-
Notifications
You must be signed in to change notification settings - Fork 0
/
datasets.py
1310 lines (1105 loc) · 57.9 KB
/
datasets.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
import os
import random
import imageio.v2 as imageio
import scipy
from PIL import Image, ImageChops, ImageFilter, ImageEnhance, ImageOps, ImageDraw
from scipy.io import loadmat
from torch.utils.data import Dataset
from torchvision import transforms
import torch
import numpy as np
import matplotlib.pyplot as plt
import re
import glob
from image_with_masks_and_attributes import ImageWithMasksAndAttributes
from categories_and_attributes import CategoriesAndAttributes, CelebAMaskHQCategoriesAndAttributes
import json
import matplotlib.patches as patches
from torchvision.transforms import functional as TF
# class CategoryType(enum.Enum):
# selective = 0
# logical = 1
# class DeepFashion2Dataset(Dataset):
# categories = [
# 'short sleeve top', 'long sleeve top', 'short sleeve outwear',
# 'long sleeve outwear', 'vest', 'sling', 'shorts',
# 'trousers', 'skirt', 'short sleeve dress',
# 'long sleeve dress', 'vest dress', 'sling dress'
# ]
# def __init__(self, image_dir, anno_dir, output_size=(300, 400)):
# self.output_size = (output_size[1], output_size[0]) # input: W, H
# self.image_dir = image_dir
# self.anno_dir = anno_dir
# self.image_filenames = [x for x in os.listdir(image_dir) if x.endswith('.jpg')]
# self.transform = transforms.Compose([
# transforms.ToTensor(), # Converts to Torch tensor and scales to [0, 1]
# ])
#
# def __len__(self):
# return len(self.image_filenames)
#
# def __getitem__(self, idx):
# # Load and transform image
# img_name = self.image_filenames[idx]
# img_path = os.path.join(self.image_dir, img_name)
# image = Image.open(img_path).convert('RGB')
#
# # Determine the crop size based on the aspect ratio of output_size
# crop_height, crop_width = self.get_max_crop(image.size, self.output_size)
#
# # Initialize mask array
# masks = torch.zeros((len(self.categories), *self.output_size), dtype=torch.float32)
# # labels = torch.zeros(len(self.categories), dtype=torch.float32)
#
# # Load annotation
# anno_path = os.path.join(self.anno_dir, img_name.replace('.jpg', '.json'))
# with open(anno_path, 'r', encoding='utf-8') as file:
# anno_data = json.load(file)
#
# # Create masks and determine labels
# for item_key, item_value in anno_data.items():
# if item_key.startswith('item'):
# category_id = item_value.get('category_id', 0) - 1
# if 0 <= category_id < len(self.categories):
# mask = Image.new('L', image.size, 0)
# draw = ImageDraw.Draw(mask)
# for segment in item_value.get('segmentation', []):
# draw.polygon(segment, fill=255)
# mask = TF.center_crop(mask, (crop_height, crop_width))
# mask = TF.resize(mask, self.output_size)
# mask_tensor = self.transform(mask).squeeze(0) # Convert to tensor and remove channel dim
# masks[category_id] = torch.max(masks[category_id], mask_tensor) # Use torch.max to combine masks
#
# # Crop the image
# image = TF.center_crop(image, (crop_height, crop_width))
#
# # Resize image to output size
# image = TF.resize(image, self.output_size)
# image_tensor = self.transform(image)
#
# # Update labels based on mask presence after cropping
# labels = (torch.sum(masks.reshape(len(self.categories), -1), dim=1) > 0).float()
#
# return image_tensor, masks, labels
#
# @staticmethod
# def get_max_crop(current_size, target_ratio):
# current_ratio = current_size[0] / current_size[1]
# target_ratio = target_ratio[0] / target_ratio[1]
# if current_ratio > target_ratio:
# # Crop width to fit target ratio
# return int(current_size[1] * target_ratio), current_size[1]
# else:
# # Crop height to fit target ratio
# return current_size[0], int(current_size[0] / target_ratio)
class DeepFashion2Dataset(Dataset):
categories = [
'short sleeve top', 'long sleeve top', 'short sleeve outwear',
'long sleeve outwear', 'vest', 'sling', 'shorts',
'trousers', 'skirt', 'short sleeve dress',
'long sleeve dress', 'vest dress', 'sling dress'
]
general_categories = ['top', 'down', 'outwear', 'dress', ]
general_categories_dict = {
'top': ['short sleeve top', 'long sleeve top', 'vest', 'sling', ],
'down': ['shorts', 'trousers', 'skirt', ],
'outwear': ['short sleeve outwear', 'long sleeve outwear', ],
'dress': ['short sleeve dress', 'long sleeve dress', 'vest dress', 'sling dress', ],
}
def __init__(self, image_dir, anno_dir, output_size=(300, 400), return_bbox=True):
self.output_size = output_size # input: W, H
self.image_dir = image_dir
self.anno_dir = anno_dir
self.image_filenames = [x for x in os.listdir(image_dir) if x.endswith('.jpg')]
self.transform = transforms.Compose([
transforms.ToTensor(), # Converts to Torch tensor and scales to [0, 1]
])
self.return_bbox = return_bbox
def __len__(self):
return len(self.image_filenames)
def __getitem__(self, idx):
# Load and transform image
img_name = self.image_filenames[idx]
img_path = os.path.join(self.image_dir, img_name)
image = Image.open(img_path).convert('RGB')
# Determine the crop size based on the aspect ratio of output_size
crop_width, crop_height = self.get_max_crop(image.size, self.output_size)
# Load annotation
anno_path = os.path.join(self.anno_dir, img_name.replace('.jpg', '.json'))
with open(anno_path, 'r', encoding='utf-8') as file:
anno_data = json.load(file)
# Initialize mask array and bbox list
masks = torch.zeros((len(self.categories), self.output_size[1], self.output_size[0]), dtype=torch.float32)
bboxes = []
# Process each item in the annotation
for item_key, item_value in anno_data.items():
if item_key.startswith('item'):
category_id = item_value.get('category_id', 0) - 1
bbox = item_value.get('bounding_box', [])
if 0 <= category_id < len(self.categories):
# Create mask for current item
mask = Image.new('L', image.size, 0)
draw = ImageDraw.Draw(mask)
for segment in item_value.get('segmentation', []):
draw.polygon(segment, fill=255)
mask = TF.center_crop(mask, (crop_height, crop_width))
mask = TF.resize(mask, (self.output_size[1], self.output_size[0]))
mask_tensor = self.transform(mask).squeeze(0) # Convert to tensor and remove channel dim
masks[category_id] = torch.max(masks[category_id], mask_tensor) # Use torch.max to combine masks
# Convert absolute bbox coordinates to relative based on crop and resize
rel_bbox = self.convert_bbox(bbox, image.size, (crop_width, crop_height), self.output_size)
bboxes.append((category_id, rel_bbox))
# Crop the image
image = TF.center_crop(image, (crop_height, crop_width))
image = TF.resize(image, (self.output_size[1], self.output_size[0]))
image_tensor = self.transform(image)
# Update labels based on mask presence after cropping
labels = (torch.sum(masks.reshape(len(self.categories), -1), dim=1) > 0).float()
if self.return_bbox:
return image_tensor, masks, labels, bboxes
else:
return image_tensor, masks, labels
def convert_bbox(self, bbox, orig_size, crop_size, output_size):
x1, y1, x2, y2 = bbox
orig_width, orig_height = orig_size
crop_width, crop_height = crop_size
output_width, output_height = output_size
# Calculate crop offsets (assuming center crop)
x_offset = (orig_width - crop_width) // 2
y_offset = (orig_height - crop_height) // 2
# Adjust bbox coordinates to cropped image
x1_cropped = x1 - x_offset
y1_cropped = y1 - y_offset
x2_cropped = x2 - x_offset
y2_cropped = y2 - y_offset
# Scale bbox coordinates to output size
x1_relative = x1_cropped / crop_width
y1_relative = y1_cropped / crop_height
x2_relative = x2_cropped / crop_width
y2_relative = y2_cropped / crop_height
return [x1_relative, y1_relative, x2_relative, y2_relative]
@staticmethod
def get_max_crop(current_size, target_size):
current_ratio = current_size[0] / current_size[1]
target_ratio = target_size[0] / target_size[1]
if current_ratio > target_ratio:
# Crop width to fit target ratio
return int(current_size[1] * target_ratio), current_size[1]
else:
# Crop height to fit target ratio
return current_size[0], int(current_size[0] / target_ratio)
class DeepFashion2DatasetGeneral(DeepFashion2Dataset):
def __getitem__(self, idx):
# Start by using the parent class method to load basic data
img_name = self.image_filenames[idx]
img_path = os.path.join(self.image_dir, img_name)
image = Image.open(img_path).convert('RGB')
crop_width, crop_height = self.get_max_crop(image.size, self.output_size)
anno_path = os.path.join(self.anno_dir, img_name.replace('.jpg', '.json'))
with open(anno_path, 'r', encoding='utf-8') as file:
anno_data = json.load(file)
# Initialize mask arrays and bbox list for all specific and general categories
masks = torch.zeros(
(len(self.categories) + len(self.general_categories), self.output_size[1], self.output_size[0]),
dtype=torch.float32)
general_masks = torch.zeros((len(self.general_categories), self.output_size[1], self.output_size[0]),
dtype=torch.float32)
general_bboxes = [(i, [0, 0, 0, 0]) for i in
range(len(self.general_categories))] # Initialize with zero-width bboxes
unwritten_bboxes = [i for i in range(len(self.general_categories))]
# Process each item in the annotation
for item_key, item_value in anno_data.items():
if item_key.startswith('item'):
category_id = item_value.get('category_id', 0) - 1
bbox = item_value.get('bounding_box', [])
if 0 <= category_id < len(self.categories):
# Create and process mask for current item
mask = Image.new('L', image.size, 0)
draw = ImageDraw.Draw(mask)
for segment in item_value.get('segmentation', []):
draw.polygon(segment, fill=255)
mask = TF.center_crop(mask, (crop_height, crop_width))
mask = TF.resize(mask, (self.output_size[1], self.output_size[0]))
mask_tensor = self.transform(mask).squeeze(0) # Convert to tensor and remove channel dim
masks[category_id + len(self.general_categories)] = torch.max(
masks[category_id + len(self.general_categories)], mask_tensor)
# Update general category masks
for i, gen_cat in enumerate(self.general_categories):
if self.categories[category_id] in self.general_categories_dict[gen_cat]:
general_masks[i] = torch.max(general_masks[i], mask_tensor)
# Update general category bounding boxes
bbox = self.convert_bbox(bbox, image.size, (crop_width, crop_height), self.output_size)
x1, y1, x2, y2 = bbox
if i not in unwritten_bboxes:
_, (gx1, gy1, gx2, gy2) = general_bboxes[i]
general_bboxes[i] = (i, [
min(gx1, x1), min(gy1, y1),
max(gx2, x2), max(gy2, y2)
])
else:
general_bboxes[i] = (i, [x1, y1, x2, y2])
# Combine specific and general masks
masks[:len(self.general_categories)] = general_masks
# Crop the image and convert to tensor
image = TF.center_crop(image, (crop_height, crop_width))
image = TF.resize(image, (self.output_size[1], self.output_size[0]))
image_tensor = self.transform(image)
# Update labels based on mask presence after cropping for all categories
labels = (torch.sum(masks.reshape(len(self.categories) + len(self.general_categories), -1), dim=1) > 0).float()
if self.return_bbox:
return image_tensor, masks, labels, general_bboxes
else:
return image_tensor, masks, labels
def collate_fn_DeepFashion2(batch):
images = []
masks = []
labels = []
bboxes = []
for image, mask, label, bbox in batch:
images.append(image)
masks.append(mask)
labels.append(label)
bboxes.append(bbox)
images = torch.stack(images, 0)
masks = torch.stack(masks, 0)
labels = torch.stack(labels, 0)
return images, masks, labels, bboxes
# def show_deepfashion2_image_masks_and_labels(dataset, index):
# # Get the data from the dataset
# image, masks, labels = dataset[index]
#
# # Convert the image tensor to PIL Image for display
# image_pil = transforms.ToPILImage()(image)
#
# # Set up the plot
# num_subplots = len(dataset.categories) + 1
# fig, axs = plt.subplots(1, num_subplots, figsize=(20, 3))
#
# # Plot the original image
# axs[0].imshow(image_pil)
# axs[0].set_title('Original Image')
# axs[0].axis('off')
#
# # Plot each mask
# for i, mask in enumerate(masks):
# axs[i + 1].imshow(mask, cmap='gray', interpolation='none')
# axs[i + 1].set_title(f'{dataset.categories[i]}: {"1" if labels[i] == 1.0 else "0"}')
# axs[i + 1].axis('off')
#
# plt.tight_layout()
# plt.show()
# def show_deepfashion2_image_masks_and_labels(image, masks, labels):
# categories = [
# 'short sleeve top', 'long sleeve top', 'short sleeve outwear',
# 'long sleeve outwear', 'vest', 'sling', 'shorts',
# 'trousers', 'skirt', 'short sleeve dress',
# 'long sleeve dress', 'vest dress', 'sling dress'
# ]
#
# # Convert the image tensor to PIL Image for display
# image_pil = transforms.ToPILImage()(image)
#
# # Set up the plot
# num_subplots = len(labels) + 1
# fig, axs = plt.subplots(1, num_subplots, figsize=(20, 3))
#
# # Plot the original image
# axs[0].imshow(image_pil)
# axs[0].set_title('Original Image')
# axs[0].axis('off')
#
# # Plot each mask
# for i, mask in enumerate(masks):
# axs[i + 1].imshow(mask, cmap='gray', interpolation='none')
# axs[i + 1].set_title(f'{categories[i]}: {"1" if labels[i] == 1.0 else "0"}')
# axs[i + 1].axis('off')
#
# plt.tight_layout()
# plt.show()
def display_deepfashion_image_with_annotations(image_folder, anno_folder):
for filename in os.listdir(anno_folder):
if filename.endswith('.json'):
json_path = os.path.join(anno_folder, filename)
image_path = os.path.join(image_folder, filename.replace('.json', '.jpg'))
try:
# Load JSON data
with open(json_path, 'r', encoding='utf-8') as file:
data = json.load(file)
# Open the corresponding image
image = Image.open(image_path)
fig, ax = plt.subplots()
ax.imshow(image)
# Prepare label text from JSON data and draw annotations
labels = []
for item_key in data:
if item_key.startswith('item'):
item = data[item_key]
category_name = item.get('category_name', 'Unknown')
category_id = item.get('category_id', 'Unknown')
style = item.get('style', 'Unknown')
labels.append(f"Category Name: {category_name}, Category ID: {category_id}, Style: {style}")
# Draw bounding box
bbox = item.get('bounding_box', [])
if bbox:
rect = patches.Rectangle((bbox[0], bbox[1]), bbox[2]-bbox[0], bbox[3]-bbox[1], linewidth=1, edgecolor='r', facecolor='none')
ax.add_patch(rect)
# Draw segmentation polygons
segments = item.get('segmentation', [])
for poly in segments:
poly_points = [(poly[i], poly[i+1]) for i in range(0, len(poly), 2)]
polygon = patches.Polygon(poly_points, linewidth=1, edgecolor='g', facecolor='none')
ax.add_patch(polygon)
# Draw landmarks
landmarks = item.get('landmarks', [])
for i in range(0, len(landmarks), 3):
x, y, v = landmarks[i], landmarks[i+1], landmarks[i+2]
if v == 2: # Visible
ax.plot(x, y, 'bo') # Blue dot for visible
elif v == 1: # Occlusion
ax.plot(x, y, 'yo') # Yellow dot for occlusion
# Show image with labels and annotations
plt.title('\n'.join(labels))
plt.axis('off') # Hide axes
plt.show()
except Exception as e:
print(f"Error processing {filename}: {e}")
class CCPDataset(Dataset):
replay_attributes = []
def __init__(self, root_dir, output_size=(512, 512), # replay=10,
categories_and_attributes: CategoriesAndAttributes = None, pixel_only=False):
self.categories_and_attributes = CelebAMaskHQCategoriesAndAttributes() if categories_and_attributes is None else categories_and_attributes
self.output_size = output_size
self.root_dir = root_dir
self.image_dir = os.path.join(root_dir, "photos")
self.mask_dir = os.path.join(root_dir, "annotations/pixel-level/")
self.label_dir = os.path.join(root_dir, "annotations/image-level/")
self.mask_path_list = glob.glob(os.path.join(self.mask_dir, '*.mat'))
self.label_path_list = glob.glob(os.path.join(self.label_dir, '*.mat'))
self.merged_path_list = self.mask_path_list if pixel_only else self.mask_path_list + self.label_path_list
# self.replay = replay
self.original_length = len(self.merged_path_list)
def __len__(self):
return self.original_length
# def __getitem__(self, idx) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
# idx_path = self.merged_path_list[idx]
# name = os.path.splitext(os.path.basename(idx_path))[0]
# image = imageio.imread(f'photos/{name}.jpg')
# labels = torch.zeros(len(self.categories_and_attributes.attributes))
# labels_str_list = self.categories_and_attributes.attributes
# # converte image into a torch tensor and resize it into output size (self.output_size)
# # convert it into colour channel first
# # make it 0~1
# # init groundtruth an integer tensor of zeros same size to image (but has only one channel)
# if idx_path in self.mask_dir:
# annotation_mat = loadmat('annotations/pixel-level/' + name + '.mat')
# groundtruth = annotation_mat['groundtruth']
# # converte groundtruth into a torch tensor and resize it into output size
# # call get_image_pixel_labels and assign 1s to labels based on its index in labels_str_list
# pixel_labels = torch.tensor(True)
# elif idx_path in self.label_dir:
# # call get_image_labels and assign 1s to labels based on its index in labels_str_list
# pixel_labels = torch.tensor(False)
# pass
def __getitem__(self, idx) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
idx_path = self.merged_path_list[idx]
name = os.path.splitext(os.path.basename(idx_path))[0]
image = imageio.imread(os.path.join(self.image_dir, f'{name}.jpg'))
# Calculate the aspect ratio of the output size
output_aspect_ratio = self.output_size[1] / self.output_size[0]
# Determine the maximum crop size of the original image that maintains this aspect ratio
orig_aspect_ratio = image.shape[1] / image.shape[0]
if orig_aspect_ratio > output_aspect_ratio:
# Width is too wide for the target aspect ratio
crop_height = image.shape[0]
crop_width = int(crop_height * output_aspect_ratio)
else:
# Height is too high for the target aspect ratio
crop_width = image.shape[1]
crop_height = int(crop_width / output_aspect_ratio)
# Randomly choose an offset for the crop (i.e., top-left corner of the crop)
x_offset = random.randint(0, image.shape[1] - crop_width)
y_offset = random.randint(0, image.shape[0] - crop_height)
# Apply crop to the original image
cropped_image = image[y_offset:y_offset + crop_height, x_offset:x_offset + crop_width]
# Image transformations: resize, convert color channel order, and normalize
transform = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize(self.output_size),
transforms.ToTensor(), # Converts to [C, H, W] & scales to [0, 1]
])
image_tensor = transform(cropped_image)
labels = torch.zeros(len(self.categories_and_attributes.attributes))
labels_str_list = self.categories_and_attributes.attributes
# Initialize groundtruth tensor
groundtruth_tensor = torch.zeros(1, *self.output_size,
dtype=torch.long) # Assuming you want integer labels for segmentation
has_pixel_labels = torch.tensor(0)
if idx_path in self.mask_path_list:
annotation_mat = loadmat(os.path.join(self.mask_dir, f'{name}.mat'))
groundtruth = annotation_mat['groundtruth']
cropped_groundtruth = groundtruth[y_offset:y_offset + crop_height, x_offset:x_offset + crop_width]
# Convert the numpy array to a PIL Image for resizing
cropped_groundtruth_image = Image.fromarray(cropped_groundtruth.astype('uint8'))
# Resize using NEAREST to avoid changing class labels unintentionally
resized_groundtruth_image = cropped_groundtruth_image.resize((self.output_size[1], self.output_size[0]), resample=Image.NEAREST)
# Convert the PIL Image back to a numpy array
resized_groundtruth_array = np.array(resized_groundtruth_image)
# Finally, convert the numpy array to a PyTorch tensor
groundtruth_tensor = torch.from_numpy(
resized_groundtruth_array).long() # Ensure it's a long tensor for indexing
# Update labels tensor based on pixel annotations
pixel_labels = self.get_image_pixel_labels(idx_path) # Assuming this function returns a list of label names
for label in pixel_labels:
if label in labels_str_list:
labels[labels_str_list.index(label)] = 1
has_pixel_labels = torch.tensor(1)
elif idx_path in self.label_path_list:
image_labels = self.get_image_labels(idx_path) # Assuming this function returns a list of label names
for label in image_labels:
if label in labels_str_list:
labels[labels_str_list.index(label)] = 1
has_pixel_labels = torch.tensor(0)
num_classes = len(labels)
# Ensure groundtruth_tensor is squeezed in case it has an unnecessary channel dimension
groundtruth_tensor = groundtruth_tensor.squeeze(0) # Assuming groundtruth_tensor is of shape [1, H, W]
# Create one-hot encoding
H, W = groundtruth_tensor.shape
one_hot_groundtruth = torch.zeros((num_classes, H, W), dtype=torch.float32)
one_hot_groundtruth.scatter_(0, groundtruth_tensor.unsqueeze(0), 1)
# # debug:::::
# # Visualization part
# fig, axs = plt.subplots(1, 3, figsize=(18, 6)) # Adjust the size as needed
#
# # Plot original image
# axs[0].imshow(image)
# axs[0].set_title('Original Image')
# axs[0].axis('off')
#
# # Highlight the cropped area in the original image
# rect = patches.Rectangle((x_offset, y_offset), crop_width, crop_height, linewidth=1, edgecolor='r',
# facecolor='none')
# axs[0].add_patch(rect)
#
# # Plot cropped and resized image
# axs[1].imshow(transforms.ToPILImage()(image_tensor))
# axs[1].set_title('Cropped & Resized Image')
# axs[1].axis('off')
#
# # Check and plot ground truth if available
# if has_pixel_labels == torch.tensor(1):
# # Assuming groundtruth_tensor is already in one-hot and resized to self.output_size
# # Convert one-hot back to class indices for visualization
# groundtruth_indices = torch.argmax(one_hot_groundtruth, dim=0)
# axs[2].imshow(groundtruth_indices, cmap='jet')
# axs[2].set_title('Ground Truth Annotation')
# axs[2].axis('off')
# else:
# # axs[2].set_visible(False)
#
# axs[2].text(0.5, 0.5, 'No pixel-level labels', horizontalalignment='center', verticalalignment='center',
# transform=axs[2].transAxes)
# axs[2].axis('off')
#
# # Add texture labels based on 'labels' tensor
# present_labels = [labels_str_list[i] for i, label in enumerate(labels) if label == 1]
# texture_labels_text = "Texture Labels: " + ", ".join(present_labels)
# # fig.suptitle(texture_labels_text, fontsize=14)
# # Adjusting the text to appear like a subtitle below the image in axs[1]
# axs[1].text(0.5, -0.5, texture_labels_text, ha='center', va='top', transform=axs[1].transAxes, fontsize=14)
#
# # axs[1].suptitle(texture_labels_text, fontsize=14)
#
# plt.show()
return image_tensor.to(dtype=torch.float32), one_hot_groundtruth.to(dtype=torch.float32), labels.to(dtype=torch.float32), has_pixel_labels.to(dtype=torch.float32)
def get_image_labels(self, im_file):
"""
Returns a list of labels contained in the given image file.
Parameters:
- im_file: String. The path to the image file's corresponding annotation file.
Returns:
- A list of label names that the image contains.
"""
# Load the label list
label_list_data = scipy.io.loadmat(os.path.join(self.root_dir, 'label_list.mat'))['label_list'][0].tolist()
# Extract the image name from the given file path
name = os.path.splitext(os.path.basename(im_file))[0]
# Load the tags from the annotation file
try:
tags = scipy.io.loadmat(os.path.join(self.label_dir, f'{name}.mat'))['tags'][0]
except FileNotFoundError:
return f"Annotation file for {name} not found."
# Translate tags to label names
label_names = [str(label_list_data[tag][0]) for tag in tags]
return label_names
def get_image_pixel_labels(self, im_file):
"""
Returns a list of labels contained in the given image file based on pixel-level annotations.
Parameters:
- im_file: String. The file name or path to the image's pixel-level annotation file.
Returns:
- A list of label names that the image contains based on pixel-level annotations.
"""
# Load the label list
label_list_data = scipy.io.loadmat(os.path.join(self.root_dir, 'label_list.mat'))['label_list'][0].tolist()
# Extract the base name for the image file
name = os.path.splitext(os.path.basename(im_file))[0]
# Load the pixel-level annotation for the given image
try:
annotation_mat = scipy.io.loadmat(os.path.join(self.mask_dir, f'{name}.mat'))
groundtruth = annotation_mat['groundtruth']
except FileNotFoundError:
return f"Pixel-level annotation file for {name} not found."
# Get unique labels from the groundtruth
cur_labels = np.unique(groundtruth)
# Translate those labels into human-readable names
label_names = [str(label_list_data[int(label)][0]) for label in cur_labels if int(label) < len(label_list_data)]
return label_names
class MergedCCPDataset(CCPDataset):
def __init__(self, root_dir, output_size=(512, 512), # replay=10,
categories_and_attributes: CategoriesAndAttributes = None, pixel_only=False):
super(MergedCCPDataset, self).__init__(
root_dir, output_size, categories_and_attributes=categories_and_attributes, pixel_only=pixel_only,
)
def __getitem__(self, idx):
image, unmerged_masks, attributes, has_pixel_labels = super(MergedCCPDataset, self).__getitem__(idx)
# Convert Tensor images to PIL for operations
image = transforms.ToPILImage()(image)
unmerged_masks = [transforms.ToPILImage()(mask) for mask in unmerged_masks]
masks = []
for category in sorted(list(self.categories_and_attributes.merged_categories.keys())):
combined_mask_np = np.zeros_like(np.array(unmerged_masks[0])) # Initialize with zeros
for sub_category in self.categories_and_attributes.merged_categories[category]:
sub_cat_idx = self.categories_and_attributes.mask_categories.index(sub_category)
mask_to_merge_np = np.array(unmerged_masks[sub_cat_idx])
# Use logical or operation to combine masks
combined_mask_np = np.logical_or(combined_mask_np, mask_to_merge_np).astype(np.uint8)
masks.append(Image.fromarray(combined_mask_np * 255, 'L')) # Convert back to PIL and append
# Convert back to Tensor
image = transforms.ToTensor()(image)
masks = torch.stack([transforms.ToTensor()(m) for m in masks], dim=0).squeeze(1)
return image, masks, attributes, has_pixel_labels
class CelebAMaskHQDataset(Dataset):
replay_attributes = ['Wearing_Hat', 'Eyeglasses', 'Blond_Hair']
def __init__(self, root_dir, output_size=(512, 512), replay=10,
categories_and_attributes: CategoriesAndAttributes = None):
self.categories_and_attributes = CelebAMaskHQCategoriesAndAttributes() if categories_and_attributes is None else categories_and_attributes
self.output_size = output_size
self.root_dir = root_dir
self.image_dir = os.path.join(root_dir, "CelebA-HQ-img")
self.mask_dir = os.path.join(root_dir, "CelebAMask-HQ-mask-anno")
self.replay = replay
# Load image list
self.image_list = sorted([f for f in os.listdir(self.image_dir) if f.endswith('.jpg')])
self.original_length = len(self.image_list)
# Build mask dictionary
self.mask_dict = {}
for folder in range(15): # There are folders from 0 to 14
folder_path = os.path.join(self.mask_dir, str(folder))
for mask_file in os.listdir(folder_path):
match = re.match(r"(\d+)_(\w+)\.png", mask_file)
if match:
img_idx, category = match.groups()
if img_idx not in self.mask_dict:
self.mask_dict[img_idx] = {}
full_path = os.path.join(folder_path, mask_file)
self.mask_dict[img_idx][category] = full_path
# Load attribute data
self.attribute_file = os.path.join(root_dir, "CelebAMask-HQ-attribute-anno.txt")
with open(self.attribute_file, 'r') as file:
lines = file.readlines()
self.attributes = lines[1].split()
self.attribute_data = {}
for line in lines[2:]:
parts = line.strip().split()
filename = parts[0]
attrs = {self.attributes[i]: int(parts[i + 1]) for i in range(len(self.attributes))}
self.attribute_data[filename] = attrs
# replay data:
for _ in range(self.replay):
for atr in self.replay_attributes:
for i in range(self.original_length):
name = self.image_list[i] # .split('.')[0]
if self.has_attribute(name, atr):
self.image_list.append(self.image_list[i])
def __getitem__(self, idx) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
# Load image
img_name = os.path.join(self.image_dir, self.image_list[idx])
image = Image.open(img_name)
# Load masks for this image
img_idx = self.image_list[idx].split('.')[0].zfill(5)
masks = []
for category in self.categories_and_attributes.mask_categories:
if category in self.mask_dict[img_idx]:
mask_path = self.mask_dict[img_idx][category]
mask = Image.open(mask_path).convert('L') # Convert to grayscale
# desired_size = image.size[:2] # Get the size (1024, 1024) from (1024, 1024, 3)
mask = mask.resize(self.output_size) # Resize to desired size
masks.append(mask)
else:
# If mask of this category is not available, use a black image
# print(category)
masks.append(Image.new('L', image.size).resize(self.output_size))
image = image.resize(self.output_size)
attributes = []
for attribute in self.categories_and_attributes.attributes:
if attribute in self.categories_and_attributes.avoided_attributes:
continue
if self.has_attribute(self.image_list[idx], attribute):
attributes.append(1.0)
else:
attributes.append(0.0)
for attribute in self.categories_and_attributes.mask_labels:
mask = masks[sorted(list(self.categories_and_attributes.merged_categories.keys())).index(attribute)]
label = transforms.ToTensor()(mask).any(dim=-1).any(dim=-1).float()
attributes.append(label)
# Convert back to Tensor
image = transforms.ToTensor()(image)
masks = torch.stack([transforms.ToTensor()(m) for m in masks], dim=0).squeeze(1)
attributes = torch.tensor(attributes, dtype=torch.float)
return image, masks, attributes
def get_image(self, idx: int):
image_tensor, masks_tensor, attributes_tensor = self.__getitem__(idx)
image_np, masks_np, attributes_np = np.array(image_tensor), np.array(masks_tensor), np.array(attributes_tensor)
masks = {}
for i, category in enumerate(self.categories_and_attributes.mask_categories):
masks[category] = masks_np[i]
attributes = {}
c = 0
for i, attribute in enumerate(self.categories_and_attributes.attributes):
if attribute in self.categories_and_attributes.avoided_attributes:
c += 1
continue
attributes[attribute] = float(attributes_np[i - c])
return ImageWithMasksAndAttributes(image_np, masks, attributes)
def __len__(self):
return len(self.image_list)
def has_attribute(self, filename, attribute):
"""Check if the specified image has the specified attribute"""
if filename in self.attribute_data and attribute in self.attribute_data[filename]:
return self.attribute_data[filename][attribute] == 1
return False
class SelectedCelebAMaskHQDataset(CelebAMaskHQDataset):
selected_categories = ['cloth', 'hair', 'hat', 'eye_g', 'skin', ]
indices = [0, 2, 9, 10, 8, ]
def __init__(self, root_dir, output_size=(512, 512), replay=10, categories_and_attributes=None):
super(SelectedCelebAMaskHQDataset, self).__init__(root_dir, output_size, replay=replay,
categories_and_attributes=categories_and_attributes)
self.selected_indices = [self.categories_and_attributes.mask_categories.index(cat) for cat in
self.selected_categories]
def __getitem__(self, idx):
image, masks_all, attributes = super(SelectedCelebAMaskHQDataset, self).__getitem__(idx)
masks_selected = masks_all[self.selected_indices]
return image, masks_selected, attributes
class MergedCelebAMaskHQDataset(CelebAMaskHQDataset): # 'brow', 'eye', 'mouth', 'nose', ]
def __init__(self, root_dir, output_size=(512, 512), replay=10, categories_and_attributes=None):
super(MergedCelebAMaskHQDataset, self).__init__(root_dir, output_size, replay=replay,
categories_and_attributes=categories_and_attributes)
def __getitem__(self, idx):
image, unmerged_masks, attributes = super(MergedCelebAMaskHQDataset, self).__getitem__(idx)
# Convert Tensor images to PIL for operations
image = transforms.ToPILImage()(image)
unmerged_masks = [transforms.ToPILImage()(mask) for mask in unmerged_masks]
masks = []
for category in sorted(list(self.categories_and_attributes.merged_categories.keys())):
combined_mask_np = np.zeros_like(np.array(unmerged_masks[0])) # Initialize with zeros
for sub_category in self.categories_and_attributes.merged_categories[category]:
sub_cat_idx = self.categories_and_attributes.mask_categories.index(sub_category)
mask_to_merge_np = np.array(unmerged_masks[sub_cat_idx])
# Use logical or operation to combine masks
combined_mask_np = np.logical_or(combined_mask_np, mask_to_merge_np).astype(np.uint8)
masks.append(Image.fromarray(combined_mask_np * 255, 'L')) # Convert back to PIL and append
# Convert back to Tensor
image = transforms.ToTensor()(image)
masks = torch.stack([transforms.ToTensor()(m) for m in masks], dim=0).squeeze(1)
return image, masks, attributes
def get_image(self, idx: int):
image_tensor, masks_tensor, attributes_tensor = self.__getitem__(idx)
image_np, masks_np, attributes_np = np.array(image_tensor), np.array(masks_tensor), np.array(attributes_tensor)
masks = {}
for i, category in enumerate(sorted(list(self.categories_and_attributes.merged_categories.keys()))):
masks[category] = masks_np[i]
attributes = {}
c = 0
for i, attribute in enumerate(self.categories_and_attributes.attributes):
if attribute in self.categories_and_attributes.avoided_attributes:
c += 1
continue
attributes[attribute] = float(attributes_np[i - c])
return ImageWithMasksAndAttributes(image_np, masks, attributes)
class AugmentedDataset(Dataset):
def __init__(self, dataset_source: Dataset, flip_prob=0.5, crop_ratio=(0.8, 0.8), scale_factor=(0.5, 2),
output_size=(512, 512),
noise_level=(0, 10), blur_radius=(0, 2), brightness_factor=(0.75, 1.25), pil=False, seed: int = None):
self.source = dataset_source
np.random.seed(seed)
self.flip_prob = flip_prob
self.crop_ratio = crop_ratio
self.output_size = output_size
self.noise_level = noise_level
self.blur_radius = blur_radius
self.brightness_factor = brightness_factor
self.scale_factor = scale_factor
self.pil = pil
def _get_transform_params(self):
new_crop_width_ratio = np.random.uniform(self.crop_ratio[0], 1)
new_crop_height_ratio = np.random.uniform(self.crop_ratio[1], 1)
new_crop_left_ratio = np.random.uniform(0, 1 - new_crop_width_ratio)
new_crop_top_ratio = np.random.uniform(0, 1 - new_crop_height_ratio)
params = {
"do_flip": np.random.rand() < self.flip_prob,
"crop_width_ratio": new_crop_width_ratio,
"crop_height_ratio": new_crop_height_ratio,
"crop_left_ratio": new_crop_left_ratio,
"crop_top_ratio": new_crop_top_ratio,
"scale_factor": np.random.uniform(self.scale_factor[0], self.scale_factor[1]),
}
if params["scale_factor"] < 1:
# Generate random padding offsets for scale_factor < 1
total_padding_width = self.output_size[0] - int(self.output_size[0] * params["scale_factor"])
total_padding_height = self.output_size[1] - int(self.output_size[1] * params["scale_factor"])
params["padding_left"] = np.random.randint(0, total_padding_width + 1)
params["padding_top"] = np.random.randint(0, total_padding_height + 1)
else:
# Generate random crop offsets for scale_factor > 1
extra_width = int(self.output_size[0] * params["scale_factor"]) - self.output_size[0]
extra_height = int(self.output_size[1] * params["scale_factor"]) - self.output_size[1]
params["crop_left"] = np.random.randint(0, extra_width + 1)
params["crop_top"] = np.random.randint(0, extra_height + 1)
return params
def _apply_common_transforms(self, image, params):
# Convert Tensor to PIL Image for transformations
image = transforms.ToPILImage()(image)
# Resize to output size
if not self.pil:
image = image.resize(self.output_size)
# Use the stored params instead of generating new ones for flip
if params["do_flip"]:
image = ImageOps.mirror(image)
new_width = int(image.width * params["crop_width_ratio"])
new_height = int(image.height * params["crop_height_ratio"])
left = int(image.width * params["crop_left_ratio"])
top = int(image.height * params["crop_top_ratio"])
image = image.crop((left, top, left + new_width, top + new_height))
# Resize to output size (again!) after cropping
image = image.resize(self.output_size)
# Apply scale
scale_factor = params["scale_factor"]
scaled_size = (int(self.output_size[0] * scale_factor), int(self.output_size[1] * scale_factor))
image = image.resize(scaled_size, resample=Image.BILINEAR)
if scale_factor < 1:
# Calculate padding sizes based on the random offsets generated
padding_l = params["padding_left"]
padding_t = params["padding_top"]
padding_r = self.output_size[0] - scaled_size[0] - padding_l
padding_b = self.output_size[1] - scaled_size[1] - padding_t
# Calculate random color for padding
random_color = self._get_random_colour(image)
# Apply padding
image = ImageOps.expand(image, border=(padding_l, padding_t, padding_r, padding_b), fill=random_color)
elif scale_factor > 1:
# Calculate the crop positions based on the random offsets generated
crop_left = params["crop_left"]
crop_top = params["crop_top"]
crop_right = crop_left + self.output_size[0]
crop_bottom = crop_top + self.output_size[1]
# Crop the image
image = image.crop((crop_left, crop_top, crop_right, crop_bottom))
return image
def _apply_image_transforms(self, image, params):
image = self._apply_common_transforms(image, params)
ori_image = image.copy()
# Add noise
noise_level = np.random.uniform(self.noise_level[0], self.noise_level[1])
if noise_level > 0:
noise = np.random.normal(0, noise_level, (image.height, image.width, 3)).astype(np.uint8)
noise_image = Image.fromarray(noise, 'RGB')
image = ImageChops.add(image, noise_image)
# Add blur
blur_value = np.random.uniform(self.blur_radius[0], self.blur_radius[1])
if blur_value > 0:
image = image.filter(ImageFilter.GaussianBlur(radius=blur_value))
# Adjust brightness
brightness_value = np.random.uniform(self.brightness_factor[0], self.brightness_factor[1])
enhancer = ImageEnhance.Brightness(image)
image = enhancer.enhance(brightness_value)
return image, ori_image
def __getitem__(self, idx):
# Unpack the first three known returns and capture any additional ones in 'extra_returns'
image, masks, attributes, *extra_returns = self.source.__getitem__(idx)
# Generate transform params for this sample
transform_params = self._get_transform_params()
# Apply the same augmentations to image and masks using the generated params
image, ori_image = self._apply_image_transforms(image, transform_params)
masks = [self._apply_common_transforms(mask, transform_params) for mask in masks]
image = transforms.ToTensor()(image)
# ori_image = transforms.ToTensor()(ori_image) # Commented out as per your code
masks = torch.stack([transforms.ToTensor()(m) for m in masks], dim=0).squeeze(1)
# Return the processed image, masks, attributes, and any extra returns
return image, masks, attributes, *extra_returns
def __len__(self):
return len(self.source)
def _get_random_colour(self, image):
mode = image.mode
if mode == 'RGB':
return tuple(np.random.randint(0, 256, size=3).tolist())
elif mode == 'RGBA':
return tuple(np.random.randint(0, 256, size=4).tolist())
elif mode == 'L':
# For Gray, ZERO!!!