-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathdataset.py
2013 lines (1758 loc) · 91.9 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
'''
Defines dataset classes for various used datasets, some are not used anymore
Contains test functions for volume visualization and slice visualization
Author: Diedre Carmo
https://github.com/dscarmo
'''
import os
from os.path import join as add_path
import glob
import pickle
from sys import argv
from matplotlib import pyplot as plt
from tqdm import tqdm
import numpy as np
import cv2 as cv
import h5py
import collections
import nibabel as nib
import psutil
import time
import copy
import json
import torch
import torch.utils.data as data
from torch.utils.data import ConcatDataset
import torchvision
from transforms import ToTensor, ToFloat32, Compose, RandomAffine, Intensity, Noisify, SoftTarget, CenterCrop
import multiprocessing as mp
from multiprocessing import Lock, Process, Queue, Manager
from utils import normalizeMri, viewnii, myrotate, int_to_onehot, chunks, HALF_MULTI_TASK_NCHANNELS, MULTI_TASK_NCHANNELS
from utils import half_multi_task_labels, limit_multi_labels, imagePrint, type_assert, one_hot, get_slice, ITKManager, split_l_r
from nathip import NatHIP, get_group
cla_lock = Lock()
adni_lock = Lock()
orientations = ["sagital", "coronal", "axial"] # original data orientations
DEFAULT_PATH = "../data"
VALID_MODES = ['train', 'validation', 'test']
# Post migration paths
default_datapath = os.path.join("/home", "diedre", "Dropbox", "bigdata", "mni_hip_data")
default_adni = os.path.join("/home", "diedre", "Dropbox", "bigdata", "manual_selection_rotated", "isometric")
mni_adni = os.path.join("/home", "diedre", "Dropbox", "bigdata", "manual_selection_rotated", "raw2mni")
mni_harp = os.path.join("/home", "diedre", "Dropbox", "bigdata", "harp", "mniharp")
default_harp = os.path.join("/home", "diedre", "Dropbox", "bigdata", "harp")
multitask_path = os.path.join("/home", "diedre", "Dropbox", "bigdata", "Hippocampus", "volbrain", "PACIENTES_E_CONTROLES")
multitask_hip_processed = os.path.join("/home", "diedre", "Dropbox", "bigdata", "Hippocampus", "processed")
multitask_hip_processed_slices = os.path.join("/home", "diedre", "Dropbox", "bigdata", "Hippocampus", "processed_slices")
HARP_CLASSES = ["cn", "mci", "ad"]
class DTISegDataset(data.Dataset):
'''
Abstracts segmentation with DTI data
'''
def __init__(self, mode, path=DEFAULT_PATH, transform=None, verbose=True, orientation=None, zero_background=True,
balance="balanced_205", norm_type="zero_to_plus", split=(0.8, 0.2), overhide_folder_name=None,
limit_masks=False, patch_size=64, register_strategy='v01', use_t1=False, displasia=False, t1_separate=False):
'''
path: folder containing data
mode: one of ['train', 'validation', 'test'], test will return volumes, other will return patches
transform: list of transforms to apply
verbose: wether to print a lot of stuff or not
orientation: one of ['sagital', 'coronal', 'axial'] or None if using test mode
nlr: wether to return left and right labels
display_samples: if different than 0, displays number of samples given
'''
super(DTISegDataset, self).__init__()
assert mode in VALID_MODES, "mode {} should be one of {}".format(mode, VALID_MODES)
assert orientation in orientations or orientation is None, "orientation {} should be one of {}".format(orientation,
orientations)
if mode == 'test':
assert orientation is None, "test mode does not support orientation other than None"
assert balance == "test", "in test mode, balance does not matter, use balance='test'"
assert norm_type in ["zero_to_plus", "minus_to_plus", "mixed"], "norm type {} not support".format(norm_type)
assert balance in ["test", "2020", "205", "51010", "355", "51515", "510"], "norm type {} not support".format(balance)
assert np.array(split).sum() == 1, "split makes no sense, should sum to 1"
assert patch_size in [32, 64]
assert register_strategy in ["v01", "v02"]
assert use_t1 in [False, "t1only", "t1dti"]
self.use_t1 = use_t1
self.mode = mode
self.orientation = orientation
self.zero_background = zero_background
self.transform = transform
self.verbose = verbose
self.limit_masks = limit_masks
self.displasia = displasia
self.t1_separate = t1_separate
separator = os.sep
path_tokens = path.split(separator)
folder_name = None
if overhide_folder_name is not None:
folder_name = overhide_folder_name
else:
if self.displasia:
if mode == "test":
folder_name = "01"
elif balance == "510":
if norm_type == "zero_to_plus":
if orientation == "sagital":
folder_name = "01"
elif orientation == "coronal":
folder_name = "02"
elif orientation == "axial":
folder_name = "03"
else:
if register_strategy == "v01":
if balance == "2020":
folder_name = "00"
elif balance == "205" or balance == "test":
if norm_type == "zero_to_plus":
folder_name = "01"
elif norm_type == "minus_to_plus":
folder_name = "02"
elif norm_type == "mixed":
folder_name = "03"
elif norm_type == "zero_to_plus":
if patch_size == 32:
if balance == "51010":
folder_name = "04"
elif balance == "355":
folder_name = "05"
elif patch_size == 64:
if balance == "51010":
folder_name = "06"
elif balance == "355":
folder_name = "07"
elif register_strategy == "v02":
if balance == "test":
folder_name = "04"
elif norm_type == "zero_to_plus":
if patch_size == 64:
if balance == "355":
folder_name = "16"
elif balance == "51010":
folder_name = "18"
elif balance == "51515":
if orientation == "sagital":
if limit_masks:
folder_name = "23"
else:
# folder_name = "19"
folder_name = "20"
elif orientation == "coronal":
folder_name = "21"
elif orientation == "axial":
folder_name = "22"
elif patch_size == 32:
if balance == "355":
folder_name = "14"
elif balance == "51010":
folder_name = "15"
if folder_name is None:
raise ValueError("Unsupported combination of patch_size, balance, norm_type and register_strategy: "
"{} {} {} {}".format(patch_size, balance, norm_type, register_strategy))
self.folder_name = folder_name
if self.displasia:
if self.mode == "test":
pre_folder = "Displasia/test"
else:
pre_folder = "Displasia/patches"
elif self.mode == "test":
pre_folder = "TestData"
else:
pre_folder = "patches"
if path_tokens[0] == "..": # Work around relative pathing
glob_args = [os.path.dirname(os.getcwd())] + path_tokens[1:] + [pre_folder, folder_name, "*.npz"]
self.items = glob.glob(os.path.join(*glob_args))
else:
glob_args = (path, pre_folder, folder_name, "*.npz")
self.items = glob.glob(os.path.join(*glob_args))
glob_args[-1] = "*.txt"
try:
print(glob_args)
readme_path = glob.glob(os.path.join(*glob_args))[0]
except IndexError:
print("Readme file for dataset not found.")
data_folder = os.path.join(*glob_args[:-1])
if self.mode != "test":
print(os.path.join(data_folder, self.mode + ".pkl"))
if os.path.isfile(os.path.join(data_folder, self.mode + ".pkl")):
with open(os.path.join(data_folder, self.mode + '.pkl'), 'rb') as saved_items:
self.items = pickle.load(saved_items)
for i, v in enumerate(self.items):
self.items[i] = os.path.join(data_folder, os.path.basename(self.items[i])) # support different folders
else:
print("PKL items file not saved, creating new ones...")
stop_point = int(len(self.items)*split[0])
print("Dividing dataset in point: {}".format(stop_point))
with open(os.path.join(data_folder, 'train.pkl'), 'wb') as to_save_items:
pickle.dump(self.items[:stop_point], to_save_items)
with open(os.path.join(data_folder, 'validation.pkl'), 'wb') as to_save_items:
pickle.dump(self.items[stop_point:], to_save_items)
if self.mode == "train":
self.items = self.items[:stop_point]
elif self.mode == "validation":
self.items = self.items[stop_point:]
print("DTISegDataset initialized with nitems: {}, mode: {}, path: {}, transform: {}, "
"orientation: {}, zero_background: {}, "
"balance: {}, norm_type: {}, limit_masks: {}"
"folder_name: {}".format(len(self.items), mode, path, transform, orientation,
zero_background, balance, norm_type, limit_masks, folder_name))
with open(readme_path) as readme_file:
readme = readme_file.read()
print(('-'*20 + "\nREADME: {}\n" + '-'*20).format(readme))
def __len__(self):
'''
Returns number of items in the dataset
'''
return len(self.items)
def __getitem__(self, i):
'''
Returns input data and target
'''
if self.verbose:
print("Dataset returning {}".format(self.items[i]))
npz = np.load(self.items[i])
if self.mode == "test":
dti, target, t1 = (npz["DTI_measures"], npz["mask_onehot"], npz["T1"])
else:
dti, target, t1 = (npz["DTI_measures"], npz[(self.mode != "test")*"patch_" + "mask_onehot"],
npz[(self.mode != "test")*"patch_" + "T1"])
if self.displasia:
t2 = npz["test_T2"] if self.mode == "test" else npz["patch_T2"]
if self.use_t1 == "t1only":
data = np.zeros((1,) + t1.shape, dtype=t1.dtype)
data[0] = t1
elif self.use_t1 == "t1dti":
data = np.zeros((5,) + t1.shape, dtype=t1.dtype)
data[:4] = dti
data[4] = t1
elif self.displasia:
data = np.zeros((3,) + t1.shape, dtype=t1.dtype)
data[0] = dti
data[1] = t1
data[2] = t2
else:
data = dti
data = data.astype(np.float32)
target = target.astype(np.float32)
if self.zero_background:
target[0] = 0
if self.limit_masks is True:
# Deprecated, causes negative inbalance
buffer = np.zeros((4,) + target.shape[1:], dtype=target.dtype)
buffer[:4] = target[1:5]
# buffer[4] = target[6] # removed hip
target = buffer
if self.mode == 'test':
target[0] = 0
target[5:9] = 0
if self.transform is not None:
data, target = self.transform(data, target)
if self.mode == 'test':
return os.path.basename(self.items[i]).split('.')[0], data, target
else:
return data, target
def get_dataloader(self, batch_size, shuffle, nworkers=0):
'''
batch_size: batch size
shuffle: follow dataset order or return randomized items
nworkers: number of processes to use
'''
return data.DataLoader(self, batch_size, shuffle, num_workers=nworkers)
def display_samples(self, nsamples, display_opencv=True):
'''
Show some random samples from the dataset
'''
if self.limit_masks:
label_names = limit_multi_labels
else:
label_names = half_multi_task_labels
sample_list = []
dataloader = self.get_dataloader(nsamples, True, 0)
batch = next(iter(dataloader))
data_batch, target_batch = batch
print("Displaying {} samples in {} mode. Dataset has {} items".format(nsamples, self.mode, len(self)))
print("Data shape: {}".format(data_batch.shape))
print("Target shape: {}".format(target_batch.shape))
for batch_count, (inp, target) in enumerate(zip(data_batch, target_batch)):
np_data = inp.numpy()
np_target = target.numpy()
data_display = np.zeros((np_data.shape[1], np_data.shape[2]*np_data.shape[0]), dtype=np_data.dtype)
target_display = np.zeros((np_target.shape[1], np_target.shape[2]*np_target.shape[0]), dtype=np_target.dtype)
for channel_count, channel in enumerate(np_data):
data_display[0:np_data.shape[1],
np_data.shape[2]*channel_count:np_data.shape[2]*(channel_count + 1)
] = channel
for channel_count, channel in enumerate(np_target):
target_display[0:np_target.shape[1],
np_target.shape[2]*channel_count:np_target.shape[2]*(channel_count + 1)
] = imagePrint(channel, label_names[channel_count], org=(5, 5), scale=0.2)
data_display = cv.resize(data_display, (0, 0), fx=2, fy=2)
target_display = cv.resize(target_display, (0, 0), fx=2, fy=2)
data_display = imagePrint(data_display, "{} sample {}".format(self.mode, batch_count), org=(10, 10), scale=0.5)
sample_list.append((data_display, target_display))
if display_opencv:
cv.imshow("Input", data_display)
cv.imshow("Target", target_display)
if cv.waitKey(0) == 27:
quit()
return sample_list
# Deprecated
class Cache():
'''
Global class holding pre-processed 3D data from multitask dataset
Dynamic cache saving last used volumes, deleting old ones if not enough RAM
Can load from the processed folder, and saves everything ever used there for use in subsequent runs.
Assumes datas come "type compressed" (float16, uint8)
Has adaptative memory limit
'''
EXPECTED_SIZE = 197
MILLION = 1000000
def __init__(self, savepath=multitask_hip_processed, decompress=True):
self.cache_lock = Lock()
self.decompress = decompress
self.data = collections.OrderedDict()
self.ram_size = psutil.virtual_memory().total / Cache.MILLION
self.data_limit = 0.10*self.ram_size
self.factor = 0.05*self.ram_size
self.savepath = savepath
self.first_limit_achieved = False
self.loaded = False
print("Initial cache memory limit: {}MB".format(int(self.data_limit)))
os.makedirs(savepath, exist_ok=True)
if self.decompress:
print("Cache will return float32 data.")
else:
print("Cache will return data as stored.")
def __getitem__(self, key):
try:
self.cache_lock.acquire()
vol, mask = self.data[key]
self.cache_lock.release()
except KeyError:
self.cache_lock.release()
print("page miss, loading from HD")
key_path = self.get_key_path(key)
npz = np.load(key_path)
vol, mask, orig = (npz['vol'], npz['mask'], npz['orig'])
self.cache_lock.acquire()
self.data[key] = (vol, mask, orig)
self.cache_lock.release()
self.check_size_limit_reached()
if self.decompress:
return (vol.astype(np.float32), mask.astype(np.float32), orig.astype(np.float32))
else:
return (vol, mask)
def __setitem__(self, key, processed_data):
'''
Protected by locks
'''
vol, mask, orig = processed_data
assert vol.dtype == np.dtype(np.float16) and mask.dtype == np.dtype(np.uint8) and orig.dtype == np.dtype(np.uint8), (
" cache needs compressed data types!")
self.cache_lock.acquire()
self.data[key] = (vol, mask, orig)
self.cache_lock.release()
self.check_size_limit_reached()
key_path = self.get_key_path(key)
if not os.path.isfile(key_path):
print("Key {} cache not found, saving to {}".format(key, key_path))
np.savez_compressed(key_path, vol=vol, mask=mask, orig=orig)
else:
print("Key {} already in HD cache".format(key))
def __len__(self):
return len(self.data)
def get_key_path(self, key):
return add_path(self.savepath, key + '.npz')
def check_size_limit_reached(self):
'''
Check and deletes oldest item if its exceeds intended size
'''
size = 0
for k, v in self.data.items():
size += v[0].nbytes + v[1].nbytes
size = size/Cache.MILLION
if self.first_limit_achieved:
free = psutil.virtual_memory().available / Cache.MILLION
if free < 0.2*self.ram_size:
self.data_limit -= self.factor
if self.data_limit <= 0:
self.data_limit = 0
else:
print("Cache Memory limit decreased to {}MB due to high RAM usage".format((self.data_limit)))
if free > 0.4*self.ram_size:
self.data_limit += self.factor
if self.data_limit >= 0.8*self.ram_size:
self.data_limit = 0.8*self.ram_size
else:
print("Cache Memory limit increased to {}MB due to low RAM usage".format((self.data_limit)))
if size >= self.data_limit:
self.first_limit_achieved = True
oldest_key = list(self.data.items())[:2]
for ok in oldest_key:
print("Deleted key {} cause of memory limits".format(ok[0]))
del self.data[ok[0]]
return True
else:
return False
def load(self):
load_deprecated = True
if not self.loaded:
print("Loading cache from HD to RAM...")
if not load_deprecated: # deprecated loading
for f in glob.glob(add_path(self.savepath, '*.npz')):
npz = np.load(f)
self.cache_lock.acquire()
self.data[os.path.basename(os.path.basename(f).split('.')[0])] = (npz['vol'], npz['mask'])
self.cache_lock.release()
free = psutil.virtual_memory().available / Cache.MILLION
print("Free memory: {}".format(free), flush=True, end='\r')
if free < 0.5*self.ram_size:
print("Pre-load finished. Managed to pre-fill {} items".format(len(self)))
break
self.loaded = True
else:
print("Cache already pre-loaded")
# Fixed mnihip ids division
fixed_mnihip_ids = {'test': ['42915', '42916', '42917', '42918', '42919', '42920', '42921', '42922', '42923', '42924', '42925',
'42926', '42927', '42928', '42929', '42930', '42931', '42933', '42934', '42912'], # 42912
'train': ['31847', '31849', '31850', '31851', '31852', '31853', '31854', '31855', '31856', '32239',
'32241', '32242', '32243', '32244', '32245', '32246', '32247', '32248', '32249', '32252',
'32254', '32255', '32256', '32257', '32258', '32260', '32261', '32262', '32263', '32264',
'32265', '32266', '32268', '32269', '32270', '32271', '32272', '32273', '32274', '32275',
'32276', '32277', '32278', '32279', '32280', '32281', '32546', '32547', '32548', '32549',
'32550', '32551', '32552', '32553', '32554', '32555', '32556', '32557', '32558', '32559', '32560',
'32561', '32562', '32563', '32564', '32565', '32566', '32567', '32568', '32569', '32570', '32571',
'32572', '32573', '32574', '32575', '32576', '32577', '32578', '32579', '32580', '32581', '32582',
'32583', '33067', '33068', '33069', '33070', '33071', '33072', '33073', '33074', '33075',
'33076', '33077', '33078', '33079', '33080', '33081', '33082', '33083', '33084', '33085', '33086',
'33088', '33089', '33091', '33092', '33093', '33094', '33097', '33098',
'33099', '33100', '33101', '33103', '33104', '33105', '33107', '33758', '33760',
'33761', '33762', '33763', '33764', '33765', '33766', '33767', '33768', '33769', '33770', '33771',
'33772', '33773', '33774', '33775', '33777', '33778', '33779', '33780', '33781', '33782',
'33783', '33784', '33785', '33786', '33787', '33788', '33789', '33790', '33791', '33792',
'33794', '33795', '33796', '33797', '34423'], # 34423 previously removed
'validation': ['34421', '34422', '42892', '42893', '42894', '42895', '42896', '42897', '42898',
'42900', '42901', '42903', '42904', '42905', '42906', '42907', '42909', '42910',
'42914', '42911'] # 42911 removed
}
fixed_mnihip_ids['all'] = fixed_mnihip_ids['train'] + fixed_mnihip_ids['validation'] + fixed_mnihip_ids['test']
def unit_test(image_dataset=True, dataset="harp", shuffle=False, ntoshow=1, show=True, plt_show=True, nworkers=0,
hiponly=True, volume=False, e2d=False):
'''
Tests vizualisation of a training batch of the dataset
'''
# Long mask deprecated
'''if adni or harp:
long_mask = False
else:
long_mask = True'''
from transforms import ReturnPatch, RandomFlip
transform_list = [Compose([ReturnPatch(ppositive=1.0, patch_size=(64, 64), debug=True)]),
Compose([ReturnPatch(ppositive=0, patch_size=(64, 64), debug=True)]),
Compose([ReturnPatch(ppositive=1.0, patch_size=(64, 64), debug=True), RandomAffine(p=1.0, rotate=20,
scale=(0.8, 1.2),
debug=True)]),
Compose([ReturnPatch(ppositive=1.0, patch_size=(64, 64), debug=True), Intensity(p=1.0, brightness=0.1,
force_max=True)]),
Compose([ReturnPatch(ppositive=1.0, patch_size=(64, 64), debug=True), Noisify(p=1.0)]),
Compose([ReturnPatch(ppositive=1.0, patch_size=(64, 64), debug=True), SoftTarget(p=1.0, order=10)]),
Compose([ReturnPatch(ppositive=1.0, patch_size=(64, 64), debug=True), RandomFlip(p=1.0,
modes=['horflip'])])]
print("Testing all orientations in all modes...")
for train_transforms in transform_list:
train_transforms.addto(ToTensor(debug=True), end=True)
for m in ["test"]:
for o in ["sagital"]:
if dataset == "harp":
test = NewHARP(group="all", mode=m, orientation=o, fold=1, transform=train_transforms)
elif dataset == "adni":
test = FloatHippocampusDataset(h5path=default_adni, mode=m, transform=train_transforms,
data_split=(0.5, 0.1, 0.4), adni=True, orientation=o, hiponly=True,
return_volume=False, e2d=True, mnireg=False)
elif dataset == "oldharp":
test = FloatHippocampusDataset(h5path=default_harp, mode=m, harp=True, transform=train_transforms,
data_split=(0.7, 0.1, 0.2), adni=False, orientation=o, hiponly=True,
return_volume=False, e2d=True, mnireg=False, return_label=False)
elif dataset == "mnihip":
# test = FloatHippocampusDataset(mode=m, transform=train_transforms, orientation=o, hiponly=hiponly,
# return_volume=volume, e2d=e2d)
test = MultiTaskDataset(verbose=True, hiponly=False, mode=m, transform=train_transforms, orientation=o,
dim='2d', e2d=True, return_onehot=True, merge_left_right=True)
else:
raise ValueError("Dataset {} does not exist".format(dataset))
test_loader = data.DataLoader(test, batch_size=ntoshow, shuffle=shuffle, num_workers=0)
batch = next(iter(test_loader))
if show is True:
display_batch(batch, str(train_transforms) + o + " dataloader test in " + m)
if plt_show:
print("Showing " + str(train_transforms))
plt.show()
def display_batch(batch, title):
'''
Displays a batch content on a grid
'''
if len(batch) == 2:
imgs, tgts = batch
elif len(batch) == 3:
imgs, tgts, clss = batch
print("display_batch input:", imgs.shape, tgts.shape)
batch_len = len(imgs)
grid_data = torch.zeros((batch_len, 1, imgs.size(2), imgs.size(3)))
tgtsmax = tgts.max().float()
if tgts.max() > 1:
if not tgts.dtype == torch.long:
raise ValueError("Tgts max higher than 1 and float?")
tgts = tgts.float()/tgtsmax
for i, (im, tg) in enumerate(zip(imgs, tgts)):
if len(tg.shape) == 3 and tg.shape[0] != 1:
buffer = torch.zeros((1, tg.shape[1], tg.shape[2]), dtype=tg.dtype)
for j in range(1, tg.shape[0]):
buffer += tg[j]
buffer[buffer > 1] = 1.0
buffer[buffer < 0] = 0.0
tg = buffer
if im.shape[0] == 3:
overlap = im[1] + tg
else:
overlap = im + tg
overlap[overlap > 1] = 1.0
overlap[overlap < 0] = 0.0
if len(batch) == 2:
clas = None
else:
clas = HARP_CLASSES[clss[i].item()]
overlap = torch.from_numpy(imagePrint(overlap.squeeze().numpy(), clas)).unsqueeze(0)
grid_data[i] = overlap
grid = torchvision.utils.make_grid(grid_data, nrow=batch_len//5).numpy().transpose(1, 2, 0)
plt.figure(num=title)
plt.title(str(batch_len) + " " + title + " samples")
plt.axis('off')
plt.imshow(grid)
def view_volumes(dataset_name="mnihip", wait=0, group=None, load_test=False, use_itk_snap=False, split=False):
'''
View volumes supplied by a dataset abstraction
'''
if dataset_name == "harp":
fhd = NewHARP("all", mode="all", verbose=True)
elif dataset_name == "mnihip":
fhd = ConcatDataset((FloatHippocampusDataset(return_volume=True, transform=None, orientation="coronal", mode="train",
verbose=True),
FloatHippocampusDataset(return_volume=True, transform=None, orientation="coronal", mode="validation",
verbose=True),
FloatHippocampusDataset(return_volume=True, transform=None, orientation="coronal", mode="test",
verbose=True)))
elif dataset_name == "oldharp":
fhd = ConcatDataset((FloatHippocampusDataset(h5path=default_harp, return_volume=True, mode="train", adni=False, harp=True,
data_split=(0.7, 0.1, 0.2), mnireg=False, return_label=False),
FloatHippocampusDataset(h5path=default_harp, return_volume=True, mode="validation", adni=False,
harp=True, data_split=(0.7, 0.1, 0.2), mnireg=False, return_label=False),
FloatHippocampusDataset(h5path=default_harp, return_volume=True, mode="test", adni=False, harp=True,
data_split=(0.7, 0.1, 0.2), mnireg=False, return_label=False)))
elif dataset_name == "mniharp":
fhd = HARP("all", mode="all", mnireg=True)
elif dataset_name == "cc359":
fhd = CC359Data()
elif dataset_name == "adni":
fhd = FloatHippocampusDataset(h5path=default_adni, return_volume=True, mode="test", adni=True, data_split=(0.0, 0.0, 1.0),
mnireg=False)
elif dataset_name == "mniadni":
fhd = FloatHippocampusDataset(h5path=mni_adni, return_volume=True, mode="test", adni=True, data_split=(0.0, 0.0, 1.0),
mnireg=True)
elif dataset_name == "multitask":
fhd = MultiTaskDataset(verbose=False, hiponly=False, return_onehot=True)
elif dataset_name == "hipmultitask":
fhd = MultiTaskDataset(verbose=True, hiponly=True, return_onehot=False)
else:
raise ValueError("Dataset {} not recognized".format(dataset_name))
itk_manager = ITKManager() if use_itk_snap else None
print("Dataset size: {}".format(len(fhd)))
for i in range(len(fhd)):
label = None
try:
data = fhd[i]
if len(data) == 2:
im, ma = data
elif len(data) == 3:
im, ma, label = data
# Handle 4D volumes (multichannel masks)
if ma.ndim == 4:
preserve_type = ma.dtype
buffer = np.zeros(ma.shape[1:])
for c in range(ma.shape[0]):
buffer = buffer + ma[c]*c
buffer = buffer/buffer.max()
ma = buffer.astype(preserve_type)
else:
ma = ma/ma.max()
if not load_test:
if hasattr(fhd, "multilabels"):
multilabel = fhd.multilabels
else:
multilabel = None
if split:
splitted_im = split_l_r(im)
splitted_ma = split_l_r(ma)
viewnii(splitted_im["left"], splitted_ma["left"], id=dataset_name + " left", wait=wait,
multi_labels=multilabel, label=label, itk_manager=itk_manager)
viewnii(splitted_im["right"], splitted_ma["right"], id=dataset_name + " right", wait=wait,
multi_labels=multilabel, label=label, itk_manager=itk_manager)
else:
viewnii(im, ma, id=dataset_name, wait=wait, multi_labels=multilabel, label=label, itk_manager=itk_manager)
except KeyboardInterrupt:
print("Dataset test interrupted by Ctrl-C")
quit()
class MultiTaskDataset(data.Dataset):
'''
Abstracts raw mnihip with all masks
Should make FloatHippocampusDataset deprecated when completed
Provides option to ramcache all data, for fast h5 file less slice extraction
'''
EXPECTED_SHAPE = (181, 217, 181)
presence_dict = None
cache = None
@staticmethod
def producer(r, data, queue):
'''
Gets data from dataset and puts in queue to go to cache
'''
for i in tqdm(r):
queue.put(data[i])
queue.put(None)
@staticmethod
def fill_hd_cache(nworkers=mp.cpu_count()):
'''
Fills ramcache with all volumes, ram cache is dict refering to ID, getitem will use correct ID
RAM CACHE is class with get and set methods that interface float32 and float16
'''
assert nworkers >= 0, "nworkers cant be negative"
data = MultiTaskDataset(verbose=True, return_id=True, use_raw_data=True, compress=True)
if len(MultiTaskDataset.cache) != len(data):
print("Cache file not found or incomplete, RAM caching being filled... This might take some time.")
if nworkers == 0:
for i in tqdm(range(len(data))):
Id, vol, mask, orig = data[i]
MultiTaskDataset.cache[Id] = (vol, mask, orig)
else:
print("Using multiprocessing with {} workers for RAM fill...".format(nworkers))
queue = Queue(maxsize=(psutil.virtual_memory().total//Cache.MILLION)//8)
data_len = len(data)
for r in chunks(list(range(data_len)), data_len//nworkers):
ps = []
p = Process(target=MultiTaskDataset.producer, args=(r, data, queue))
ps.append(p)
p.start()
done = 0
print("Consuming data fetched by {} workers".format(nworkers))
while True:
data = queue.get()
print("Get Got!")
if data is None:
done += 1
else:
vol_id, vol, mask, orig = data
MultiTaskDataset.cache[vol_id] = (vol, mask, orig)
if done == nworkers:
return
for p in ps:
p.join()
else:
print("RAM cache already filled with {} entries".format(len(MultiTaskDataset.cache)))
def __init__(self, path=multitask_path, group='both', mode='all', data_split=(0.8, 0.1, 0.1), transform=None, verbose=True,
hiponly=False, use_raw_data=False, dim='3d', orientation=None, e2d=False, return_id=False, compress=False,
return_onehot=True, zero_background=True, merge_left_right=True):
'''
path: path that contains folders ['controls', 'patients']
group: select one of 'controls', 'patients', or 'both'
mode: select one of 'all', 'train', 'validation', or 'test'
data_split: how to separate data between modes, same separation not guaranteed if not using default value
transform: transform to apply to the data, make sure its compatible with dim
verbose: more prints if True
hiponly: returns slices that dont have mask when dim == '2d' if True
ramcache: stores all volumes on RAM in initialization phase if True
dim: return slices if '2d', or volumes if '3d'
orientation: when using dim='2d', selects orientation to return
e2d: return input as 3 neighbour slices
return_id: returns id, data, target instead of data, target
compress: returns in a compressed datatype instead of float32
return_onehot: returns target in onehot format
zero_background: zero out onehot background (cross entropy should ignore index 0)
merge_left_right: if true, do not return different label for left and right
'''
super(MultiTaskDataset, self).__init__()
# Assert arguments make sense
valid_modes = ['all', 'train', 'validation', 'test']
valid_groups = ['both', 'controls', 'patients']
valid_dims = ['2d', '3d']
assert mode in valid_modes and group in valid_groups and dim in valid_dims, ("arguments to MultiTask dataset make no"
"sense check documentation")
if dim == '3d':
assert e2d is False and orientation is None, ("e2d and orientation makes no sense when dim == '3d', also hiponly has"
"to be False")
assert hiponly != merge_left_right or (merge_left_right is False and hiponly is False), ("choose one: merge sides or"
"hiponly")
assert return_onehot, "option to not return one hot is currently disabled"
self.transform = transform
self.verbose = verbose
self.hiponly = hiponly
self.use_raw_data = use_raw_data
self.e2d = e2d
self.orientation = orientation
self.return_onehot = return_onehot
self.zero_background = zero_background
self.merge_left_right = merge_left_right
self.multilabels = HALF_MULTI_TASK_NCHANNELS if merge_left_right else MULTI_TASK_NCHANNELS
if orientation is not None:
self.n_slices = MultiTaskDataset.EXPECTED_SHAPE[orientations.index(orientation)]
self.slice_mode = True
# Initialize presence dict if not initialized already
if MultiTaskDataset.presence_dict is None:
with open(add_path(multitask_hip_processed, 'hip_presence.json'), 'r') as f:
MultiTaskDataset.presence_dict = json.load(f)
self.presence_dict = copy.deepcopy(MultiTaskDataset.presence_dict)
print("Presence dict initialized with {} volumes".format(len(self.presence_dict)))
else:
self.slice_mode = False
self.dim = dim
self.return_id = return_id
self.compress = compress
self.reconstruction_orientations = orientations
if compress:
self.vol_type, self.mask_type, self.orig_type = np.dtype(np.float16), np.dtype(np.uint8), np.dtype(np.uint8)
print("WARNING: Returning compressed data")
else:
self.vol_type, self.mask_type, self.orig_type = np.dtype(np.float32), np.dtype(np.float32), np.dtype(np.float32)
os_separator = "\\" if os.name == 'nt' else '/'
files = list(glob.glob(add_path(path, "**", "*.nii"), recursive=True))
if not self.slice_mode:
MultiTaskDataset.cache = Cache()
if len(glob.glob(add_path(MultiTaskDataset.cache.savepath, '*.npz'))) == Cache.EXPECTED_SIZE:
MultiTaskDataset.cache.load()
print("Datacache pre-loaded to limit, from {}".format(MultiTaskDataset.cache.savepath))
# Store tuples dict['id'] = (vol, mask)
controls = {}
patients = {}
# Get information from nii files
for f in files:
ftokens = f.split(os_separator)
if not ('SOBROU' in ftokens):
bname = os.path.basename(f)
number = bname[-9:-4]
if bname[0:3] == 'lab':
index = 1
elif bname[0] == 'n':
index = 0
else:
continue
if "PACIENTES" in ftokens:
try:
patients[number]
except KeyError:
patients[number] = ['', '']
patients[number][index] = f
elif "CONTROLES" in ftokens:
try:
controls[number]
except KeyError:
controls[number] = ['', '']
controls[number][index] = f
# Remove ids not in selected mode
for k in list(controls.keys()):
if k not in fixed_mnihip_ids[mode]:
controls.pop(k)
for k in list(patients.keys()):
if k not in fixed_mnihip_ids[mode]:
patients.pop(k)
citems = list(controls.items())
pitems = list(patients.items())
# Remove ids not in selected group
if group == 'both':
self.final_items = citems + pitems
if self.verbose:
print("{} control + patients items".format(len(self.final_items)))
elif group == 'controls':
self.final_items = citems
if self.verbose:
print("{} control items".format(len(self.final_items)))
elif group == 'patients':
self.final_items = pitems
if self.verbose:
print("{} patient items".format(len(self.final_items)))
self.n_vols = len(self.final_items)
if self.slice_mode:
self.slice_paths = []
for k, v in self.final_items:
for slic in self.presence_dict[k][self.orientation]:
self.slice_paths.append(add_path(multitask_hip_processed_slices, k + '_' + self.orientation[0] + str(slic) +
".npz"))
print(("Multitask init in group {}, mode {}, datasplit {}, use_raw_data {}, compress {} return_onehot {}"
"zero background {}, merge_left_right {}, done.").format(group, mode, data_split, use_raw_data, compress,
return_onehot, zero_background, merge_left_right))
def __len__(self):
if self.slice_mode:
return len(self.slice_paths)
else:
return self.n_vols
def __getitem__(self, i):
'''
Masks are one hots
'''
gitem_time = time.time()
if self.slice_mode:
slice_path = self.slice_paths[i]
if self.verbose:
print(slice_path)
mtask_id = slice_path.split('/')[-1].split('_')[0]
self.slice_index = i
if self.verbose:
print("Slice came from {}, path {}".format(mtask_id, slice_path))
else:
mtask_id, (volpath, maskpath) = self.final_items[i]
if self.verbose:
print("Subject {} Volpath: {}".format(mtask_id, volpath))
# Select slices if in slice_mode
if self.slice_mode:
npz = np.load(self.slice_paths[i])
center_img = npz['vol_slice']
target = npz['mask_slice'] if self.return_onehot else npz['orig_slice']
if self.e2d:
preindex = self.slice_index - 1
if preindex < 0:
preindex = self.slice_index
postindex = self.slice_index + 1
if postindex == len(self):
postindex = self.slice_index
image = np.zeros((3, center_img.shape[0], center_img.shape[1]), dtype=center_img.dtype)
image[0] = np.load(self.slice_paths[preindex])['vol_slice']
image[1] = center_img
image[2] = np.load(self.slice_paths[postindex])['vol_slice']
else:
image, target = center_img, target
image, target = image.astype(self.vol_type), target.astype(self.mask_type if self.return_onehot else self.orig_type)
else:
# Get the volume from memory
if self.use_raw_data:
# Gets raw data, process and returns
vol = nib.load(volpath).get_fdata()
mask = nib.load(maskpath).get_fdata()
vs = vol.shape
ms = mask.shape
if vs != ms:
print("WARNING: shapes of mask and volume are not equal!")
norm_vol = normalizeMri(vol.astype(np.float32)).astype(self.vol_type)
norm_mask = int_to_onehot(mask, onehot_type=self.mask_type)
norm_orig = mask.astype(self.orig_type)
else:
# get already processed data
norm_vol, norm_mask, norm_orig = MultiTaskDataset.cache[mtask_id]
image = norm_vol
if self.return_onehot:
target = norm_mask
else:
target = norm_orig
target_type = self.mask_type if self.return_onehot else self.orig_type
assert image.dtype == self.vol_type and target.dtype == target_type, (" wrong vol or mask dtype before transforms,"
"should be {} and {}").format(self.vol_type,
self.mask_type)
if self.hiponly:
if self.return_onehot:
target = target[11] + target[12]
else:
target = np.ma.masked_not_equal(target, 11).filled(0) + np.ma.masked_not_equal(target, 12).filled(0)
elif self.merge_left_right:
if self.return_onehot:
new_target = np.zeros((HALF_MULTI_TASK_NCHANNELS,) + target.shape[1:], dtype=target.dtype)
new_target[0] = target[0]
for i in range(1, HALF_MULTI_TASK_NCHANNELS):
new_target[i] = target[2*i-1] + target[2*i]
else:
new_target = np.zeros_like(target)
new_target = (target + 1)//2
target = new_target
# Remove background if returning one hot. If not returning one hot, make sure ignore_index 0 is on cross entropy.
if self.return_onehot and self.zero_background:
if target.ndim == 3:
target[0, :, :] = 0
elif target.ndim == 4:
target[0, :, :, :] = 0
else: