-
Notifications
You must be signed in to change notification settings - Fork 5
/
tools.py
215 lines (198 loc) · 8.91 KB
/
tools.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
import tensorflow as tf
from tf_image.application.augmentation_config import ColorAugmentation, AugmentationConfig, AspectRatioAugmentation
from tf_image.core.bboxes.clip import clip_random_with_bboxes
from tf_image.core.bboxes.erase import multiple_erase, calculate_bboxes_max_erase_area
from tf_image.core.bboxes.flip import flip_left_right, flip_up_down
from tf_image.core.bboxes.resize import random_aspect_ratio_deformation, random_pad_to_square
from tf_image.core.bboxes.rotate import random_rotate, rot90, rot45
from tf_image.core.clip import clip_random
from tf_image.core.colors import channel_drop, grayscale, channel_swap, rgb_shift
from tf_image.core.convert_type_decorator import convert_type
from tf_image.core.quality import gaussian_noise
from tf_image.core.random import random_function
from tf_image.core.random import random_function_bboxes
def random_augmentations(image, augmentation_config: AugmentationConfig, bboxes=None, prob_demanding_ops: float = 0.5):
"""
Apply augmentations in random order.
WARNING: this is just a testing class and it is likely to change.
:param image: 3-D Tensor of shape (height, width, channels).
:param augmentation_config: Config defining which augmentations can be applied.
:param bboxes: 2-D Tensor of shape (box_number, 4) containing bounding boxes in format [ymin, xmin, ymin, xmax]
:param prob_demanding_ops: Probability that a time consuming operation (like rotation) will be performed.
:return: augmented image or (augmented image, bboxes) if bboxes parameter is not None
"""
has_bboxes = bboxes is not None
if not has_bboxes:
bboxes = tf.reshape([], (0, 4))
# convert_dtype decorator needs this special argument order (converting now saves us converting in each operation)
image, bboxes = _random_augmentations(image, bboxes, augmentation_config, prob_demanding_ops)
if has_bboxes:
return image, bboxes
return image
@tf.function
@convert_type
def _random_augmentations(image, bboxes, augmentation_config: AugmentationConfig, prob_demanding_ops: float):
@tf.function
def apply(idx, image, bboxes):
# List of tuples (precondition, augmentation), augmentation will be applied only if precondition is True.
functions = [
(
tf.math.equal(augmentation_config.color, ColorAugmentation.AGGRESSIVE),
lambda: (
(
random_function(image, rgb_shift, 0.2, **{"r_shift": 0.15, "g_shift": 0.15, "b_shift": 0.15}),
bboxes,
)
),
),
(
tf.math.equal(augmentation_config.color, ColorAugmentation.AGGRESSIVE),
lambda: (
random_function(image, channel_swap, 0.1),
bboxes,
),
),
(
tf.math.equal(augmentation_config.color, ColorAugmentation.AGGRESSIVE),
lambda: (random_function(image, grayscale, 0.1), bboxes),
),
(
tf.math.equal(augmentation_config.color, ColorAugmentation.AGGRESSIVE),
lambda: (random_function(image, channel_drop, 0.1), bboxes),
),
(
tf.math.greater_equal(augmentation_config.color, ColorAugmentation.LIGHT),
lambda: (tf.image.random_brightness(image, 0.2), bboxes),
),
(
tf.math.greater_equal(augmentation_config.color, ColorAugmentation.LIGHT),
lambda: (tf.image.random_contrast(image, 0.8, 1.2), bboxes),
),
(
tf.math.greater_equal(augmentation_config.color, ColorAugmentation.MEDIUM),
lambda: (tf.image.random_saturation(image, 0.8, 1.2), bboxes),
),
(
tf.math.greater_equal(augmentation_config.color, ColorAugmentation.MEDIUM),
lambda: (tf.image.random_hue(image, 0.2), bboxes),
),
(
tf.math.equal(augmentation_config.crop, True),
lambda: tf.cond(
tf.greater(tf.shape(bboxes)[0], 0),
lambda: clip_random_with_bboxes(image, bboxes),
lambda: (
clip_random(
image,
min_shape=(
tf.cast(tf.cast(tf.shape(image)[0], dtype=tf.float32) * 0.9, dtype=tf.int32),
tf.cast(tf.cast(tf.shape(image)[1], dtype=tf.float32) * 0.9, dtype=tf.int32),
),
),
bboxes,
),
),
),
(
tf.math.equal(augmentation_config.distort_aspect_ratio, AspectRatioAugmentation.NORMAL),
lambda: random_function_bboxes(
image,
bboxes,
random_aspect_ratio_deformation,
prob=prob_demanding_ops,
unify_dims=False,
max_squeeze=0.6,
max_stretch=1.3,
),
),
(
tf.math.equal(augmentation_config.distort_aspect_ratio, AspectRatioAugmentation.TOWARDS_SQUARE),
lambda: random_function_bboxes(
image,
bboxes,
random_aspect_ratio_deformation,
prob=prob_demanding_ops,
unify_dims=True,
max_squeeze=0.6,
max_stretch=1.3,
),
),
(
tf.math.equal(augmentation_config.quality, True),
lambda: (random_function(image, gaussian_noise, prob=0.15, stddev_max=0.05), bboxes),
),
(
tf.math.equal(augmentation_config.erasing, True),
lambda: multiple_erase(
image,
bboxes,
iterations=tf.random.uniform((), 0, 7, tf.int32),
max_area=calculate_bboxes_max_erase_area(bboxes, max_area=0.1),
),
),
(tf.math.equal(augmentation_config.rotate90, True), lambda: rot90(image, bboxes)),
(
tf.math.equal(augmentation_config.rotate45, True),
lambda: (random_function_bboxes(image, bboxes, rot45, prob=0.5)),
),
(
tf.math.greater(augmentation_config.rotate_max, 0),
lambda: random_function_bboxes(
image,
bboxes,
random_rotate,
prob=prob_demanding_ops,
min_rotate=-augmentation_config.rotate_max,
max_rotate=augmentation_config.rotate_max,
),
),
(
tf.math.equal(augmentation_config.flip_horizontal, True),
lambda: random_function_bboxes(image, bboxes, flip_left_right, 0.5),
),
(
tf.math.equal(augmentation_config.flip_vertical, True),
lambda: random_function_bboxes(image, bboxes, flip_up_down, 0.5),
),
]
# We cannot simply index by i, this loop will find the given augmentation
# and perform it if the precondition is satisfied.
for i in range(len(functions)):
image, bboxes = tf.cond(
tf.math.logical_and(tf.equal(i, idx), functions[i][0]), functions[i][1], lambda: (image, bboxes)
)
return image, bboxes
# TODO we had some problems if random_jpeg_quality was inside the random operations ... find out why
image = tf.cond(
tf.math.equal(augmentation_config.quality, True),
lambda: tf.image.random_jpeg_quality(image, 35, 98),
lambda: image,
)
# Randomize the sequence of augmentation indices.
augmentation_count = 17
order = tf.random.shuffle(tf.range(augmentation_count))
# Loop over all augmentation and apply them.
i = tf.constant(0, dtype=tf.int32)
condition = lambda i, _image, _bboxes: tf.greater(augmentation_count, i)
body = lambda i, image, bboxes: (i + 1, *apply(order[i], image, bboxes))
_, image, bboxes = tf.while_loop(
condition,
body,
(i, image, bboxes),
shape_invariants=(
i.get_shape(),
tf.TensorShape([None, None, None]),
tf.TensorShape([None, 4]),
),
)
# this ned to be at the end, otherwise we are not guaranteed to get the square
# (and it could interact with the other augmentation in such way that we would have too much empty space)
image, bboxes = tf.cond(
tf.math.equal(augmentation_config.padding_square, True),
lambda: random_pad_to_square(image, bboxes),
lambda: (
image,
bboxes,
),
)
return image, bboxes