-
Notifications
You must be signed in to change notification settings - Fork 0
/
mixup.py
162 lines (150 loc) · 7.57 KB
/
mixup.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
import numpy as np
import tensorflow as tf
def one_hot(x, num_classes, on_value=1., off_value=0.):
return tf.one_hot(x, num_classes, on_value=on_value, off_value=off_value)
def mixup_target(target, num_classes, lam=1., smoothing=0.0):
off_value = smoothing / num_classes
on_value = 1. - smoothing + off_value
y1 = one_hot(target, num_classes, on_value=on_value, off_value=off_value)
y2 = one_hot(tf.reverse(target, axis=[0]), num_classes, on_value=on_value, off_value=off_value)
return y1 * lam + y2 * (1. - lam)
def rand_bbox(img_shape, lam, margin=0., count=None):
ratio = np.sqrt(1 - lam)
img_h, img_w = img_shape[-2:]
cut_h, cut_w = int(img_h * ratio), int(img_w * ratio)
margin_y, margin_x = int(margin * cut_h), int(margin * cut_w)
cy = np.random.randint(0 + margin_y, img_h - margin_y, size=count)
cx = np.random.randint(0 + margin_x, img_w - margin_x, size=count)
yl = np.clip(cy - cut_h // 2, 0, img_h)
yh = np.clip(cy + cut_h // 2, 0, img_h)
xl = np.clip(cx - cut_w // 2, 0, img_w)
xh = np.clip(cx + cut_w // 2, 0, img_w)
return yl, yh, xl, xh
def rand_bbox_minmax(img_shape, minmax, count=None):
assert len(minmax) == 2
img_h, img_w = img_shape[-2:]
cut_h = np.random.randint(int(img_h * minmax[0]), int(img_h * minmax[1]), size=count)
cut_w = np.random.randint(int(img_w * minmax[0]), int(img_w * minmax[1]), size=count)
yl = np.random.randint(0, img_h - cut_h, size=count)
xl = np.random.randint(0, img_w - cut_w, size=count)
yu = yl + cut_h
xu = xl + cut_w
return yl, yu, xl, xu
def cutmix_bbox_and_lam(img_shape, lam, ratio_minmax=None, correct_lam=True, count=None):
if ratio_minmax is not None:
yl, yu, xl, xu = rand_bbox_minmax(img_shape, ratio_minmax, count=count)
else:
yl, yu, xl, xu = rand_bbox(img_shape, lam, count=count)
if correct_lam or ratio_minmax is not None:
bbox_area = (yu - yl) * (xu - xl)
lam = 1. - bbox_area / float(img_shape[-2] * img_shape[-1])
return (yl, yu, xl, xu), lam
class Mixup:
def __init__(self, mixup_alpha=1., cutmix_alpha=0., cutmix_minmax=None, prob=1.0, switch_prob=0.5,
mode='batch', correct_lam=True, label_smoothing=0.1, num_classes=1000):
self.mixup_alpha = mixup_alpha
self.cutmix_alpha = cutmix_alpha
self.cutmix_minmax = cutmix_minmax
if self.cutmix_minmax is not None:
assert len(self.cutmix_minmax) == 2
self.cutmix_alpha = 1.0
self.mix_prob = prob
self.switch_prob = switch_prob
self.label_smoothing = label_smoothing
self.num_classes = num_classes
self.mode = mode
self.correct_lam = correct_lam
self.mixup_enabled = True
def _params_per_elem(self, batch_size):
lam = np.ones(batch_size, dtype=np.float32)
use_cutmix = np.zeros(batch_size, dtype=np.bool)
if self.mixup_enabled:
if self.mixup_alpha > 0. and self.cutmix_alpha > 0.:
use_cutmix = np.random.rand(batch_size) < self.switch_prob
lam_mix = np.where(
use_cutmix,
np.random.beta(self.cutmix_alpha, self.cutmix_alpha, size=batch_size),
np.random.beta(self.mixup_alpha, self.mixup_alpha, size=batch_size))
elif self.mixup_alpha > 0.:
lam_mix = np.random.beta(self.mixup_alpha, self.mixup_alpha, size=batch_size)
elif self.cutmix_alpha > 0.:
use_cutmix = np.ones(batch_size, dtype=np.bool)
lam_mix = np.random.beta(self.cutmix_alpha, self.cutmix_alpha, size=batch_size)
else:
assert False, "One of mixup_alpha > 0., cutmix_alpha > 0., cutmix_minmax not None should be true."
lam = np.where(np.random.rand(batch_size) < self.mix_prob, lam_mix.astype(np.float32), lam)
return lam, use_cutmix
def _params_per_batch(self):
lam = 1.
use_cutmix = False
if self.mixup_enabled and np.random.rand() < self.mix_prob:
if self.mixup_alpha > 0. and self.cutmix_alpha > 0.:
use_cutmix = np.random.rand() < self.switch_prob
lam_mix = np.random.beta(self.cutmix_alpha, self.cutmix_alpha) if use_cutmix else \
np.random.beta(self.mixup_alpha, self.mixup_alpha)
elif self.mixup_alpha > 0.:
lam_mix = np.random.beta(self.mixup_alpha, self.mixup_alpha)
elif self.cutmix_alpha > 0.:
use_cutmix = True
lam_mix = np.random.beta(self.cutmix_alpha, self.cutmix_alpha)
else:
assert False, "One of mixup_alpha > 0., cutmix_alpha > 0., cutmix_minmax not None should be true."
lam = float(lam_mix)
return lam, use_cutmix
def _mix_elem(self, x):
batch_size = len(x)
lam_batch, use_cutmix = self._params_per_elem(batch_size)
x_orig = tf.identity(x) # need to keep an unmodified original for mixing source
for i in range(batch_size):
j = batch_size - i - 1
lam = lam_batch[i]
if lam != 1.:
if use_cutmix[i]:
(yl, yh, xl, xh), lam = cutmix_bbox_and_lam(
x[i].shape, lam, ratio_minmax=self.cutmix_minmax, correct_lam=self.correct_lam)
x[i][..., yl:yh, xl:xh] = x_orig[j][..., yl:yh, xl:xh]
lam_batch[i] = lam
else:
x[i] = x[i] * lam + x_orig[j] * (1 - lam)
return tf.convert_to_tensor(lam_batch, dtype=x.dtype)
def _mix_pair(self, x):
batch_size = len(x)
lam_batch, use_cutmix = self._params_per_elem(batch_size // 2)
x_orig = tf.identity(x) # need to keep an unmodified original for mixing source
for i in range(batch_size // 2):
j = batch_size - i - 1
lam = lam_batch[i]
if lam != 1.:
if use_cutmix[i]:
(yl, yh, xl, xh), lam = cutmix_bbox_and_lam(
x[i].shape, lam, ratio_minmax=self.cutmix_minmax, correct_lam=self.correct_lam)
x[i][:, yl:yh, xl:xh] = x_orig[j][:, yl:yh, xl:xh]
x[j][:, yl:yh, xl:xh] = x_orig[i][:, yl:yh, xl:xh]
lam_batch[i] = lam
else:
x[i] = x[i] * lam + x_orig[j] * (1 - lam)
x[j] = x[j] * lam + x_orig[i] * (1 - lam)
lam_batch = np.concatenate((lam_batch, lam_batch[::-1]))
return tf.convert_to_tensor(lam_batch, dtype=x.dtype)
def _mix_batch(self, x):
lam, use_cutmix = self._params_per_batch()
if lam == 1.:
return 1.
if use_cutmix:
(yl, yh, xl, xh), lam = cutmix_bbox_and_lam(
x.shape, lam, ratio_minmax=self.cutmix_minmax, correct_lam=self.correct_lam)
x[..., yl:yh, xl:xh] = tf.reverse(x, axis=[0])[..., yl:yh, xl:xh]
else:
x_flipped = tf.reverse(x, axis=[0]) * (1. - lam)
x = x * lam + x_flipped
return lam
def __call__(self, x, target):
assert len(x) % 2 == 0, 'Batch size should be even when using this'
if self.mode == 'elem':
lam = self._mix_elem(x)
elif self.mode == 'pair':
lam = self._mix_pair(x)
else:
lam = self._mix_batch(x)
target = mixup_target(target, self.num_classes, lam, self.label_smoothing)
return x, target