-
Notifications
You must be signed in to change notification settings - Fork 1
/
data_loader.py
1533 lines (1106 loc) · 62.5 KB
/
data_loader.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
# we are (loosely) following the YOLO format for data formatting and bounding box annotations
# see here: https://github.com/AlexeyAB/Yolo_mark/issues/60 for a description
# external imports
import os
import random
import time
from tqdm import tqdm # progress bar
import torch
from torch.utils.data.sampler import SubsetRandomSampler # RandomSampling
from torch.utils.data import DataLoader, Dataset
import torchvision.transforms.functional as TF
import torchvision.transforms as T
from torchvision import utils
import pandas as pd
import numpy as np
import cv2 # imread
import matplotlib.pyplot as plt
import matplotlib.patches as patches
import warnings # silent warning on (torch.stack(images))
import scipy # https://docs.opencv.org/4.5.2/d4/d13/tutorial_py_filtering.html
from skimage import io
# internal imports
import config as c
import gvars as g
import arguments as a
import utils
warnings.filterwarnings("ignore", message=r"Passing", category=FutureWarning)
proj_dir = c.proj_dir
random_flag = False
points_flag = True
demo = False
density_demo = True
mk_size = 25
#### DLR ACD classes
class DLRACD(Dataset):
# TODO add image name so concordance with gsd table is possible
def __init__(self, root_dir,overlap=50,transform = None):
self.root_dir = root_dir + '/DLRACD/'
# this dict structure is only used for processing in init method,
# switch to flat list at end
self.count = False
self.dlr_acd = True
self.classification = False # TODO - no null filtering - yet
self.transform = transform
self.overlap = overlap
self.images = {'Train':[],'Test':[]}
self.counts = {'Train':[],'Test':[]}
self.density_counts = {'Train':[],'Test':[]}
self.annotations = {'Train':[]}
self.anno_path = {'Train':[]}
self.train_indices = []
self.test_indices = []
self.patches = {'Train':[],'Test':[]}
self.patch_densities = {'Train':[],'Test':[]}
self.patch_path = {'Train':[],'Test':[]}
self.point_maps = {'Train':[]}
self.point_map_path = {'Train':[]}
gsd_table_df = pd.read_csv(self.root_dir+"dlracd_gsds.txt")
print('Initialising dataset...')
### Patches section
for mode in ['Test','Train']:
i_paths = os.listdir(os.path.join(self.root_dir,mode+"/Images/"))
image_files = [f for f in i_paths if os.path.isfile(os.path.join(self.root_dir,mode+"/Images/", f))]
images_list = sorted(image_files)
if overlap == 0:
for img_path in tqdm(images_list, desc="Loading and splitting {} images...".format(mode)):
im = cv2.imread(os.path.join(self.root_dir,mode+"/Images/"+img_path))
self.images[mode].append(im)
im_patches = utils.split_image(im, patch_size = 320,save = False, overlap=0)
self.patches[mode].extend(im_patches)
self.patch_path[mode].extend([img_path]*len(im_patches))
else:
patch_save = not os.path.exists(self.root_dir+mode+"/Images/Patches/")
im_patch_path = self.root_dir+mode+"/Images/Patches/"
if patch_save:
os.makedirs(im_patch_path, exist_ok = False)
# create image patches
for img_path in tqdm(images_list[:99], desc="Loading and splitting {} images...".format(mode)):
im = cv2.imread(os.path.join(self.root_dir,mode+"/Images/"+img_path))
im_patches = utils.split_image(im, save = patch_save, overlap=overlap,name = img_path[:-4],path = im_patch_path,frmt = 'jpg',dlr=True)
patches_list = sorted(os.listdir(os.path.join(self.root_dir,mode+"/Images/Patches/")))
for patch_path in tqdm(patches_list[:99], desc="Loading {} patches...".format(mode)):
im_patch = cv2.imread(os.path.join(self.root_dir,mode+"/Images/Patches/"+patch_path))
self.patches[mode].append(im_patch) # (320, 320, 3)
self.patch_path[mode].append(os.path.join(self.root_dir,mode+"/Images/Patches/"+patch_path))
if mode == 'Train': # only Train data available
### Annotations section
a_paths = os.listdir(os.path.join(self.root_dir,mode+"/Annotation/"))
anno_files = [f for f in a_paths if os.path.isfile(os.path.join(self.root_dir,mode+"/Annotation/", f))]
annotation_list = sorted(anno_files)
if overlap == 0:
for anno_path in tqdm(annotation_list[:99], desc="Loading and splitting annotations..."):
# read in as greyscale (0 arg)
annotation = cv2.imread(os.path.join(self.root_dir,mode+"/Annotation/"+anno_path),0)
self.annotations[mode].append(annotation)
point_maps = utils.split_image(annotation, patch_size = 320,save = False, overlap=0)
for pm in point_maps:
pm = pm / 255
self.counts[mode].append(pm.sum())
self.point_maps[mode].append(pm)
self.point_map_path[mode].append(anno_path)
# NB convert pd series to list, 0th element
gsd_sigma = gsd_table_df[gsd_table_df['name'] == anno_path[:-4]]['gsd'].values[0]
if c.debug:
print('Anno path is {}'.format(anno_path))
print('GSD is {}'.format(gsd_sigma))
if a.args.model_name in ['UNet_seg','LCFCN']:
p_d = scipy.ndimage.filters.maximum_filter(pm,size = (a.args.max_filter_size,a.args.max_filter_size))
else:
p_d = scipy.ndimage.filters.gaussian_filter(pm, sigma = gsd_sigma, mode='constant')
self.density_counts[mode].append(p_d.sum())
self.patch_densities[mode].append(p_d)
else:
patch_save = not os.path.exists(self.root_dir+mode+"/Annotation/Patches/")
anno_patch_path = self.root_dir+mode+"/Annotation/Patches/"
if patch_save:
os.makedirs(anno_patch_path, exist_ok = False)
# create annotation patches
for anno_path in tqdm(annotation_list[:99], desc="Loading and splitting images..."):
anno = cv2.imread(os.path.join(self.root_dir,mode+"/Annotation/"+anno_path))
anno_patches = utils.split_image(anno,save=patch_save, overlap=overlap,name = anno_path[:-4],path = anno_patch_path,frmt = 'png',dlr=True)
self.anno_path[mode].extend([anno_path]*len(anno_patches))
anno_patches_list = sorted(os.listdir(os.path.join(self.root_dir,mode+"/Annotation/Patches/")))
for anno_patch_path,anno_path in tqdm(zip(anno_patches_list[:99],self.anno_path[mode]), desc="Loading annotation patches..."):
# read in as greyscale (0 arg)
pm = cv2.imread(os.path.join(self.root_dir,mode+"/Annotation/Patches/"+anno_patch_path),0)
pm = pm / 255
self.counts[mode].append(pm.sum())
self.point_maps[mode].append(pm)
self.point_map_path[mode].append(anno_patch_path)
# NB convert pd series to list, 0th element
gsd_sigma = gsd_table_df[gsd_table_df['name'] == anno_path[:-4]]['gsd'].values[0]
if c.debug:
print('Anno path is {}'.format(anno_path))
print('Anno patch path is {}'.format(anno_patch_path))
print('GSD is {}'.format(gsd_sigma))
p_d = scipy.ndimage.filters.gaussian_filter(pm, sigma = gsd_sigma, mode='constant')
self.density_counts[mode].append(p_d.sum())
self.patch_densities[mode].append(p_d)
## jointly shuffle data indices
joint_lists = list(zip(self.patches['Train'],self.patch_densities['Train'],
self.counts['Train'],self.density_counts['Train'],self.point_maps['Train']))
random.shuffle(joint_lists)
self.patches['Train'],self.patch_densities['Train'],self.counts['Train'],self.density_counts['Train'],self.point_maps['Train'] = zip(*joint_lists)
break_point = round(0.7*len(self.patches['Train']))
self.train_indices = range(0,break_point)
self.test_indices = range(break_point,len(self.patches['Train']))
#self.test_indices = range(len(self.train_indices),len(self.train_indices)+len(self.patches['Test']))
self.patches = [*self.patches['Train'],*self.patches['Test']]
self.patch_path = [*self.patch_path['Train'],*self.patch_path['Test']]
self.patch_densities = [*self.patch_densities['Train']] # no test annotations provided
self.counts = [*self.counts['Train']]
self.density_counts = [*self.density_counts['Train']]
self.point_maps = [*self.point_maps['Train']]
self.point_map_path = [*self.point_map_path['Train']]
self.hparam_dict = None
self.metric_dict = None
print("Number of DLR ACD image patches: {}".format(len(self.patches)))
print("Number of DLR ACD point map patches: {}".format(len(self.point_maps)))
print("Number of DLR ACD density map patches: {}".format(len(self.patch_densities)))
time.sleep(1)
print('\nInitialisation finished')
def __len__(self):
return len(self.patches['Train'])+len(self.patches['Test'])
def __getitem__(self, idx):
if torch.is_tensor(idx):
idx = idx.tolist()
sample = {}
sample['patch'] = self.patches[idx]
if idx < len(self.patch_densities):
# if self.overlap == 0:
sample['patch_density'] = self.patch_densities[idx]
sample['counts'] = self.counts[idx]
sample['density_counts'] = self.density_counts[idx]
sample['point_map'] = self.point_maps[idx]
else:
sample['patch_density'] = None
sample['counts'] = None
sample['density_counts'] = None
sample['point_map'] = None
if self.transform:
sample = self.transform(sample)
return sample
def show_annotations(self, idx):
sample = self[idx]
print("True Count:")
print(sample["counts"])
print('Paths')
print((self.patch_path[idx]+" | "+self.point_map_path[idx]))
patch = sample['patch'].permute((1, 2, 0)).cpu().numpy()
point_map = sample['point_map'] .cpu().numpy()
gt_coords = np.argwhere(point_map == 1)
patch_densities = sample['patch_density'].cpu().numpy()
fig, ax = plt.subplots(1,3,figsize=(24, 8))
ax[0].imshow(patch_densities, cmap='viridis', interpolation='nearest', aspect="auto")
ax[1].imshow((patch * 255).astype(np.uint8), aspect="auto")
ax[2].imshow((point_map * 255).astype(np.uint8), aspect="auto")
if len(gt_coords) != 0:
ax[2].scatter(gt_coords[:,1], gt_coords[:,0],c='red',marker='1',s=30,
label='Ground truth coordinates')
#ax[2].set_ylim([0, 320])
#ax[2].set_xlim([0, 320])
fig.tight_layout()
fig.subplots_adjust(top=0.90)
fig.suptitle('Sample {}: \nDensity Count: {}, True Count: {}'.format(idx,round(sample['density_counts'],2),sample['counts']),y=1.0,fontsize=24)
def custom_collate_aerial(self,batch):
patches = list()
patch_densities = list()
counts = list()
point_maps = list()
for b in batch:
patches.append(b['patch'])
patch_densities.append(b['patch_density'])
counts.append(b['counts'])
point_maps.append(b['point_map'])
patches = torch.stack(patches,dim = 0)
patch_densities = torch.stack(patch_densities,dim = 0)
counts = torch.stack(counts,dim = 0)
point_maps = torch.stack(point_maps,dim = 0)
collated_batch = patches,patch_densities,counts,point_maps
return collated_batch
class DLRACDToTensor(object):
"""Convert ndarrays in sample to Tensors, move to GPU."""
def __call__(self, sample):
patch = sample['patch']
# swap color axis because
# numpy image: H x W x C
# torch image: C x H x W
patch = torch.from_numpy(patch)
patch = patch.permute(2,0,1)
patch = patch.float().div(255)#.to(c.device)
sample['patch'] = patch.float()
if 'patch_density' in sample.keys():
sample['patch_density'] = torch.from_numpy(sample['patch_density'])#.to(c.device)
if 'counts' in sample.keys():
sample['counts'] = torch.from_numpy(np.array(sample['counts']).astype(float))#.to(c.device)
if 'point_map' in sample.keys():
sample['point_map'] = torch.from_numpy(sample['point_map'] )#.to(c.device)
return sample
class DLRACDAddUniformNoise(object):
"""Add uniform noise to Dmaps to stabilise training."""
def __init__(self, r1=0., r2=a.args.noise):
self.r1 = r1
self.r2 = r2
def __call__(self, sample):
# uniform tensor in pytorch:
# https://stackoverflow.com/questions/44328530/how-to-get-a-uniform-distribution-in-a-range-r1-r2-in-pytorch
if 'patch_density' in sample.keys():
pass
sample['patch_density'] = sample['patch_density'] + torch.FloatTensor(sample['patch_density'].size()).uniform_(self.r1, self.r2)#.to(c.device)
if c.debug:
print("uniform noise ({},{}) added to patch density".format(self.r1, self.r2))
return sample
class DLRACDCropRotateFlipScaling(object):
"""Randomly rotate, flip and scale aerial image and density map."""
def __call__(self, sample):
# Left - Right, Up - Down flipping
# 1/4 chance of no flip, 1/4 chance of no rotation, 1/16 chance of no flip or rotate
# resize = T.Resize(size=(a.args.image_size,a.args.image_size))
# i, j, h, w = T.RandomCrop.get_params(sample['patch'], output_size=(a.args.image_size, a.args.image_size))
# sample['patch'] = TF.crop(sample['patch'].unsqueeze(0), i, j, h, w)
# sample['patch_density'] = TF.crop(sample['patch_density'].unsqueeze(0).unsqueeze(0), i, j, h, w)
# sample['point_map'] = TF.crop(sample['point_map'] .unsqueeze(0).unsqueeze(0), i, j, h, w)
# sample['patch'] = resize(sample['patch'])
# sample['patch_density'] = resize(sample['patch_density'])
# sample['point_map'] = resize(sample['point_map'] )
sample['patch'] = sample['patch'].unsqueeze(0)
sample['patch_density']= sample['patch_density'].unsqueeze(0).unsqueeze(0)
sample['point_map'] = sample['point_map'].unsqueeze(0).unsqueeze(0)
if random.randint(0,1):
sample['patch'] = torch.flip(sample['patch'],(3,))
sample['patch_density'] = torch.flip(sample['patch_density'],(3,))
sample['point_map'] = torch.flip(sample['point_map'] ,(3,))
if random.randint(0,1):
sample['patch'] = torch.flip(sample['patch'],(2,))
sample['patch_density'] = torch.flip(sample['patch_density'],(2,))
sample['point_map'] = torch.flip(sample['point_map'] ,(2,))
rangle = float(random.randint(0,3)*90)
sample['patch'] = TF.rotate(sample['patch'],angle=rangle)
sample['patch_density'] = TF.rotate(sample['patch_density'],angle=rangle)
sample['point_map'] = TF.rotate(sample['point_map'] ,angle=rangle)
sample['patch'] = sample['patch'].squeeze()
sample['patch_density'] = sample['patch_density'].squeeze().squeeze()
sample['point_map'] = sample['point_map'] .squeeze().squeeze()
return sample
#### Cow Flow Classes
# GSD = 0.3m
class CowObjectsDataset(Dataset):
"""Cow Objects dataset."""
def __init__(self, root_dir,transform=None,convert_to_points=False,generate_density=False,model_name=a.args.model_name,
count=False,classification=False,ram=False,holdout=a.args.holdout,sat=False,resize=a.args.resize):
"""
Args:
root_dir (string): Directory with the following structure:
object.names file
object.data file
obj directory containing imgs and txt annotations
each img has a corresponding txt file
test.txt file
train.txt file
transform (callable, optional): Optional transform to be applied
on a sample.
"""
self.root_dir = root_dir
self.transform = transform
self.points = convert_to_points
self.density = generate_density
self.count = count
self.classification = classification
self.ram = ram
self.holdout = holdout
self.sat = sat
if not self.density:
self.sigma = 0
else:
self.sigma = a.args.sigma
self.images = []
self.annotations_list = []
self.point_maps = []
self.density_list = []
self.labels_list = []
self.count_list = []
self.binary_labels_list = []
self.im_names = []
self.resize = resize
names = []
with open(os.path.join(self.root_dir,"object.names")) as f:
names_input = f.readlines()
for line in names_input:
names.append(line.strip())
data = {}
with open(os.path.join(self.root_dir,"object.data")) as f:
data_input = f.readlines()
for line in data_input:
data[line.split("=")[0].strip()] = line.split("=")[1].strip()
self.names = names
self.data = data
self.train_path = os.path.join(self.root_dir,os.path.split(self.data["train"])[1])
self.holdout_path = os.path.join(self.root_dir,os.path.split(self.data["holdout"])[1])
self.sat_path = os.path.join(self.root_dir,os.path.split(self.data["satellite"])[1])
im_paths = []
if self.holdout:
path = self.holdout_path
elif self.sat:
path = self.sat_path
else:
path = self.train_path
with open(os.path.join(self.root_dir,path)) as f:
im_names_input = f.readlines()
for line in im_names_input:
im_paths.append(os.path.join(self.root_dir,'obj/',line.strip()))
self.im_names.append(line)
self.im_paths = im_paths
def compute_labels(idx,resize=False):
# this will break with use of config file.... # TODO URGENT
if a.args.model_name in ['UNet_seg','LCFCN']: #a.args.dmap_type == 'max':
dmap_type = '_max'
else:
dmap_type = '' # _gauss
if resize:
dmap_size = '_resized'
else:
dmap_size = '' # _gauss
""" computes and returns labels for a single annotation file"""
labels = []
# print('len(self.im_paths)')
# print(len(self.im_paths))
# print('idx')
# print(idx)
img_path = os.path.join(self.root_dir,
self.im_paths[idx])
if self.sat:
txt_path = img_path[:-4]+'txt'
else:
txt_path = img_path[:-3]+'txt'
image = io.imread(img_path)
# check if annotations file is empty
header_list = ['class', 'x', 'y', 'width', 'height']
with open(txt_path) as annotations:
count = len(annotations.readlines())
dmap_path = g.DMAP_DIR+self.im_names[idx]
# using ray breaks relative pathing above somehow
# if a.args.mode == 'search':
# dmap_path = "/home/mks29/clones/cow_flow"+dmap_path[1:]
if self.ram and not a.args.mode == 'store':
if not os.path.exists(dmap_path):
ValueError("Dmaps must have been previously stored!")
store = np.load(dmap_path[:-5]+dmap_type+dmap_size+'.npz',allow_pickle=True)
density_map = store['arr_0']
# import matplotlib.pyplot as plt
# fig, ax = plt.subplots(1,1)
# ax.imshow(density_map)
# 1/0
labels = store['arr_1']
annotations = store['arr_2']
point_map = store['arr_3']
else:
if count == 0:
annotations = np.array([])
density_map = create_maps() #, point_map
# add points onto basemap
base_map, point_flag = utils.create_point_map(mdl=None,annotations=annotations,resize=resize)
point_map = base_map
labels.extend(point_flag)
else:
annotations = pd.read_csv(txt_path,names=header_list,delim_whitespace=True)
annotations = annotations.to_numpy()
if self.points:
# convert annotations to points
# delete height and width columns, which won't be used in density map
# becomes: <object-class> <centre-x> <centre-y>
annotations = np.delete(arr = annotations,obj = [3,4],axis = 1) # np 2d: row = axis 0, col = axis 1
if self.density:
if not self.points:
print("Generation of maps requires conversion to points")
assert self.points
# running gaussian filter over points as in crowdcount mcnn
# https://github.com/svishwa/crowdcount-mcnn/blob/master/data_preparation/get_density_map_gaussian.m
# https://stackoverflow.com/questions/17190649/how-to-obtain-a-gaussian-filter-in-python
# map generation
# set density map image size equal to data image size
# TODO - this is possibly massively inefficient
density_map = create_maps(resize=resize) # , point_map
# add points onto basemap
base_map, point_flag = utils.create_point_map(mdl=None,annotations=annotations,resize=resize)
point_map = base_map
labels.extend(point_flag)
# if a.args.model_name in ['LCFCN','UNet_seg'] or a.args.rrc:
# point_map = base_map
# else:
# point_map = None
if a.args.model_name in ['UNet_seg','LCFCN']:
density_map += scipy.ndimage.filters.maximum_filter(base_map,size = (4,4)) # a.args.max_filter_size,a.args.max_filter_size
else:
density_map += scipy.ndimage.filters.gaussian_filter(base_map, sigma = 4, mode='constant') # a.args.sigma
# import matplotlib.pyplot as plt
# fig, ax = plt.subplots(1,1)
# ax.imshow(density_map)
# 1/0
labels = np.array(labels) # list into default collate function produces empty tensors
# store dmaps/labels/annotations
if a.args.mode == 'store':
if not os.path.exists(g.DMAP_DIR):
os.makedirs(g.DMAP_DIR)
np.savez_compressed(dmap_path[:-5]+dmap_type+dmap_size,density_map,labels,annotations,point_map,allow_pickle=True)
sample = {'image': image}
if self.density and not self.count:
sample['density'] = density_map; sample['labels'] = labels
sample['point_map'] = point_map
if self.count:
sample['counts'] = torch.as_tensor(count).float()
if self.classification:
positive = (len(annotations) == 0)
sample['binary_labels'] = torch.tensor(positive).type(torch.LongTensor)
sample['annotations'] = annotations
return sample
self.compute_labels = compute_labels
if self.ram:
if not a.args.mode == 'store' and self.ram:
desc = "Loading images, annotations, dmaps and labels into RAM"
elif self.ram and a.args.mode == 'store':
desc = "Storing dmaps, pmaps, annotations and labels to file"
for idx in tqdm(range(len(self.im_paths)),desc=desc):
#for idx in [5895]:
sample = compute_labels(idx)
if not a.args.mode == 'store':
self.images.append(sample['image'])
if self.density and not self.count:
self.density_list.append(sample['density'])
self.point_maps.append(sample['point_map'])
self.labels_list.append(sample['labels'])
if self.count:
self.count_list.append(sample['counts'])
if self.classification:
self.binary_labels_list.append(sample['binary_labels'])
self.annotations_list.append(sample['annotations'])
def __len__(self):
if self.ram:
out = len(self.images)
else:
if self.holdout:
txt_file = "holdout.txt"
elif self.sat:
txt_file = "satellite.txt"
else:
txt_file = "train.txt"
with open(os.path.join(self.root_dir,txt_file)) as f:
train_ims = f.readlines()
out = len(train_ims)
return out
def __getitem__(self, idx):
if torch.is_tensor(idx):
idx = idx.tolist()
sample = {}
# yolo object annotation columns
# <object-class> <x> <y> <width> <height>
if self.ram:
sample['image'] = self.images[idx]
if self.density and not self.count:
sample['density'] = self.density_list[idx]
sample['point_map'] = self.point_maps[idx]
sample['labels'] = self.labels_list[idx]
if self.count:
sample['counts'] = self.count_list[idx]
if self.classification:
sample['binary_labels'] = self.binary_labels_list[idx]
sample['annotations'] = self.annotations_list[idx]
else:
sample = self.compute_labels(idx,resize=self.resize)
if self.transform:
sample = self.transform(sample)
return sample
def get_annotations(self, idx):
"""retrieve just the annotations associated with the sample"""
if torch.is_tensor(idx):
idx = idx.tolist()
img_path = os.path.join(self.root_dir,
self.im_paths[idx])
txt_path = img_path[:-3]+'txt'
# yolo object annotation columns
# <object-class> <x> <y> <width> <height>
# check if annotations file is empty
header_list = ['class', 'x', 'y', 'width', 'height']
if os.stat(txt_path).st_size == 0:
annotations = np.array([])
else:
annotations = pd.read_csv(txt_path,names=header_list,delim_whitespace=True)
annotations = annotations.to_numpy()
return annotations
# Helper function to show a batch
def show_annotations_batch(self,sample_batched,debug = False):
"""Show image with annotations for a batch of samples."""
batch_size = len(sample_batched[0])
if self.density:
ncol = 2
else:
ncol = 1
fig, ax = plt.subplots(batch_size,ncol,figsize=(8,13))
plt.ioff()
fig.suptitle('Batch from dataloader',y=0.95,fontsize=20)
#fig.set_size_inches(8*ncol,6*batch_size)
#fig.set_dpi(100)
images_batch = sample_batched[0]
if self.density:
density_batch = sample_batched[1]
for i in range(batch_size):
density = density_batch[i]
image = images_batch[i]
image = image.permute((1, 2, 0)).squeeze().cpu().numpy()
ax[i,0].axis('off')
ax[i,1].axis('off')
ax[i,0].imshow(density, cmap='hot', interpolation='nearest')
ax[i,1].imshow((image * 255).astype(np.uint8))
else:
annotations_batch = sample_batched[1]
for i in range(batch_size):
grid = utils.make_grid(images_batch[i])
ax[i].imshow(grid.numpy().transpose((1, 2, 0)))
ax[i].axis('off')
for i in range(batch_size):
# sample_batched: image, annotations, classes
an = sample_batched[1][i]
if len(an) != 0:
an = an.numpy()
if not self.points:
for j in range(0,len(an)):
x = an[j,0]*c.img_size[0]-0.5*an[j,2]*c.img_size[0]
y = an[j,1]*c.img_size[1]-0.5*an[j,3]*c.img_size[1]
rect = patches.Rectangle(xy = (x,y),
width = an[j,2]*c.img_size[0],
height = an[j,3]*c.img_size[1],
linewidth=1, edgecolor='r', facecolor='none')
# Add the patch to the Axes
ax[i].add_patch(rect)
else:
num_cols = an.shape[1]
if num_cols != 2:
ax[i].scatter(an[:,1]*c.img_size[0],an[:,2]*c.img_size[1], s=mk_size,color = 'red')
else:
ax[i].scatter(an[:,0]*c.img_size[0],an[:,1]*c.img_size[1], s=mk_size,color = 'red')
plt.show()
# helper function to show annotations + example
def show_annotations(self,sample_no,title = ""):
sample = self[sample_no]
if self.count:
print("Count:")
print(sample["counts"])
return
# TODO - plot annotations as well (i.e. pre dmap)
if self.density:
im = sample['image']
dmap = sample['density'].cpu().numpy()
unnorm = UnNormalize(mean =tuple(c.norm_mean),std=tuple(c.norm_std))
im = unnorm(im)
im = im.permute(1,2,0).cpu().numpy()
fig, ax = plt.subplots(1,2)
# TODO - bug here! mismatch impaths and data
#fig.suptitle(os.path.basename(os.path.normpath(self.im_paths[sample_no])))
ax[0].imshow(dmap, cmap='viridis', interpolation='nearest')
ax[1].imshow((im * 255).astype(np.uint8))
else:
"""Show image with landmarks"""
image = sample['image']
an = sample['annotations']
fig, ax = plt.subplots(figsize=(3,4))
ax.imshow(image)
# define patch using x1,x2,y1,y2 coords -> map to height and width
# <object-class> <centre-x> <centre-y> <width> <height>
if len(an) != 0:
if not self.points:
for i in range(0,len(an)):
rect = patches.Rectangle(xy = (an[i,1]*c.img_size[0]-0.5*an[i,3]*c.img_size[0],
an[i,2]*c.img_size[1]-0.5*an[i,4]*c.img_size[1]),
width = an[i,3]*c.img_size[0],height = an[i,4]*c.img_size[1],
linewidth=1, edgecolor='r', facecolor='none')
# Add the patch to the Axes
ax.add_patch(rect)
else:
num_cols = an.shape[1]
if num_cols != 2:
ax.scatter(an[:,1]*c.img_size[0],an[:,2]*c.img_size[1], s=mk_size,color = 'red')
else:
ax.scatter(an[:,0]*c.img_size[0],an[:,1]*c.img_size[1], s=mk_size,color = 'red')
# turn off whitespace etc
plt.axis('off')
if self.density:
for i in range(0,len(ax)):
ax[i].get_xaxis().set_visible(False)
ax[i].get_yaxis().set_visible(False)
else:
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
if title != "" and self.density:
ax[0].set_title(title)
elif title != "":
ax.set_title(title)
print('Image size: {}'.format(im.shape))
print('Density map size: {}'.format(dmap.shape))
print('Density map sum: {}'.format(dmap.sum()))
print('Label count: {}'.format(len(self[sample_no]['labels'])))
return
# because batches have varying numbers of bounding boxes, need to define custom collate func
# https://discuss.pytorch.org/t/dataloader-gives-stack-expects-each-tensor-to-be-equal-size-due-to-different-image-has-different-objects-number/91941/5
# method to stack tensors of different sizes:
# https://github.com/sgrvinod/a-PyTorch-Tutorial-to-Object-Detection/blob/master/datasets.py
def custom_collate_aerial(self,batch,debug = False):
# only needed to collate annotations
self.b_images = list()
self.b_density = list()
self.b_point_map = list()
self.b_labels = list()
self.b_annotations = list()
if self.classification:
self.b_binary_labels = list()
if self.count:
self.b_counts = list()
# custom memory pinning method on custom type
def pin_memory(self):
self.b_images = self.b_images.pin_memory()
self.b_density = self.b_density.pin_memory()
self.b_point_map = self.b_point_map.pin_memory()
self.b_labels = self.b_labels.pin_memory()
self.b_annotations = self.b_annotations.pin_memory()
if self.count:
self.b_counts = self.b_counts.pin_memory()
if self.classification:
self.b_binary_labels = self.b_binary_labels.pin_memory()
return self
for b in batch:
self.b_images.append(b['image'])
if 'density' in b.keys():
self.b_density.append(b['density'])
if 'point_map' in b.keys():
self.b_point_map.append(b['point_map'])
if 'labels' in b.keys():
self.b_labels.append(b['labels'])
if 'annotations' in b.keys():
self.b_annotations.append(b['annotations'])
if self.count:
self.b_counts.append(b['counts'])
if self.classification:
self.b_binary_labels.append(b['binary_labels'])
self.b_images = torch.stack(self.b_images,dim = 0)
if 'density' in b.keys():
self.b_density = torch.stack(self.b_density,dim = 0)
if 'point_map' in b.keys():
self.b_point_map = torch.stack(self.b_point_map,dim = 0)
if 'labels' in b.keys():
self.b_labels = self.b_labels
if 'annotations' in b.keys():
self.b_annotations = self.b_annotations
if self.count:
self.b_counts = torch.stack(self.b_counts,dim = 0)
if self.classification:
self.b_binary_labels = torch.stack(self.b_binary_labels,dim = 0)
out = self.b_images,self.b_density,self.b_labels
if self.count:
out = out + (self.b_counts,)
if self.classification:
out = out + (self.b_binary_labels,)
out = out + (self.b_annotations,)
out = out + (self.b_point_map,)
return out
# def custom_collate_aerial(batch):
# return AerialCustomBatch(batch)
# class AerialCustomBatch:
# def __init__(self,data):
# self.b_images = list()
# self.b_densities = list()
# self.b_point_map = list()
# self.b_labels = list()
# self.b_annotations = list()
# self.b_binary_labels = list()
# if self.count:
# self.b_counts = list()
# for b in data:
# self.b_images.append(b['image'])
# if 'density' in b.keys():
# self.b_densities.append(b['density'])
# if 'point_map' in b.keys():
# self.b_point_map.append(b['point_map'])
# if 'labels' in b.keys():
# self.b_labels.append(b['labels'])