forked from lxasqjc/Deformation-Segmentation
-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataset.py
1039 lines (933 loc) · 46.4 KB
/
dataset.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
import os, os.path, random
import json
import torch
from torch.nn import functional as F
from torchvision import transforms
import numpy as np
from PIL import Image
import cv2
import albumentations as A
def img_transform(img):
# 0-255 to 0-1
img = np.float32(np.array(img)) / 255.
img = img.transpose((2, 0, 1))
img = torch.from_numpy(img.copy())
return img
def imresize(im, size, interp='bilinear'):
if interp == 'nearest':
resample = Image.NEAREST
elif interp == 'bilinear':
resample = Image.BILINEAR
elif interp == 'bicubic':
resample = Image.BICUBIC
else:
raise Exception('resample method undefined!')
return im.resize(size, resample)
def b_imresize(im, size, interp='bilinear'):
return F.interpolate(im, size, mode=interp)
# from HRnet
def multi_scale_aug(image, label=None):
# print('image_shape: ', image.shape)
# print('label_shape: ', label.shape)
rand_scale = 0.5 + random.randint(0, 16) / 10.0
long_size = np.int(2048 * rand_scale + 0.5)
w, h = image.shape[-2:]
if h > w:
new_h = long_size
new_w = np.int(w * long_size / h + 0.5)
else:
new_w = long_size
new_h = np.int(h * long_size / w + 0.5)
image = F.interpolate(image, (new_w, new_h), mode='bilinear')
if label is not None:
label = F.interpolate(label.unsqueeze(1).float(), (new_w, new_h), mode='nearest').squeeze(1).long()
else:
return image
return image, label
def patch_loader(X, Y, xi, yi, cfg, train_mode=False, select_scale=None):
X_batch = X.clone()
Y_batch = Y.clone()
if select_scale is not None:
select_scale_batch = select_scale.clone()
len_batch = len(select_scale_batch)
else:
len_batch = X.size(0)
# print('---------select_scale_batch-----------\n', select_scale_batch)
for b in range(len_batch):
X = X_batch[b].unsqueeze(0)
Y = Y_batch[b].unsqueeze(0)
if select_scale is not None:
select_scale = int(select_scale_batch[b])
# print('---------select_scale-----------\n', select_scale)
fov_map_scale = cfg.MODEL.fov_map_scale
ori_w, ori_h = X.shape[-2:]
if cfg.DATASET.multi_scale_aug and train_mode:
X, Y = multi_scale_aug(X, Y)
ori_cx_lr = xi*fov_map_scale
ori_cy_lr = yi*fov_map_scale*cfg.MODEL.patch_ap
scaled_cx_lr = ori_cx_lr * (X.shape[-2]/ori_w)
scaled_cy_lr = ori_cy_lr * (X.shape[-1]/ori_h)
xi, yi = scaled_cx_lr//fov_map_scale, scaled_cy_lr//(fov_map_scale*cfg.MODEL.patch_ap)
# X: b,c,w,h
xi_ori, yi_ori = xi, yi
if train_mode == False:
patch_bank = list((float(cfg.VAL.expand_prediection_rate_patch)*np.array(cfg.MODEL.patch_bank)).astype(int))
else:
patch_bank = cfg.MODEL.patch_bank
segm_downsampling_rate = cfg.DATASET.segm_downsampling_rate
fov_padding = cfg.MODEL.fov_padding
# single_gpu_size = str(torch.cuda.get_device_properties('cuda:0')).split('total_memory=')[1].split('MB')[0]
single_gpu_size = 10000
if int(single_gpu_size) < 1:
fov_padding_cpu = True
else:
fov_padding_cpu = False
if 'CITYSCAPES' in cfg.DATASET.root_dataset or 'CITYSCAPE' in cfg.DATASET.list_train:
ignore_label=20-1
elif 'Digest' in cfg.DATASET.root_dataset:
ignore_label=-2
else:
if cfg.DATASET.ignore_index != -2:
ignore_label=cfg.DATASET.ignore_index
else:
ignore_label=-2
X_patches = []
X_patches_cords = []
for s in range(len(patch_bank)):
if cfg.VAL.F_Xlr_only and s>0:
X_patches.append(X_patches[0])
continue
if select_scale != None and s != select_scale and s != 0:
continue
patch_size = patch_bank[s]
patch_size_x = patch_size
patch_size_y = patch_size*cfg.MODEL.patch_ap
# TODO: debug, current adjusting xi, yi approach only available when X.shape can devide by fov_map_scale
if cfg.DATASET.adjust_crop_range:
xi = int(xi_ori*(X.shape[-2]-patch_size_x)/X.shape[-2])
yi = int(yi_ori*(X.shape[-1]-patch_size_y)/X.shape[-1])
# correction on residual
if xi_ori >= round(ori_w/fov_map_scale)-1:
xi += 1
if yi_ori >= round(ori_h/(fov_map_scale*cfg.MODEL.patch_ap))-1:
yi += 1
# if X.shape[2] < patch_size_x or X.shape[3] < patch_size_y:
# raise Exception('Patch size {}x{} exceed image size {}'.format(patch_size_x, patch_size_y, X.shape))
if fov_padding:
if cfg.DATASET.adjust_crop_range:
p_h = max(patch_size_x-X.shape[2], 0)
p_w = max(patch_size_y-X.shape[3], 0)
# p = max(X_p_x, X_p_y)
p_y_h = max(patch_bank[0]-X.shape[2], 0)
p_y_w = max(patch_bank[0]*cfg.MODEL.patch_ap-X.shape[3], 0)
# p_y = max(Y_p_x, Y_p_y)
else:
p_w = patch_size_y-1
p_h = patch_size_x-1
p_y_w = patch_bank[0]*cfg.MODEL.patch_ap-1
p_y_h = patch_bank[0]-1
if cfg.DATASET.mirror_padding:
if fov_padding_cpu:
X_pad = F.pad(X, (p_w,p_w,p_h,p_h), mode='reflect').cpu()
else:
X_pad = F.pad(X, (p_w,p_w,p_h,p_h), mode='reflect')
if train_mode and segm_downsampling_rate != 1:
if fov_padding_cpu:
Y_pad = F.pad(Y.unsqueeze(1).float(), (p_y_w//segm_downsampling_rate,p_y_w//segm_downsampling_rate,p_y_h//segm_downsampling_rate,p_y_h//segm_downsampling_rate), mode='reflect').cpu()
else:
Y_pad = F.pad(Y.unsqueeze(1).float(), (p_y_w//segm_downsampling_rate,p_y_w//segm_downsampling_rate,p_y_h//segm_downsampling_rate,p_y_h//segm_downsampling_rate), mode='reflect')
else:
if fov_padding_cpu:
Y_pad = F.pad(Y.unsqueeze(1).float(), (p_y_w,p_y_w,p_y_h,p_y_h), mode='reflect').cpu()
else:
Y_pad = F.pad(Y.unsqueeze(1).float(), (p_y_w,p_y_w,p_y_h,p_y_h), mode='reflect')
Y_pad = Y_pad.squeeze(1).long()
else:
if fov_padding_cpu:
X_pad = F.pad(X, (p_w,p_w,p_h,p_h)).cpu()
else:
X_pad = F.pad(X, (p_w,p_w,p_h,p_h))
if train_mode and segm_downsampling_rate != 1:
if fov_padding_cpu:
Y_pad = F.pad(Y, (p_y_w//segm_downsampling_rate,p_y_w//segm_downsampling_rate,p_y_h//segm_downsampling_rate,p_y_h//segm_downsampling_rate), value=ignore_label).cpu()
else:
Y_pad = F.pad(Y, (p_y_w//segm_downsampling_rate,p_y_w//segm_downsampling_rate,p_y_h//segm_downsampling_rate,p_y_h//segm_downsampling_rate), value=ignore_label)
else:
if fov_padding_cpu:
Y_pad = F.pad(Y, (p_y_w,p_y_w,p_y_h,p_y_h), value=ignore_label).cpu()
else:
Y_pad = F.pad(Y, (p_y_w,p_y_w,p_y_h,p_y_h), value=ignore_label)
cx_lr = xi*fov_map_scale # upper left corner of current X_lr pixel
if cfg.DATASET.adjust_crop_range:
if cx_lr > (X.shape[2]-patch_size_x):
cx_lr = X.shape[2]-patch_size_x
cx = cx_lr + patch_bank[0]//2 - patch_size_x//2 # upper left corner of current patch size with same center of X_lr pixel
if cfg.DATASET.adjust_crop_range:
if cx < 0:
cx = 0
if fov_padding:
# cx_lr = cx_lr+p_y
cx_p = cx+p_h
if cfg.DATASET.multi_scale_aug:
cx_p_y = cx+p_y_h
else:
cx_p_y = cx_lr+p_y_h
if cfg.DATASET.multi_scale_aug and train_mode:
if cx_p < 0:
cx_p = 0
elif cx_p > (X.shape[2]-patch_size_x):
cx_p = X.shape[2]-patch_size_x
if cx_p_y < 0:
cx_p_y = 0
elif cx_p_y > (X.shape[2]-patch_size_x):
cx_p_y = X.shape[2]-patch_size_x
if cfg.DATASET.adjust_crop_range:
if cx_p_y < 0:
cx_p_y = 0
elif cx_p_y > (X.shape[2]-patch_size_x):
cx_p_y = X.shape[2]-patch_size_x
else:
if cx < 0:
cx = 0
elif cx > (X.shape[2]-patch_size_x):
cx = X.shape[2]-patch_size_x
cy_lr = yi*(fov_map_scale*cfg.MODEL.patch_ap) # upper left corner of current X_lr pixel
if cfg.DATASET.adjust_crop_range:
if cy_lr > (X.shape[3]-patch_size_y):
cy_lr = X.shape[3]-patch_size_y
cy = cy_lr + (patch_bank[0]*cfg.MODEL.patch_ap)//2 - patch_size_y//2 # upper left corner of current patch size with same center of X_lr pixel
if cfg.DATASET.adjust_crop_range:
if cy < 0:
cy = 0
if fov_padding:
# cy_lr = cy_lr+p_y
cy_p = cy+p_w
if cfg.DATASET.multi_scale_aug:
cy_p_y = cy+p_y_w
else:
cy_p_y = cy_lr+p_y_w
if cfg.DATASET.multi_scale_aug and train_mode:
if cy_p < 0:
cy_p = 0
elif cy_p > (X.shape[3]-patch_size_y):
cy_p = X.shape[3]-patch_size_y
if cy_p_y < 0:
cy_p_y = 0
elif cy_p_y > (X.shape[3]-patch_size_y):
cy_p_y = X.shape[3]-patch_size_y
if cfg.DATASET.adjust_crop_range:
if cy_p_y < 0:
cy_p_y = 0
elif cy_p_y > (X.shape[3]-patch_size_y):
cy_p_y = X.shape[3]-patch_size_y
else:
if cy < 0:
cy = 0
elif cy > (X.shape[3]-patch_size_y):
cy = X.shape[3]-patch_size_y
if fov_padding:
crop_patch = X_pad[:, :, cx_p:cx_p+patch_size_x, cy_p:cy_p+patch_size_y].to(X.device)
else:
crop_patch = X[:, :, cx:cx+patch_size_x, cy:cy+patch_size_y]
assert min(crop_patch.shape)!=0, "crop_patch size wrong ({}) cropped from X_pad ({}), X({}) at (cx_p={},cy_p={},xi={},yi={}, patch_size_x={}, patch_size_y={})".format(crop_patch.shape, X_pad.shape, X.shape, cx_p, cy_p, xi, yi, patch_size_x, patch_size_y)
if not (select_scale != None and select_scale != 0 and s == 0):
X_patches.append(b_imresize(crop_patch, (patch_bank[0],patch_bank[0]*cfg.MODEL.patch_ap), interp='bilinear'))
if not train_mode:
if not (select_scale != None and select_scale != 0 and s == 0):
X_patches_cords.append((cx, cy, patch_size, p_w, p_h)) # TODO: patch_size to be corrected in visualisation in eval/eval_multipro
if s == 0:
if segm_downsampling_rate != 1 and train_mode:
patch_size = patch_bank[0] // segm_downsampling_rate
cx_p_y = cx_p_y // segm_downsampling_rate
cy_p_y = cy_p_y // segm_downsampling_rate
else:
patch_size = patch_bank[0]
patch_size_x = patch_size
patch_size_y = patch_size*cfg.MODEL.patch_ap
if not train_mode:
Y_patch_cord = (cx_p_y, cy_p_y, patch_size, p_y_w, p_y_h)
# Y.shape b,w,h, NOTE Y has different size in train and val mode
if fov_padding:
Y_patch = Y_pad[:, cx_p_y:cx_p_y+patch_size_x, cy_p_y:cy_p_y+patch_size_y].to(Y.device)
else:
Y_patch = Y[:, cx_lr:cx_lr+patch_size_x, cy_lr:cy_lr+patch_size_y]
if b == 0:
X_patches_batch = X_patches
Y_patch_batch = Y_patch
else:
for p in range(len(X_patches_batch)):
X_patches_batch[p] = torch.cat([X_patches_batch[p], X_patches[p]])
Y_patch_batch = torch.cat([Y_patch_batch, Y_patch])
if train_mode:
return X_patches_batch, Y_patch_batch
else:
# print('Y_patch_cord: ', Y_patch_cord)
return X_patches_batch, Y_patch_cord, X_patches_cords, Y_patch_batch
class BaseDataset(torch.utils.data.Dataset):
def __init__(self, odgt, opt, **kwargs):
# NOTE: dataset classes SHOULD be mapped with minimum starting from [1:] NOT [0:]
# remapping labels reflecting servie degree of GS
if opt.root_dataset == '/scratch0/chenjin/GLEASON2019_DATA/Data/' or \
opt.root_dataset == '/home/chenjin/Chen_UCL/Histo-MRI-mapping/GLEASON2019_DATA/Data/' or \
opt.root_dataset == '/SAN/medic/Histo_MRI_GPU/chenjin/Data/GLEASON2019_DATA/Data/' or \
'GLEASON2019_DATA' in opt.root_dataset:
# four class mapping
if opt.class_mapping == 0:
self.label_mapping = {0: 1,
1: 1, 2: 1,
3: 2, 4: 3,
5: 4, 6: 1,
}
# three class mapping exclude class5
elif opt.class_mapping == 30:
self.label_mapping = {0: 1,
1: 1, 2: 1,
3: 2, 4: 3,
5: 1, 6: 1,
}
# gs3 vs all
elif opt.class_mapping == 3:
self.label_mapping = {0: 1,
1: 1, 2: 1,
3: 2, 4: 1,
5: 1, 6: 1,
}
# gs4 vs all
elif opt.class_mapping == 4:
self.label_mapping = {0: 1,
1: 1, 2: 1,
3: 1, 4: 2,
5: 1, 6: 1,
}
# gs5 vs all
elif opt.class_mapping == 5:
self.label_mapping = {0: 1,
1: 1, 2: 1,
3: 1, 4: 1,
5: 2, 6: 1,
}
# benine vs all
elif opt.class_mapping == 6:
self.label_mapping = {0: 1,
1: 1, 2: 1,
3: 2, 4: 2,
5: 2, 6: 1,
}
# mean and std
self.normalize = transforms.Normalize(
# gleason2019 322 train mean and std applied
mean=[0.748, 0.611, 0.823],
std=[0.146, 0.245, 0.119])
elif opt.root_dataset == '/home/chenjin/Chen_UCL/Histo-MRI-mapping/DigestPath2019/' or 'Digest' in opt.list_train:
self.label_mapping = {0: 1,
255: 2,
}
# mean and std
self.normalize = transforms.Normalize(
# gleason2019 322 train mean and std applied
mean=[0.816, 0.697, 0.792],
std=[0.160, 0.277, 0.198])
elif 'ADE20K' in opt.root_dataset or 'ADE' in opt.list_train:
self.label_mapping = {}
# mean and std
self.normalize = transforms.Normalize(
# gleason2019 322 train mean and std applied
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
elif 'CITYSCAPES' in opt.root_dataset or 'CITYSCAPE' in opt.list_train:
# following HRNet-Semantic-Segmentation setting
# but starting from 1 instead of 0, seems 0 leads to bug in criterion.OhemCrossEntropy implementation
# debug note 24/12/19 seems label must start from 1 and must be continues, otherwise lead inconsistence between pred by view(-1) and seg_label
ignore_label=20
if opt.binary_class != -1:
self.label_mapping = {}
for i in range(-1,34):
if i in (-1,0,1,2,3,4,5,6,9,10,14,15,16,18,29,30):
self.label_mapping[i] = ignore_label
if i == opt.binary_class:
self.label_mapping[i] = 2
else:
self.label_mapping[i] = 1
else:
self.label_mapping = {-1: ignore_label, 0: ignore_label,
1: ignore_label, 2: ignore_label,
3: ignore_label, 4: ignore_label,
5: ignore_label, 6: ignore_label,
7: 1, 8: 2, 9: ignore_label,
10: ignore_label, 11: 3, 12: 4,
13: 5, 14: ignore_label, 15: ignore_label,
16: ignore_label, 17: 6, 18: ignore_label,
19: 7, 20: 8, 21: 9, 22: 10, 23: 11, 24: 12,
25: 13, 26: 14, 27: 15, 28: 16,
29: ignore_label, 30: ignore_label,
31: 17, 32: 18, 33: 19}
# mean and std
self.normalize = transforms.Normalize(
# gleason2019 322 train mean and std applied
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
elif 'Histo' in opt.root_dataset or 'histomri' in opt.list_train or 'histomri' in opt.root_dataset:
self.label_mapping = {}
# mean and std
self.normalize = transforms.Normalize(
# gleason2019 322 train mean and std applied
mean=[0.8223, 0.7783, 0.7847],
std=[0.210, 0.216, 0.241])
elif 'DeepGlob' in opt.root_dataset or 'DeepGlob' in opt.list_train:
# ignore_label=7
if opt.ignore_index == 0:
self.label_mapping = {0: 2,
1: 3, 2: 4,
3: 5, 4: 6,
5: 7, 6: 1,
}
elif opt.ignore_index == 6:
self.label_mapping = {0: 1,
1: 2, 2: 3,
3: 4, 4: 5,
5: 6, 6: 7,
}
# mean and std
self.normalize = transforms.Normalize(
# gleason2019 322 train mean and std applied
mean=[0.282, 0.379, 0.408],
std=[0.089, 0.101, 0.127])
elif 'Face_single_example' in opt.root_dataset or 'Face_single_example' in opt.list_train:
# ignore_label=7
self.label_mapping = {201: 1,
184: 2, 173: 3,
191: 4, 194: 5
}
# mean and std
self.normalize = transforms.Normalize(
# gleason2019 322 train mean and std applied
mean=[0.282, 0.379, 0.408],
std=[0.089, 0.101, 0.127])
else:
raise Exception('Unknown root for mapping and normalisation!')
# parse options
self.imgSizes = opt.imgSizes
self.imgMaxSize = opt.imgMaxSize
# max down sampling rate of network to avoid rounding during conv or pooling
self.padding_constant = opt.padding_constant
# parse the input list
self.parse_input_list(odgt, **kwargs)
def convert_label(self, label, inverse=False):
label = np.array(label)
temp = label.copy()
if inverse:
for v, k in self.label_mapping.items():
label[temp == k] = v
else:
for k, v in self.label_mapping.items():
label[temp == k] = v
return label
def convert_label_single_face(self, label, inverse=False):
label = np.array(label)
temp = label.copy()
if inverse:
pass
# for v, k in self.label_mapping.items():
# label[temp == k] = v
else:
print('max label before mapping: {}'.format(np.max(label)))
for k, v in self.label_mapping.items():
label[temp == k] = v
label[label == temp] = 1
print('max label after mapping: {}'.format(np.max(label)))
# label_base = np.zeros(label.shape)
# label = np.maximum(label, label_base)
return label
def parse_input_list(self, odgt, max_sample=-1, start_idx=-1, end_idx=-1):
if isinstance(odgt, list):
self.list_sample = odgt
elif isinstance(odgt, str):
self.list_sample = [json.loads(x.rstrip()) for x in open(odgt, 'r')]
if max_sample > 0:
self.list_sample = self.list_sample[0:max_sample]
if start_idx >= 0 and end_idx >= 0: # divide file list
self.list_sample = self.list_sample[start_idx:end_idx]
self.num_sample = len(self.list_sample)
assert self.num_sample > 0
print('# samples: {}'.format(self.num_sample))
def img_transform(self, img):
# 0-255 to 0-1
img = np.float32(np.array(img)) / 255.
img = img.transpose((2, 0, 1))
img = self.normalize(torch.from_numpy(img.copy()))
return img
def img_transform_unnorm(self, img):
# 0-255 to 0-1
img = np.float32(np.array(img)) / 255.
img = img.transpose((2, 0, 1))
img = torch.from_numpy(img.copy())
return img
def img_transform_rev(self, img):
# 0-255 to 0-1
img = np.float32(np.array(img)) / 255.
img = img.transpose((2, 0, 1))
img = torch.from_numpy(img.copy())
return img
def segm_transform(self, segm):
# to tensor, -1 to 149
# !!!!! JC: This is why all data need to mapped to 1-numClass
# and because of this, ignore_index (in CrossEntropy/OhemCrossEntropy/IoU) = ignore_label (in dataset class_mapping)-1
segm = torch.from_numpy(np.array(segm)).long() - 1
return segm
# Round x to the nearest multiple of p and x' >= x
def round2nearest_multiple(self, x, p):
return ((x - 1) // p + 1) * p
class TrainDataset(BaseDataset):
def __init__(self, root_dataset, odgt, opt, batch_per_gpu=1, cal_REV=False, **kwargs):
super(TrainDataset, self).__init__(odgt, opt, **kwargs)
self.root_dataset = root_dataset
# down sampling rate of segm labe
self.segm_downsampling_rate = opt.segm_downsampling_rate
self.batch_per_gpu = batch_per_gpu
# classify images into two classes: 1. h > w and 2. h <= w
self.batch_record_list = [[], []]
# override dataset length when trainig with batch_per_gpu > 1
self.cur_idx = 0
self.if_shuffled = False
# augmentation
self.augmentation = opt.random_flip
self.balance_sam_idx = 0
self.num_class = opt.num_class
self.cal_REV = cal_REV
self.opt = opt
if self.opt.repeat_sample != 0:
self.repeat_sample_count = 0
def _get_sub_batch(self):
while True:
# get a sample record
this_sample = self.list_sample[self.cur_idx]
# when gt_gradient, make sure loaded sample got enough foreground pixels
if self.opt.gt_gradient_rm_under_repre > 0 and self.opt.num_class == 2:
search_well_repre_class = True
s_idx = self.cur_idx
while search_well_repre_class:
search_sample = self.list_sample[s_idx]
s_idx += 1
if s_idx >= self.num_sample:
s_idx = 0
segm_path = os.path.join(self.root_dataset, search_sample['fpath_segm'])
segm = self.convert_label(Image.open(segm_path))
segm -= 1
hist, _ = np.histogram(segm, bins=self.num_class, range=(0, self.num_class-1))
if (hist[-1] / np.sum(hist)) > self.opt.gt_gradient_rm_under_repre:
this_sample = search_sample
search_well_repre_class = False
self.cur_idx = s_idx
if self.augmentation == 'balance_sample' and self.balance_sam_idx > 2:
# search gs-5 and reset idx every 3 steps represent
# severe rare gs-4 in contrast to balanced other 3 classes
search_rare_class = True
s_idx = self.cur_idx
while search_rare_class:
search_sample = self.list_sample[s_idx]
s_idx += 1
if s_idx >= self.num_sample:
s_idx = 0
segm_path = os.path.join(self.root_dataset, search_sample['fpath_segm'])
segm = self.convert_label(Image.open(segm_path))
hist, _ = np.histogram(segm, bins=self.num_class, range=(0, self.num_class-1))
if (hist[-1] / np.sum(hist)) > 0.25:
this_sample = search_sample
search_rare_class = False
self.balance_sam_idx = 0
self.balance_sam_idx += 1
# print('this sample fpath_segm: {}'.format(this_sample['fpath_segm']))
if this_sample['height'] > this_sample['width']:
self.batch_record_list[0].append(this_sample) # h > w, go to 1st class
else:
self.batch_record_list[1].append(this_sample) # h <= w, go to 2nd class
# update current sample pointer
if self.opt.repeat_sample != 0:
self.repeat_sample_count += 1
if self.repeat_sample_count % self.opt.repeat_sample == 0:
self.cur_idx += 1
self.repeat_sample_count = 0
else:
self.cur_idx += 1
if self.cur_idx >= self.num_sample:
self.cur_idx = 0
np.random.shuffle(self.list_sample)
if len(self.batch_record_list[0]) == self.batch_per_gpu:
batch_records = self.batch_record_list[0]
self.batch_record_list[0] = []
break
elif len(self.batch_record_list[1]) == self.batch_per_gpu:
batch_records = self.batch_record_list[1]
self.batch_record_list[1] = []
break
return batch_records
def __getitem__(self, index):
# NOTE: random shuffle for the first time. shuffle in __init__ is useless
if not self.if_shuffled and self.opt.shuffle_list:
np.random.seed(index)
np.random.shuffle(self.list_sample)
self.if_shuffled = True
# get sub-batch candidates
batch_records = self._get_sub_batch()
# resize all images' short edges to the chosen size
if isinstance(self.imgSizes, list) or isinstance(self.imgSizes, tuple):
this_short_size = np.random.choice(self.imgSizes)
else:
this_short_size = self.imgSizes
# calculate the BATCH's height and width
# since we concat more than one samples, the batch's h and w shall be larger than EACH sample
batch_widths = np.zeros(self.batch_per_gpu, np.int32)
batch_heights = np.zeros(self.batch_per_gpu, np.int32)
for i in range(self.batch_per_gpu):
img_height, img_width = batch_records[i]['height'], batch_records[i]['width']
# print('batch_records_height: {}, batch_records_width: {}'.format(batch_records[i]['height'], batch_records[i]['width']))
if self.imgMaxSize == 1:
# discard rescalel, keep original image size
this_scale = 1
else:
this_scale = min(
this_short_size / min(img_height, img_width), \
self.imgMaxSize / max(img_height, img_width))
batch_widths[i] = img_width * this_scale
batch_heights[i] = img_height * this_scale
# print('this_scale: {}'.format(this_scale))
# print('batch_widths[i]: {}, batch_heights[i]: {}'.format(batch_widths[i], batch_heights[i]))
# Here we must pad both input image and segmentation map to size h' and w' so that p | h' and p | w'
batch_width = np.max(batch_widths)
batch_height = np.max(batch_heights)
batch_width = int(self.round2nearest_multiple(batch_width, self.padding_constant))
batch_height = int(self.round2nearest_multiple(batch_height, self.padding_constant))
assert self.padding_constant >= self.segm_downsampling_rate, \
'padding constant must be equal or large than segm downsamping rate'
batch_images = torch.zeros(
self.batch_per_gpu, 3, batch_height, batch_width)
batch_segms = torch.zeros(
self.batch_per_gpu,
batch_height // self.segm_downsampling_rate,
batch_width // self.segm_downsampling_rate).long()
for i in range(self.batch_per_gpu):
this_record = batch_records[i]
# load image and label
image_path = os.path.join(self.root_dataset, this_record['fpath_img'])
segm_path = os.path.join(self.root_dataset, this_record['fpath_segm'])
# skip non exitst Training
if not os.path.isfile(segm_path):
continue
img = Image.open(image_path).convert('RGB')
segm = Image.open(segm_path)
if self.opt.check_dataload and i == 0:
img_ori = self.img_transform(img)
assert(segm.mode == "L")
assert(img.size[0] == segm.size[0])
assert(img.size[1] == segm.size[1])
# print(img.size)
# random_flip
if self.augmentation == 'Flip':
if np.random.choice([0, 1]):
img = img.transpose(Image.FLIP_LEFT_RIGHT)
segm = segm.transpose(Image.FLIP_LEFT_RIGHT)
elif self.augmentation == 'Flip_Pixel':
aug = A.Compose([
A.GaussNoise(),
A.RandomBrightnessContrast(),
A.Flip()
],p=1)
img = np.array(img)
segm = np.array(segm)
augmented = aug(image=img, mask=segm)
img = Image.fromarray(augmented['image'])
segm = Image.fromarray(augmented['mask'])
elif self.augmentation == 'cityHRaug':
aug = A.Compose([
A.RandomScale(scale_limit=[0.5,2.0]),
A.RandomCrop(512, 1024),
A.HorizontalFlip()
],p=1)
img = np.array(img)
segm = np.array(segm)
augmented = aug(image=img, mask=segm)
img = Image.fromarray(augmented['image'])
segm = Image.fromarray(augmented['mask'])
elif self.augmentation == 'balance_sample' and (i+1) % 4 == 0:
aug = A.Compose([
A.RandomCrop(self.imgSizes[0], self.imgSizes[1]),
A.Flip()
],p=1)
img = np.array(img)
segm = np.array(segm)
search_rare = True
while search_rare:
augmented = aug(image=img, mask=segm)
segm_s = self.convert_label(augmented['mask'])
hist, _ = np.histogram(segm_s, bins=self.num_class, range=(0, self.num_class-1))
if (hist[-1] / np.sum(hist)) > 0.25:
img = Image.fromarray(augmented['image'])
segm = Image.fromarray(augmented['mask'])
search_rare = False
elif self.augmentation == 'fullFoV_balance_sample' and (i+1) % 4 == 0:
img = imresize(img, (batch_widths[i], batch_heights[i]), interp='bilinear')
segm = imresize(segm, (batch_widths[i], batch_heights[i]), interp='nearest')
aug = A.Compose([
A.RandomCrop(batch_widths[i], batch_heights[i]),
A.Flip()
],p=1)
img = np.array(img)
segm = np.array(segm)
search_rare = True
while search_rare:
augmented = aug(image=img, mask=segm)
segm_s = self.convert_label(augmented['mask'])
hist, _ = np.histogram(segm_s, bins=self.num_class, range=(0, self.num_class-1))
if (hist[-1] / np.sum(hist)) > 0.25:
img = Image.fromarray(augmented['image'])
segm = Image.fromarray(augmented['mask'])
search_rare = False
elif self.augmentation.split("_")[0] == 'Crop':
if self.augmentation == 'Crop_Flip':
crop_w = self.imgSizes[0]
crop_h = self.imgSizes[1]
else:
crop = int(self.augmentation.split("_")[-1])
if img.size[0] < crop or img.size[1] < crop:
crop_h, crop_w = img.size[0], img.size[1]
else:
crop_w, crop_h = crop, crop
if self.augmentation.split("_")[1] == 'aug':
aug = A.Compose([
# A.RandomSizedCrop((1250, 2500), 2500, 2500),
# A.ShiftScaleRotate(),
A.RandomCrop(crop_w, crop_h),
A.RandomBrightnessContrast(),
A.MultiplicativeNoise(),
A.Flip()
# A.RGBShift(),
# A.Blur(),
# A.GaussNoise(),
# A.ElasticTransform(),
# A.Cutout(p=1)
],p=1)
else:
aug = A.Compose([
# A.RandomSizedCrop((1250, 2500), 2500, 2500),
# A.ShiftScaleRotate(),
A.RandomCrop(crop_w, crop_h),
A.Flip()
# A.RGBShift(),
# A.Blur(),
# A.GaussNoise(),
# A.ElasticTransform(),
# A.Cutout(p=1)
],p=1)
img = np.array(img)
segm = np.array(segm)
augmented = aug(image=img, mask=segm)
img = Image.fromarray(augmented['image'])
segm = Image.fromarray(augmented['mask'])
# print(img.size)
# note that each sample within a mini batch has different scale param
img = imresize(img, (batch_widths[i], batch_heights[i]), interp='bilinear')
segm = imresize(segm, (batch_widths[i], batch_heights[i]), interp='nearest')
# further downsample seg label, need to avoid seg label misalignment
segm_rounded_width = self.round2nearest_multiple(segm.size[0], self.segm_downsampling_rate)
segm_rounded_height = self.round2nearest_multiple(segm.size[1], self.segm_downsampling_rate)
segm_rounded = Image.new('L', (segm_rounded_width, segm_rounded_height), 0)
segm_rounded.paste(segm, (0, 0))
segm = imresize(
segm_rounded,
(segm_rounded.size[0] // self.segm_downsampling_rate, \
segm_rounded.size[1] // self.segm_downsampling_rate), \
interp='nearest')
# image transform, to torch float tensor CxHxW
if self.cal_REV:
img = self.img_transform_rev(img)
else:
img = self.img_transform(img)
# JC: re-ordering label according to servie degree
if 'Face_single_example' in self.opt.root_dataset or 'Face_single_example' in self.opt.list_train:
segm = self.convert_label_single_face(segm)
else:
segm = self.convert_label(segm)
# JC: for dataset like gleason2019 that background labelled 0 need to be considered
if ('GLEASON' in self.root_dataset and np.min(segm) == 0) or 'histomri' in self.opt.list_train or 'histomri' in self.opt.root_dataset:
segm = Image.fromarray(np.add(segm, 1))
# segm transform, to torch long tensor HxW
segm = self.segm_transform(segm)
# put into batch arrays
# print('train batch_image_{}_size: {}\n'.format(i, img.shape))
# print('train batch_segm_{}_size: {}\n'.format(i, segm.shape))
batch_images[i][:, :img.shape[1], :img.shape[2]] = img
batch_segms[i][:segm.shape[0], :segm.shape[1]] = segm
output = dict()
output['img_data'] = batch_images
output['seg_label'] = batch_segms # torch.nn.functional.one_hot(batch_segms, num_classes=7)
if self.opt.check_dataload:
output['img_ori'] = img_ori
# print('seg_label shape: {}'.format(output['seg_label'].shape))
# print('seg_label size: {}'.format(output['seg_label'].shape))
if self.cal_REV:
return batch_images
else:
return output
def __len__(self):
return int(1e10) # It's a fake length due to the trick that every loader maintains its own list
#return self.num_sampleclass
class ValDataset(BaseDataset):
def __init__(self, root_dataset, odgt, opt, cfg, **kwargs):
super(ValDataset, self).__init__(odgt, opt, **kwargs)
self.root_dataset = root_dataset
self.cfg = cfg
if cfg.VAL.expand_prediection_rate != 1:
self.imgSizes_val = tuple(cfg.VAL.expand_prediection_rate*np.array(self.imgSizes))
self.imgMaxSize_val = cfg.VAL.expand_prediection_rate*self.imgMaxSize
else:
self.imgSizes_val = self.imgSizes
self.imgMaxSize_val = self.imgMaxSize
def __getitem__(self, index):
this_record = self.list_sample[index]
# load image and label
image_path = os.path.join(self.root_dataset, this_record['fpath_img'])
segm_path = os.path.join(self.root_dataset, this_record['fpath_segm'])
img = Image.open(image_path).convert('RGB')
segm = Image.open(segm_path)
if self.cfg.DATASET.val_central_crop:
width, height = img.size
new_width, new_height = self.cfg.DATASET.val_central_crop_shape
left = (width - new_width)/2
top = (height - new_height)/2
right = (width + new_width)/2
bottom = (height + new_height)/2
img = img.crop((left, top, right, bottom))
segm = segm.crop((left, top, right, bottom))
self.imgMaxSize_val = 1
assert(segm.mode == "L")
assert(img.size[0] == segm.size[0])
assert(img.size[1] == segm.size[1])
ori_width, ori_height = img.size
# print('ori_width = {}, ori_height = {}'.format(ori_width, ori_height))
img_resized_list = []
img_resized_list_unnorm = []
for this_short_size in self.imgSizes_val:
if self.imgMaxSize_val == 1:
# discard 1st downsample in foveation model, i.e. creat foveation map on original image
scale = 1
else:
# calculate target height and width
scale = min(this_short_size / float(min(ori_height, ori_width)),
self.imgMaxSize_val / float(max(ori_height, ori_width)))
target_height, target_width = int(ori_height * scale), int(ori_width * scale)
# print('scaled target_height = {}, target_width = {}'.format(target_height, target_width))
# to avoid rounding in network
target_width = self.round2nearest_multiple(target_width, self.padding_constant)
target_height = self.round2nearest_multiple(target_height, self.padding_constant)
# print('rounded target_height = {}, target_width = {}'.format(target_height, target_width))
# resize images
img_resized = imresize(img, (target_width, target_height), interp='bilinear')
# image transform, to torch float tensor 3xHxW
img_resized_unnorm = self.img_transform_unnorm(img_resized)
img_resized = self.img_transform(img_resized)
img_resized = torch.unsqueeze(img_resized, 0)
img_resized_list.append(img_resized)
img_resized_unnorm = torch.unsqueeze(img_resized_unnorm, 0)
img_resized_list_unnorm.append(img_resized_unnorm)
# JC: re-ordering label according to servie degree
if 'Face_single_example' in self.cfg.DATASET.root_dataset or 'Face_single_example' in self.cfg.DATASET.list_train:
segm = self.convert_label_single_face(segm)
else:
segm = self.convert_label(segm)
# JC: for dataset like gleason2019 that background labelled 0 need to be considered
if ('GLEASON' in self.root_dataset and np.min(segm) == 0) or 'histomri' in self.cfg.DATASET.list_train or 'histomri' in self.cfg.DATASET.root_dataset:
segm = Image.fromarray(np.add(segm, 1))
# segm transform, to torch long tensor HxW
segm = self.segm_transform(segm)
batch_segms = torch.unsqueeze(segm, 0)
# batch_segms = torch.nn.functional.one_hot(batch_segms, num_classes=7)
output = dict()
output['img_ori'] = np.array(img)
output['img_data'] = [x.contiguous() for x in img_resized_list]
output['img_data_unnorm'] = [x.contiguous() for x in img_resized_list_unnorm]
output['seg_label'] = batch_segms.contiguous()
output['info'] = this_record['fpath_img']
return output
def __len__(self):
return self.num_sample
class TestDataset(BaseDataset):
def __init__(self, odgt, opt, cfg, **kwargs):
super(TestDataset, self).__init__(odgt, opt, **kwargs)
self.opt = opt
self.cfg = cfg
def __getitem__(self, index):
def crop_image(image_path):
list_crop_imgs = []
coordinate_list = []
img_path = image_path
if 'gleason2019' in self.opt.list_train:
patch_size = int(self.opt.list_train.split('train268_')[1].split('_')[0])
else:
patch_size = 5000
patch_size_x = patch_size
patch_size_y = patch_size
overlap = 0
count = 0
print(image_path)
image = cv2.imread(image_path)
ori_size = (image.shape[1], image.shape[0])
x_iter_num = (image.shape[1]+overlap)//(patch_size_x-overlap)
y_iter_num = (image.shape[0]+overlap)//(patch_size_y-overlap)
for xi in range(x_iter_num+1):
for yi in range(y_iter_num+1):
if xi == 0 or image.shape[1] < patch_size_x:
cx = 0
elif xi == x_iter_num:
cx = image.shape[1]-patch_size_x
else:
cx = xi*patch_size_x-overlap
if yi == 0 or image.shape[0] < patch_size_y:
cy = 0
elif yi == y_iter_num:
cy = image.shape[0]-patch_size_y
else:
cy = yi*patch_size_y-overlap
if image.shape[1] < patch_size_x:
patch_size_x_cur = image.shape[1]
else:
patch_size_x_cur = patch_size_x
if image.shape[0] < patch_size_y:
patch_size_y_cur = image.shape[0]
else:
patch_size_y_cur = patch_size_y
crop_ti = image[cy:cy+patch_size_y_cur, cx:cx+patch_size_x_cur]
crop_ti = Image.fromarray(crop_ti)
coordinate_list.append([cx, cy])
list_crop_imgs.append(crop_ti)
return list_crop_imgs, coordinate_list, ori_size