-
Notifications
You must be signed in to change notification settings - Fork 1
/
dataset.py
338 lines (281 loc) · 11.5 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
import math
import os
import cv2
import numpy as np
from functools import partial
from numba import jit
from PIL import Image
from torch.utils.data import DataLoader
from torchvision import datasets
import torchvision.transforms as T
import albumentations as A
from albumentations.pytorch import ToTensor
import xml.etree.ElementTree as ET
IMG_SIZE = 64
IMG_SIZE_2 = IMG_SIZE * 2
IMG_EXTENSIONS = ('.jpg', '.jpeg', '.png', '.ppm', '.bmp', '.pgm', '.tif', '.tiff', '.webp')
class MyImg:
def __init__(self, img, tfm):
self.px = np.array(img)
self.tfm = tfm
@property
def size(self):
h, w, _ = self.px.shape
return min(w, h)
def pad(img, padding_mode='reflect'):
p = math.ceil((max(img.size) - min(img.size)) / 2)
p_horr = p if img.width < img.height else 0
p_vert = p if img.height < img.width else 0
img = T.Pad((p_horr, p_vert), padding_mode=padding_mode)(img)
if img.width != img.height:
s = min(img.size)
img = img.crop((0, 0, s, s))
return img
def take_top(img):
size = min(img.size)
bbox = (0, 0, size, size)
return img.crop(bbox)
def take_diagonal(img):
w, h = img.size
size = min(w, h)
bbox_l = (0, 0, size, size)
bbox_r = (w - size, h - size, w, h)
return [img.crop(bbox_l), img.crop(bbox_r)]
resize = T.Resize(IMG_SIZE, interpolation=Image.LANCZOS)
resize2x = T.Resize(IMG_SIZE_2, interpolation=Image.LANCZOS)
center_crop = T.Compose([resize, T.CenterCrop(IMG_SIZE)])
center_crop2x = T.Compose([resize2x, T.CenterCrop(IMG_SIZE_2)])
top_crop = T.Compose([T.Lambda(take_top), resize])
top_crop2x = T.Compose([T.Lambda(take_top), resize2x])
two_crops = T.Compose([resize, T.Lambda(take_diagonal)])
two_crops2x = T.Compose([resize2x, T.Lambda(take_diagonal)])
pad_only = T.Compose([T.Lambda(pad), resize])
pad_only2x = T.Compose([T.Lambda(pad), resize2x])
@jit(nopython=True)
def calc_one_axis(clow, chigh, pad, cmax):
clow = max(0, clow - pad)
chigh = min(cmax, chigh + pad)
return clow, chigh, chigh - clow
def calc_bbox(obj, img_w, img_h, zoom=0.0, try_square=True):
bndbox = obj.find('bndbox')
xmin = int(bndbox.find('xmin').text)
ymin = int(bndbox.find('ymin').text)
xmax = int(bndbox.find('xmax').text)
ymax = int(bndbox.find('ymax').text)
# occasionally i get bboxes which exceed img size
xmin, xmax, obj_w = calc_one_axis(xmin, xmax, 0, img_w)
ymin, ymax, obj_h = calc_one_axis(ymin, ymax, 0, img_h)
if zoom != 0.0:
pad_w = obj_w * zoom / 2
pad_h = obj_h * zoom / 2
xmin, xmax, obj_w = calc_one_axis(xmin, xmax, pad_w, img_w)
ymin, ymax, obj_h = calc_one_axis(ymin, ymax, pad_h, img_h)
if try_square:
# try pad both sides equaly
if obj_w > obj_h:
pad = (obj_w - obj_h) / 2
ymin, ymax, obj_h = calc_one_axis(ymin, ymax, pad, img_h)
elif obj_h > obj_w:
pad = (obj_h - obj_w) / 2
xmin, xmax, obj_w = calc_one_axis(xmin, xmax, pad, img_w)
# if it's still not square, try pad where possible
if obj_w > obj_h:
pad = obj_w - obj_h
ymin, ymax, obj_h = calc_one_axis(ymin, ymax, pad, img_h)
elif obj_h > obj_w:
pad = obj_h - obj_w
xmin, xmax, obj_w = calc_one_axis(xmin, xmax, pad, img_w)
return int(xmin), int(ymin), int(xmax), int(ymax)
@jit(nopython=True)
def bb2wh(bbox):
width = bbox[2] - bbox[0]
height = bbox[3] - bbox[1]
return width, height
def make_x2res(img, bbox):
if min(bb2wh(bbox)) < IMG_SIZE_2:
return
ar = img.width / img.height
if ar == 1.0:
tfm_img = resize2x(img)
elif 1.0 < ar < 1.15:
tfm_img = center_crop2x(img)
elif 1.15 < ar < 1.25:
tfm_img = pad_only2x(img)
elif 1.25 < ar < 1.5:
tfm_img = two_crops2x(img)
elif 1.0 < 1 / ar < 1.6:
tfm_img = top_crop2x(img)
else:
tfm_img = None
return tfm_img
def add_sample(samples, label, tfm, imgs, labels):
if not samples:
return
elif isinstance(samples, Image.Image):
imgs.append(MyImg(samples, tfm))
labels.append(label)
elif isinstance(samples, list):
imgs.extend([MyImg(s, tfm) for s in samples])
labels.extend([label] * len(samples))
else:
assert False
def is_valid_file(x):
return datasets.folder.has_file_allowed_extension(x, IMG_EXTENSIONS)
class DogsDataSet(datasets.vision.VisionDataset):
def __init__(self, root, label_root, transforms, target_transform=None, max_samples=None):
super().__init__(root, transform=None)
assert isinstance(transforms, list) and len(transforms) == 3
self.transforms = transforms
self.target_transform = target_transform
self.max_samples = max_samples
self.classes = {}
imgs, labels = self._load_subfolders_images(self.root, label_root)
assert len(imgs) == len(labels)
if len(imgs) == 0:
raise RuntimeError(f'Found 0 files in subfolders of: {self.root}')
self.imgs = imgs
self.labels = labels
def _create_or_get_class(self, name):
try:
label = self.classes[name]
except KeyError:
label = len(self.classes)
self.classes[name] = label
return label
def _load_subfolders_images(self, root, label_root):
light_zoom, medium_zoom = 0.08, 0.12
n_pad, n_center, n_top, n_2crops, n_skip, n_dup, n_noop = 0, 0, 0, 0, 0, 0, 0
imgs, labels, paths = [], [], []
add_sample_ = partial(add_sample, imgs=imgs, labels=labels)
for root, _, fnames in sorted(os.walk(root)):
for fname in sorted(fnames):
path = os.path.join(root, fname)
paths.append(path)
if self.max_samples:
paths = paths[:self.max_samples]
for path in paths:
if not is_valid_file(path):
continue
img = datasets.folder.default_loader(path)
annotation_basename = os.path.splitext(os.path.basename(path))[0]
annotation_dirname = next(
dirname for dirname in os.listdir(label_root) if dirname.startswith(annotation_basename.split('_')[0]))
annotation_filename = os.path.join(label_root, annotation_dirname, annotation_basename)
tree = ET.parse(annotation_filename)
root = tree.getroot()
objects = root.findall('object')
for o in objects:
name = o.find('name').text
label = self._create_or_get_class(name)
prev_bbox, tfm_imgs = None, None
bbox = calc_bbox(o, img_w=img.width, img_h=img.height, zoom=light_zoom)
obj_img = img.crop(bbox)
add_sample_(make_x2res(obj_img, bbox), label, 2)
bbox = calc_bbox(o, img_w=img.width, img_h=img.height)
if min(bb2wh(bbox)) < IMG_SIZE:
# don't want pixel mess in gen imgs
n_skip += 1
continue
obj_img = img.crop(bbox)
ar = obj_img.width / obj_img.height
if ar == 1.0:
tfm_imgs = [resize(obj_img)]
n_noop += 1
elif 1.0 < ar < 1.3:
tfm_imgs = [center_crop(obj_img), pad_only(obj_img)]
n_center += 1
n_pad += 1
elif 1.3 <= ar < 1.5:
tfm_imgs = two_crops(obj_img) + [pad_only(obj_img)]
n_2crops += 2
n_pad += 1
elif 1.5 <= ar < 2.0:
tfm_imgs = two_crops(obj_img)
n_2crops += 2
elif 1.0 < 1 / ar < 1.5:
tfm_imgs = [top_crop(obj_img), pad_only(obj_img)]
n_top += 1
n_pad += 1
elif 1.5 <= 1 / ar < 1.8:
tfm_imgs = [top_crop(obj_img)]
n_top += 1
else:
tfm_imgs = None
n_skip += 1
add_sample_(tfm_imgs, label, 0)
add_sample_(make_x2res(obj_img, bbox), label, 1)
prev_bbox = bbox
bbox = calc_bbox(o, img_w=img.width, img_h=img.height, zoom=medium_zoom, try_square=False)
if bbox == prev_bbox:
n_dup += 1
continue
if min(bb2wh(bbox)) < IMG_SIZE_2: continue
obj_img = img.crop(bbox)
ar = obj_img.width / obj_img.height
if 1.3 < ar < 1.5:
tfm_imgs = two_crops(obj_img)
n_2crops += 2
elif 1.05 < 1 / ar < 1.6: # maybe tall
tfm_imgs = top_crop(obj_img)
n_top += 1
else:
continue
add_sample_(tfm_imgs, label, 0)
add_sample_(make_x2res(obj_img, bbox), label, 1)
prev_bbox = bbox
n_x1, n_x2 = 0, 0
for i, img in enumerate(imgs):
if img.size == IMG_SIZE:
n_x1 += 1
else:
n_x2 += 1
print(f'Found {len(self.classes)} classes\nLoaded 64x64 {n_x1} images\n'
f'Loaded 128x128 {n_x2} images\n')
print(f'Pad only: {n_pad}\nCrop center: {n_center}\n'
f'Crop top: {n_top}\nCrop 2 times: {n_2crops}\n'
f'Take as-is: {n_noop}\nSkip: {n_skip}\nSame bbox: {n_dup}')
return imgs, labels
def __getitem__(self, index):
img = self.imgs[index]
label = self.labels[index]
tfms = self.transforms[img.tfm]
img = tfms(image=img.px)['image']
if self.target_transform:
label = self.target_transform(label)
return img, label
def __len__(self):
return len(self.imgs)
def create_runtime_tfms():
mean, std = [0.5] * 3, [0.5] * 3
resize_to_64 = A.SmallestMaxSize(IMG_SIZE, interpolation=cv2.INTER_AREA)
out = [A.HorizontalFlip(p=0.5), A.Normalize(mean=mean, std=std), ToTensor()]
rand_crop = A.Compose([
A.SmallestMaxSize(IMG_SIZE + 8, interpolation=cv2.INTER_AREA),
A.RandomCrop(IMG_SIZE, IMG_SIZE)
])
affine_1 = A.ShiftScaleRotate(
shift_limit=0, scale_limit=0.1, rotate_limit=8,
interpolation=cv2.INTER_CUBIC,
border_mode=cv2.BORDER_REFLECT_101, p=1.0)
affine_1 = A.Compose([affine_1, resize_to_64])
affine_2 = A.ShiftScaleRotate(
shift_limit=0.06, scale_limit=(-0.06, 0.18), rotate_limit=6,
interpolation=cv2.INTER_CUBIC,
border_mode=cv2.BORDER_REFLECT_101, p=1.0)
affine_2 = A.Compose([affine_2, resize_to_64])
tfm_0 = A.Compose(out)
tfm_1 = A.Compose([A.OneOrOther(affine_1, rand_crop, p=1.0), *out])
tfm_2 = A.Compose([affine_2, *out])
return [tfm_0, tfm_1, tfm_2]
def get_data_loaders(data_root=None, label_root=None, batch_size=32, num_workers=2, shuffle=True,
pin_memory=True, drop_last=True):
print('Using dataset root location %s' % data_root)
train_set = DogsDataSet(data_root, label_root, create_runtime_tfms())
# Prepare loader; the loaders list is for forward compatibility with
# using validation / test splits.
loaders = []
loader_kwargs = {'num_workers': num_workers, 'pin_memory': pin_memory,
'drop_last': drop_last} # Default, drop last incomplete batch
train_loader = DataLoader(train_set, batch_size=batch_size, shuffle=shuffle, **loader_kwargs)
loaders.append(train_loader)
return loaders