points/data/ShapeNet/ShapeNetCore.v2.5000p",
mesh_dir="/home/mnt/points/data/ShapeNet/ShapeNetCore.v2.remesh",
num_samples=-1, normalization="BoundingBox", knn=False,
num_neighbors=40,
class_choice=None, phase="train", normal=False, npoints=5000, sample=True,
data_augmentation_Z_rotation=False, data_augmentation_Z_rotation_range=360,
data_augmentation_3D_rotation=False, random_translation=False, anisotropic_scaling=False,
shuffle=False, use_fixed_pairs=False, **kwargs):
torch.utils.data.Dataset.__init__(self)
self.mesh_data = False
self.mesh_dir = mesh_dir
self.dataset_string_args = str(phase) + "_" + \
str(class_choice) + \
"_" + str(num_samples) + \
"_" + str(5000)
self.path_dataset = os.path.join("./data/processed_shapenetv2", self.dataset_string_args)
os.makedirs(os.path.dirname(self.path_dataset), exist_ok=True)
self.shuffle = shuffle
self.num_samples = num_samples
self.anisotropic_scaling = anisotropic_scaling
self.knn = knn
self.num_neighbors = num_neighbors
self.normalization = normalization
self.random_translation = random_translation
self.data_augmentation_Z_rotation = data_augmentation_Z_rotation
self.data_augmentation_Z_rotation_range = data_augmentation_Z_rotation_range # range in degree of random rotation
self.data_augmentation_3D_rotation = data_augmentation_3D_rotation
self.sym_plane = ("yz", )
self.npoints = npoints
self.sample = sample
self.normal = normal
self.phase = phase
self.root = root_dir
self.datapath = [] # List to store all path of dataset files
self.catfile = os.path.join(self.root, 'synsetoffset2category.txt')
self.namecat2numbercat = {}
self.numbercat2namecat = {}
self.numsamples_by_cat = {}
self.class_choice = class_choice
self.meta = {}
self.data = []
self.use_fixed_pairs = use_fixed_pairs
# if self.shuffle and self.use_fixed_pairs:
# logger.warn("{}: shuffle and use_fixed_pairs is exclusive, shuffle to False".format(__name__))
# if not os.path.exists(self.root):
# print("Downloading Shapetnet for segmentation...")
# os.system('chmod +x ./data/download_dataset_shapenet.sh')
# os.system('./data/download_dataset_shapenet.sh')
# ----------------------------------------------------------#
## Create dictionaries with keys : name of class, value : name of folder and vice-versa
with open(self.catfile, 'r') as f:
for line in f:
ls = line.strip().split()
self.numsamples_by_cat[ls[0]] = 0
self.namecat2numbercat[ls[0]] = ls[1]
self.numbercat2namecat[ls[1]] = ls[0]
self.meta[ls[0]] = [] # List to store all path of files per category
# ----------------------------------------------------------#
# create list of files datapaths
# cat/name 0/1 (0:train 1:test)
file_list = []
split_file = None
if class_choice is None:
split_files = glob(os.path.join(self.root, "*_shuffled_train_test_split.txt"))
else:
split_files = [os.path.join(self.root, "{}_shuffled_train_test_split.txt".format(synnum))
for synname, synnum in self.namecat2numbercat.items() if synname in class_choice]
for split_file in split_files:
if os.path.isfile(split_file):
phase_token = "0" if self.phase=="train" else "1"
with open(split_file, "r") as f:
for l in f:
if l[-2:-1] == phase_token:
file_list.append(l.split()[0])
else:
with open(split_file, "w") as f:
remove_str = len("/models/model_normalized.pts")
for k in self.numbercat2namecat:
all_files_in_cat = find_files(os.path.join(self.root, k), ["ply","pts"])
np.random.shuffle(all_files_in_cat)
num_test = max(int(0.05 * len(all_files_in_cat)), 10)
for i, l in enumerate(all_files_in_cat):
l = os.path.relpath(l[:-remove_str], self.root)
f.write(l+" "+str(int(i<num_test))+"\n")
if (self.phase == "train" and i >= num_test) or (self.phase=="test" and i<num_test):
file_list.append(l)
for file in file_list:
# Typical example : 03001627/355fa0f35b61fdd7aa74a6b5ee13e775 so remove 'shape_data/' and add '.txt'
file_path = glob(os.path.join(self.root, file, "models", "model_normalized.*"))[0]
number_category = file[:8]
if (class_choice is None) or (self.numbercat2namecat[number_category] in class_choice):
self.meta[self.numbercat2namecat[number_category]].append(
(file_path, self.numbercat2namecat[number_category]))
self.numsamples_by_cat[self.numbercat2namecat[number_category]] = self.numsamples_by_cat[
self.numbercat2namecat[
number_category]] + 1
# ----------------------------------------------------------#
# add all paths to the same list, keep track of sizes
for cat in self.meta.keys():
for file in self.meta[cat]:
self.datapath.append(file)
logger.info("size of " + self.phase + " dataset : " + str(len(self.datapath)))
self.preprocess()
if self.knn:
start = time.time()
logger.info(None,
"Computing nearest neighbors graph... (can take some time if it's not already precomputed.)")
self.compute_nearest_neighbors_graph()
logger.success('Done!')
end = time.time()
logger.info("Ellapsed time : " + '"%.2f' % (end - start))
# fixed pairs? create or load a dict of source_target_pairs storing the cat/file_name: cat/file_name
if self.use_fixed_pairs and self.phase in ("val", "test"):
if os.path.exists(self.path_dataset + "_pairs.txt"):
source_target_file = np.loadtxt(self.path_dataset + "_pairs.txt", dtype=str)
self.source_target_pairs = dict(source_target_file)
else:
self.source_target_pairs = dict()
for i in range(len(self.datas)):
_, cat, filename = self.datas[i]
remove_str = len("/models/model_normalized.pts")
filename = os.path.basename(filename[:-remove_str])
if self.knn:
j = self.indices[i][np.random.randint(self.num_neighbors)]
else:
j = np.random.randint(len(self.datas))
_, cat2, filename2 = self.datas[j]
filename2 = os.path.basename(filename2[:-remove_str])
self.source_target_pairs[self.namecat2numbercat[cat]+"/"+filename] = self.namecat2numbercat[cat2]+"/"+filename2
with open(self.path_dataset + "_pairs.txt","w") as f:
for key, value in self.source_target_pairs.items():
f.write("{} {}\n".format(key, value))
# Shuffle_list
self.len_data = len(self.datas)
self.shuffle_list = [i for i in range(self.len_data)]
if self.shuffle:
np.random.shuffle(self.shuffle_list)
@staticmethod
def normalize(x):
pc = pc_utils.normalize_to_box(x)[0]
return pc
def getAnItem(self, index):
"""
returns:
points : float (N,6) positions and normals
cat : string category
file_path : string
"""
# ----------------------------------------------------------#
points, cat, file_path = self.datas[index]
remove_str = len("/models/model_normalized.pts")
file_name = os.path.basename(file_path[:-remove_str])
points = points.clone()
# Resample
if self.sample:
choice = np.random.choice(points.size(0), self.npoints, replace=True)
points = points[choice, :]
rot_matrix = pc_utils.uniform_rotation_axis_matrix(axis=1, range_rot=self.data_augmentation_Z_rotation_range)
if self.anisotropic_scaling:
# Data augmentation : anisotropic scaling
points[:, :3] = pc_utils.anisotropic_scaling(points[:, :3]).contiguous()
points[:, :3] = self.normalize(points[:, :3])
# Remark : if input of data_augmentation is normalized to unit ball and centered, the output rotated is as well.
if self.random_translation:
points = pc_utils.add_random_translation(points, scale=0.03)
# also load mesh during test
if self.phase == "test" and self.mesh_dir is not None:
mesh_path = os.path.join(self.mesh_dir, self.namecat2numbercat[cat], file_name, "models", "model_normalized.obj")
V_mesh, F_mesh = read_trimesh(mesh_path)
V_mesh = V_mesh[:,:3]
F_mesh = F_mesh[:,:3]
# V_mesh = (V_mesh - center)/scale
V_mesh = self.normalize(V_mesh)
V_mesh = torch.from_numpy(V_mesh).to(dtype=torch.float)
F_mesh = torch.from_numpy(F_mesh).to(dtype=torch.int64)
return points, V_mesh, F_mesh, cat, file_name
else:
return points, cat, file_name
def __getitem__(self, index):
if self.use_fixed_pairs:
_, cat, filename = self.datas[index]
remove_str = len("/models/model_normalized.pts")
target = self.source_target_pairs[self.namecat2numbercat[cat]+'/'+os.path.basename(filename[:-remove_str])]
index_2 = [i for i, tupl in enumerate(self.datas) if ((self.namecat2numbercat[tupl[1]]+'/'+os.path.basename(tupl[2][:-remove_str])) == target)]
assert(len(index_2) == 1)
index_2 = index_2[0]
else:
index = self.shuffle_list[index]
if self.knn:
index_2 = self.indices[index][np.random.randint(self.num_neighbors)]
else:
index_2 = np.random.randint(self.__len__())
source_data = self.getAnItem(index)
target_data = self.getAnItem(index_2)
if len(source_data)==5 and len(target_data) == 5:
return {"source_shape": source_data[0], "source_cat": source_data[3], "source_file": source_data[4], "source_mesh": source_data[1], "source_face": source_data[2],
"target_shape": target_data[0], "target_cat": target_data[3], "target_file": target_data[4], "target_mesh": target_data[1], "target_face": target_data[2]}
else:
return {"source_shape": source_data[0], "source_cat": source_data[1], "source_file": source_data[2],
"target_shape": target_data[0], "target_cat": target_data[1], "target_file": target_data[2]}
@staticmethod
def render_result(shape_dir, **kwargs):
"""render shapes inside a directory with thea"""
forward = (-1,-0.65,1)
pos = (3.05,1.83,-3.05)
up = (0.1,0.4,-0.1)
renderMeshes(shape_dir, up=up, forward=forward, pos=pos, **kwargs)
class PairedSurreal(torch.utils.data.Dataset):
def __init__(self, root_dir, phase="train", npoints=6890, regular_sampling=False, normal=False, data_augmentation_Z_rotation=False,
data_augmentation_Z_rotation_range=360, data_augmentation_3D_rotation=False, max=-1,
template="./data/surreal_template_v77.ply", source="./data/surreal_template.ply"):
super().__init__()
self.data_augmentation_Z_rotation = data_augmentation_Z_rotation
self.data_augmentation_Z_rotation_range = data_augmentation_Z_rotation_range
self.data_augmentation_3D_rotation = data_augmentation_3D_rotation
self.sym_plane = None
self.normal = normal
self.phase = phase
self.train = self.phase == "train"
self.regular_sampling = regular_sampling # sample points uniformly or proportionaly to their adjacent area
self.npoints = npoints
template = template or "./data/surreal_template_v77.ply"
source = source or "./data/surreal_template.ply"
self.datas = []
start = time.time()
if self.phase == "train":
self.datas = torch.load(os.path.join(root_dir, "datas_surreal_train.pth"))
else:
self.datas = torch.load(os.path.join(root_dir, "datas_surreal_test.pth"))
end = time.time()
print("Ellapsed time to load dataset: ", end - start)
# template
# if not os.path.exists("./data/template/template.ply"):
# os.system("chmod +x ./data/download_template.sh")
# os.system("./data/download_template.sh")
self.mesh_data = True
if self.phase == "train":
self.len = len(self.datas) if max < 0 else max # self.datas.shape[0] 230000
else:
self.len = 100 if max < 0 else max
self.mesh_vertex, self.mesh_face = read_trimesh(source)
self.mesh_vertex = self.mesh_vertex[:,:3].astype(np.float32)
self.mesh_face = torch.from_numpy(self.mesh_face[:,:3].astype(np.int64))
self.mesh_vertex = torch.from_numpy(self.mesh_vertex)
V, center, scale = pc_utils.center_bounding_box(self.mesh_vertex)
self.mesh_vertex = V.squeeze(0)
center = center.squeeze(0)
scale = scale.squeeze(0)
self.cage_vertex, self.cage_face = read_trimesh(template, normal=True)
if "surreal_template_" in template :
self.cage_vertex = self.cage_vertex[:,:3].astype(np.float32) + 0.05*self.cage_vertex[:,3:6].astype(np.float32)
else:
self.cage_vertex = self.cage_vertex[:,:3].astype(np.float32)
self.cage_face = torch.from_numpy(self.cage_face[:,:3].astype(np.int64))
self.cage_vertex = torch.from_numpy(self.cage_vertex)
self.cage_vertex = (self.cage_vertex - center)
_, farea = compute_face_normals_and_areas(self.mesh_vertex.unsqueeze(0), self.mesh_face.unsqueeze(0))
v_area = scatter_add(farea.view(-1,1).expand(-1,3).contiguous().view(-1), self.mesh_face.view(-1), 0, out_size=(self.mesh_vertex.shape[0],))
self.prob = (v_area / torch.sum(v_area)).numpy()
def getPairIdx(self, index):
target_index = index
source_index = np.random.choice(self.len)
return (source_index, target_index)
def __getitem__(self, index):