-
Notifications
You must be signed in to change notification settings - Fork 0
/
cropmix.py
161 lines (140 loc) · 6.14 KB
/
cropmix.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
# Implementation of CropMix: Sampling a Rich Input Distribution via Multi-Scale Cropping
import random
import numpy as np
import torchvision.transforms as transforms
def mixup(view1, view2, lam, inter_aug):
if inter_aug:
permute = [0, 1, 2]
random.shuffle(permute)
if lam > 0.5:
view1 = lam * view1 + (1 - lam) * view2[permute]
return view1
else:
view1 = lam * view1[permute] + (1 - lam) * view2
return view1
else:
return lam * view1 + (1 - lam) * view2
def cutmix(view1, view2, inter_aug):
lam = np.random.uniform(low=0.0, high=1.0)
lam2 = np.random.uniform(low=0.0, high=1.0)
def random_bbox(lam, H, W):
cut_rat = np.sqrt(1.0 - lam)
cut_w = np.int(W * cut_rat)
cut_h = np.int(H * cut_rat)
cx = np.random.randint(W)
cy = np.random.randint(H)
bbx1 = np.clip(cx - cut_w // 2, 0, W)
bby1 = np.clip(cy - cut_h // 2, 0, H)
bbx2 = np.clip(cx + cut_w // 2, 0, W)
bby2 = np.clip(cy + cut_h // 2, 0, H)
return bbx1, bby1, bbx2, bby2
_, h, w = view1.shape
bbx1, bby1, bbx2, bby2 = random_bbox(lam, h, w)
if inter_aug:
permute = [0, 1, 2]
random.shuffle(permute)
if lam2 > 0.5:
view2 = view2[permute]
view1[:, bbx1:bbx2, bby1:bby2] = view2[:, bbx1:bbx2, bby1:bby2]
return view1
else:
view1 = view1[permute]
view2[:, bbx1:bbx2, bby1:bby2] = view1[:, bbx1:bbx2, bby1:bby2]
return view2
else:
if lam2 > 0.5:
view1[:, bbx1:bbx2, bby1:bby2] = view2[:, bbx1:bbx2, bby1:bby2]
return view1
else:
view2[:, bbx1:bbx2, bby1:bby2] = view1[:, bbx1:bbx2, bby1:bby2]
return view2
class CropMix:
def __init__(self, scale, mix_ratio, number, operation, inter_aug):
self.scale = scale
self.mix_ratio = mix_ratio
self.number = number
self.operation = operation
self.inter_aug = inter_aug
def __call__(self, x):
if self.number == 234:
self.number = random.choice([2, 3, 4])
if self.number == 2:
t1 = transforms.Compose([
transforms.RandomResizedCrop(224, scale=(self.scale, self.scale + (1 - self.scale) / self.number)),
transforms.ToTensor(),
])
t2 = transforms.Compose([
transforms.RandomResizedCrop(224, scale=(self.scale + (1 - self.scale) / self.number, 1)),
transforms.ToTensor(),
])
view1 = t1(x)
view2 = t2(x)
if self.operation == 0:
lam = np.random.beta(self.mix_ratio / self.number, self.mix_ratio / self.number)
mixed = mixup(view1, view2, lam, self.inter_aug)
else:
mixed = cutmix(view1, view2, self.inter_aug)
elif self.number == 3:
t1 = transforms.Compose([
transforms.RandomResizedCrop(224, scale=(self.scale, self.scale + (1 - self.scale) / self.number)),
transforms.ToTensor(),
])
t2 = transforms.Compose([
transforms.RandomResizedCrop(224, scale=(
(self.scale + (1 - self.scale) / self.number), (self.scale + 2 * (1 - self.scale) / self.number))),
transforms.ToTensor(),
])
t3 = transforms.Compose([
transforms.RandomResizedCrop(224, scale=((self.scale + 2 * (1 - self.scale) / self.number), 1)),
transforms.ToTensor(),
])
view1 = t1(x)
view2 = t2(x)
view3 = t3(x)
views = [view1, view2, view3]
random.shuffle(views)
if self.operation == 0:
lam = np.random.beta(self.mix_ratio / self.number, self.mix_ratio / self.number)
mixed = mixup(views[0], views[1], lam, self.inter_aug)
lam = np.random.beta(self.mix_ratio / self.number, self.mix_ratio / self.number)
mixed = mixup(mixed, views[2], lam, self.inter_aug)
else:
mixed = cutmix(views[0], views[1], self.inter_aug)
mixed = cutmix(mixed, views[2], self.inter_aug)
elif self.number == 4:
t1 = transforms.Compose([
transforms.RandomResizedCrop(224, scale=(self.scale, self.scale + (1 - self.scale) / self.number)),
transforms.ToTensor(),
])
t2 = transforms.Compose([
transforms.RandomResizedCrop(224, scale=(
(self.scale + (1 - self.scale) / self.number), (self.scale + 2 * (1 - self.scale) / self.number))),
transforms.ToTensor(),
])
t3 = transforms.Compose([
transforms.RandomResizedCrop(224, scale=(
(self.scale + 2 * (1 - self.scale) / self.number), (self.scale + 3 * (1 - self.scale) / self.number))),
transforms.ToTensor(),
])
t4 = transforms.Compose([
transforms.RandomResizedCrop(224, scale=((self.scale + 3 * (1 - self.scale) / self.number), 1)),
transforms.ToTensor(),
])
view1 = t1(x)
view2 = t2(x)
view3 = t3(x)
view4 = t4(x)
views = [view1, view2, view3, view4]
random.shuffle(views)
if self.operation == 0:
lam = np.random.beta(self.mix_ratio / self.number, self.mix_ratio / self.number)
mixed = mixup(views[0], views[1], lam, self.inter_aug)
lam = np.random.beta(self.mix_ratio / self.number, self.mix_ratio / self.number)
mixed = mixup(mixed, views[2], lam, self.inter_aug)
lam = np.random.beta(self.mix_ratio / self.number, self.mix_ratio / self.number)
mixed = mixup(mixed, views[3], lam, self.inter_aug)
else:
mixed = cutmix(views[0], views[1], self.inter_aug)
mixed = cutmix(mixed, views[2], self.inter_aug)
mixed = cutmix(mixed, views[3], self.inter_aug)
return mixed