This repository has been archived by the owner on Nov 16, 2023. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 73
/
augmentation.py
339 lines (312 loc) · 17.8 KB
/
augmentation.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
import numpy as np
import math
# from mxnet import nd
from mxnet.gluon import nn
class Augmentation(nn.HybridBlock):
def __init__(self, angle_range, zoom_range, translation_range, target_shape, orig_shape, batch_size,
aspect_range = None, relative_angle = 0, relative_scale = (1, 1), relative_translation = 0):
super().__init__()
self._angle_range = tuple(map(lambda x : x / 180 * math.pi, angle_range) )
self._scale_range = zoom_range
try:
translation_range = tuple(translation_range)
if len(translation_range) != 2:
raise ValueError('expect translation range to have shape [2,], but got {}'.format(translation_range))
except TypeError:
translation_range = (-translation_range, translation_range)
self._translation_range = tuple(map(lambda x : x * 2, translation_range))
self._target_shape = np.array(target_shape)
self._orig_shape = np.array(orig_shape)
self._batch_size = batch_size
self._unit = np.flip(self._target_shape - 1, axis=0).reshape([2,1]) / np.flip(self._orig_shape - 1, axis=0).reshape([1,2])
self._relative_scale = relative_scale
self._relative_angle = tuple(map(lambda x : x / 180 * math.pi * relative_angle, angle_range) )
self._relative_translation = (-relative_translation * 2, relative_translation * 2)
self._aspect_range = aspect_range
def _get_relative_transform(self, F):
aspect_ratio = (self._target_shape[0] - 1) / (self._target_shape[1] - 1)
rotation = F.random.uniform(*self._relative_angle, shape=(self._batch_size))
scale = F.random.uniform(*self._relative_scale, shape=(self._batch_size))
affine_params = [scale * rotation.cos(), scale * -rotation.sin() * aspect_ratio, F.zeros_like(scale),
scale * rotation.sin() / aspect_ratio, scale * rotation.cos(), F.zeros_like(scale),
F.zeros_like(scale), F.zeros_like(scale), F.ones_like(scale)]
affine = F.reshape(F.stack(*affine_params, axis=1), [0, 3, 3])
return affine
def hybrid_forward(self, F, img1, img2):
rotation = F.random.uniform(*self._angle_range, shape=(self._batch_size))
scale = F.random.uniform(*self._scale_range, shape=(self._batch_size))
if self._aspect_range is not None:
aspect_ratio = F.random.uniform(*self._aspect_range, shape=(self._batch_size))
else:
aspect_ratio = 1
pad_x, pad_y = 1 - scale * self._unit[0, 0], 1 - scale * self._unit[1, 1]
translation_x = F.random.uniform(-1, 1, shape=(self._batch_size,)) * pad_x + F.random.uniform(*self._translation_range, shape=(self._batch_size))
translation_y = F.random.uniform(-1, 1, shape=(self._batch_size,)) * pad_y + F.random.uniform(*self._translation_range, shape=(self._batch_size))
affine_params = [scale * aspect_ratio * rotation.cos() * self._unit[0, 0], scale * aspect_ratio * -rotation.sin() * self._unit[1, 0], translation_x,
scale * rotation.sin() * self._unit[0, 1], scale * rotation.cos() * self._unit[1, 1], translation_y]
affine_params = F.stack(*affine_params, axis=1)
rel_affine = self._get_relative_transform(F)
affine_2 = F.reshape(F.batch_dot(F.reshape(affine_params, [0, 2, 3]), rel_affine), [0, 6])
rel_translation = [F.zeros((self._batch_size,)), F.zeros((self._batch_size,)), F.random.uniform(*self._relative_translation, shape=(self._batch_size,)),
F.zeros((self._batch_size,)), F.zeros((self._batch_size,)), F.random.uniform(*self._relative_translation, shape=(self._batch_size))]
rel_translation = F.stack(*rel_translation, axis = 1)
affine_2 = affine_2 + rel_translation
grid = F.GridGenerator(data=affine_params, transform_type='affine', target_shape=list(self._target_shape))
img1 = F.BilinearSampler(data=img1, grid=grid)
grid_2 = F.GridGenerator(data=affine_2, transform_type='affine', target_shape=list(self._target_shape))
img2 = F.BilinearSampler(data=img2, grid=grid_2)
return img1, img2
'''
class ChromaticBrightnessAugmentation(nn.HybridBlock):
def __init__(self, brightness = 0.5, batch_size = 1, **kwargs):
super().__init__(**kwargs)
self.brightness = brightness
self.batch_size = batch_size
def hybrid_forward(self, F, img):
aug = img
alpha = 1.0 + F.random.uniform(-self.brightness, self.brightness, shape = (self.batch_size, 1, 1, 1))
aug = F.broadcast_mul(aug, alpha)
return aug
class ChromaticContrastAugmentation(nn.HybridBlock):
def __init__(self, contrast = 0.5, batch_size = 1, **kwargs):
super().__init__(**kwargs)
self.contrast = contrast
self.coefficient = [0.299, 0.587, 0.114]
self.batch_size = batch_size
def hybrid_forward(self, F, img):
aug = img
alpha = 1.0 + F.random.uniform(-self.contrast, self.contrast, shape = (self.batch_size, 1, 1, 1))
gray = F.concat(*[img.slice_axis(axis = 1, begin = k, end = k + 1) * self.coefficient[k] for k in range(3)], dim = 1)
mean = F.mean(gray, keepdims = True, axis = (1, 2, 3))
gray = 3.0 * (1.0 - alpha) * mean
aug = F.broadcast_mul(aug, alpha)
aug = F.broadcast_add(aug, gray)
return aug
'''
class ChromaticSHAugmentation(nn.HybridBlock):
def __init__(self, saturation = 0.5, hue = 0.5, batch_size = 1, **kwargs):
super().__init__(**kwargs)
self.saturation = saturation
self.hue = hue
self.matrix_yiq = [ [ 0.299, 0.587, 0.114],
[ 0.596, -0.274, -0.321],
[ 0.211, -0.523, -0.311]]
self.matrix_rgb = [ [ 1. , 0.956, 0.621],
[ 1. , -0.272, -0.647],
[ 1. , -1.107, 1.705]]
self.batch_size = batch_size
def hybrid_forward(self, F, img):
aug = img
alpha = 1.0 + F.random.uniform(-self.saturation, self.saturation, shape = (self.batch_size, 1, 1, 1))
theta = F.random.uniform(-self.hue * np.pi, self.hue * np.pi, shape = (self.batch_size, 1, 1, 1))
su = alpha * F.cos(theta)
sw = alpha * F.sin(theta)
matrix = [ [0.299 + 0.701 * su + 0.168 * sw, 0.587 - 0.587 * su + 0.330 * sw, 0.114 - 0.114 * su - 0.497 * sw],
[0.299 - 0.299 * su - 0.328 * sw, 0.587 + 0.413 * su + 0.035 * sw, 0.114 - 0.114 * su + 0.292 * sw],
[0.299 - 0.300 * su + 1.250 * sw, 0.587 - 0.588 * su - 1.050 * sw, 0.114 + 0.886 * su - 0.203 * sw]]
aug = F.concat(*[sum([F.broadcast_mul(aug.slice_axis(axis = 1, begin = j, end = j + 1), matrix[i][j]) for j in range(3)]) for i in range(3)], dim = 1)
return aug
'''
class ChromaticGammaAugmentation(nn.HybridBlock):
def __init__(self, gamma = (0.7, 1.5), batch_size = 1, **kwargs):
super().__init__(**kwargs)
self.gamma_min, self.gamma_max = gamma
self.batch_size = batch_size
def hybrid_forward(self, F, img):
aug = img
alpha = F.random.uniform(self.gamma_min, self.gamma_max, shape = (self.batch_size, 1, 1, 1))
aug = F.broadcast_power(aug, alpha)
return aug
class ChromaticEigenAugmentation(nn.HybridBlock):
def __init__(self, batch_size = 1, **kwargs):
super().__init__(**kwargs)
self.batch_size = batch_size
def hybrid_forward(self, F, img):
spin_angle = F.random.uniform(low = -np.pi, high = np.pi, shape = (self.batch_size, 3, 1, 1))
cos_ = [F.cos(spin_angle).slice_axis(axis = 1, begin = k, end = k + 1) for k in range(3)]
sin_ = [F.sin(spin_angle).slice_axis(axis = 1, begin = k, end = k + 1) for k in range(3)]
spin_matrix = [ [ cos_[0] * cos_[1], sin_[1] * cos_[2] + sin_[0] * cos_[1] * sin_[2], sin_[1] * sin_[2] - sin_[0] * cos_[1] * cos_[2]],
[- cos_[0] * sin_[1], cos_[1] * cos_[2] - sin_[0] * sin_[1] * sin_[2], cos_[1] * sin_[2] + sin_[0] * sin_[1] * cos_[2]],
[ sin_[0] , - cos_[0] * sin_[2] , cos_[0] * cos_[2] ]]
aug = F.concat(*[sum([F.broadcast_mul(img.slice_axis(axis = 1, begin = j, end = j + 1), spin_matrix[i][j]) for j in range(3)]) for i in range(3)], dim = 1)
return aug
class ChromaticComposeAugmentation(nn.Block):
def __init__(self, brightness = 0.2, contrast = 0.5, saturation = 0.5, hue = 0.5, gamma = (0.7, 1.5), batch_size = 1, **kwargs):
super().__init__(**kwargs)
self.brightness = brightness
self.contrast = contrast
self.saturation = saturation
self.hue = hue
self.gamma = gamma
self.batch_size = batch_size
self.aug_brightness = ChromaticBrightnessAugmentation(self.brightness, self.batch_size)
self.aug_contrast = ChromaticContrastAugmentation(self.contrast, self.batch_size)
self.aug_sh = ChromaticSHAugmentation(self.saturation, self.hue, self.batch_size)
self.augs = [self.aug_brightness, self.aug_contrast, self.aug_sh]
self.Gamma = ChromaticGammaAugmentation(self.gamma, self.batch_size)
def forward(self, img1, img2):
aug = nd.concat(img1, img2, dim = 2)
augs = random.sample(self.augs, 3)
for aug_op in augs:
aug = aug_op(aug)
aug = aug.clip(0, 1)
aug = self.Gamma(aug)
return nd.split(aug, axis = 2, num_outputs = 2)
'''
class ColorAugmentation(nn.HybridBlock):
def __init__(self, contrast_range, brightness_sigma, channel_range, batch_size, shape, noise_range,
saturation, hue, gamma_range = None, eigen_aug = False, **kwargs):
super().__init__(**kwargs)
self._contrast_range = contrast_range
self._brightness_sigma = brightness_sigma
self._channel_range = channel_range
self._batch_size = batch_size
self._shape = shape
self._noise_range = noise_range
self._gamma_range = gamma_range
self._eigen_aug = eigen_aug
self._saturation = saturation
self._hue = hue
def hybrid_forward(self, F, img1, img2):
contrast = F.random.uniform(*self._contrast_range, shape=(self._batch_size, 1, 1, 1)) + 1
brightness = F.random.normal(scale=self._brightness_sigma, shape=(self._batch_size, 1, 1, 1))
channel = F.random.uniform(*self._channel_range, shape=(self._batch_size, 3, 1, 1))
noise_sigma = F.random.uniform(*self._noise_range)
if self._gamma_range is not None:
gamma = F.random.uniform(*self._gamma_range, shape = (self._batch_size, 1, 1, 1))
contrast = contrast.repeat(repeats=3, axis=1)
brightness = brightness.repeat(repeats=3, axis=1)
alpha = 1.0 + F.random.uniform(-self._saturation, self._saturation, shape = (self._batch_size, 1, 1, 1))
theta = F.random.uniform(-self._hue * np.pi, self._hue * np.pi, shape = (self._batch_size, 1, 1, 1))
su = alpha * F.cos(theta)
sw = alpha * F.sin(theta)
sh_matrix = [ [0.299 + 0.701 * su + 0.168 * sw, 0.587 - 0.587 * su + 0.330 * sw, 0.114 - 0.114 * su - 0.497 * sw],
[0.299 - 0.299 * su - 0.328 * sw, 0.587 + 0.413 * su + 0.035 * sw, 0.114 - 0.114 * su + 0.292 * sw],
[0.299 - 0.300 * su + 1.250 * sw, 0.587 - 0.588 * su - 1.050 * sw, 0.114 + 0.886 * su - 0.203 * sw]]
if self._eigen_aug:
spin_angle = F.random.uniform(low = -np.pi, high = np.pi, shape = (self._batch_size, 3, 1, 1))
cos_ = [F.cos(spin_angle).slice_axis(axis = 1, begin = k, end = k + 1) for k in range(3)]
sin_ = [F.sin(spin_angle).slice_axis(axis = 1, begin = k, end = k + 1) for k in range(3)]
spin_matrix = [ [ cos_[0] * cos_[1], sin_[1] * cos_[2] + sin_[0] * cos_[1] * sin_[2], sin_[1] * sin_[2] - sin_[0] * cos_[1] * cos_[2]],
[-cos_[0] * sin_[1], cos_[1] * cos_[2] - sin_[0] * sin_[1] * sin_[2], cos_[1] * sin_[2] + sin_[0] * sin_[1] * cos_[2]],
[ sin_[0] ,-cos_[0] * sin_[2] , cos_[0] * cos_[2] ]]
ret = []
for img in (img1, img2):
aug = img
aug = F.concat(*[sum([F.broadcast_mul(aug.slice_axis(axis = 1, begin = j, end = j + 1), sh_matrix[i][j]) for j in range(3)]) for i in range(3)], dim = 1)
noise = F.random.normal(scale=1, shape=(self._batch_size, 3) + tuple(self._shape))
aug = aug + noise * noise_sigma
mean = F.mean(aug, keepdims=True, axis=(2,3))
aug = F.broadcast_minus(aug, mean)
aug = F.broadcast_mul(aug, contrast * channel)
if self._eigen_aug:
aug = F.concat(*[sum([F.broadcast_mul(aug.slice_axis(axis = 1, begin = j, end = j + 1), spin_matrix[i][j]) for j in range(3)]) for i in range(3)], dim = 1)
aug = F.broadcast_add(aug, mean * channel + brightness)
aug = F.clip(aug, 0, 1)
if self._gamma_range is not None:
aug = F.broadcast_power(aug, F.exp(gamma))
ret.append(aug)
return ret
class GeometryAugmentation(nn.HybridBlock):
def __init__(self, angle_range, zoom_range, translation_range, target_shape, orig_shape, batch_size,
aspect_range = None, relative_angle=None, relative_scale=None, relative_translation=None):
super().__init__()
self._angle_range = tuple(map(lambda x : x / 180 * math.pi, angle_range) )
self._scale_range = zoom_range
try:
translation_range = tuple(translation_range)
if len(translation_range) != 2:
raise ValueError('expect translation range to have shape [2,], but got {}'.format(translation_range))
except TypeError:
translation_range = (-translation_range, translation_range)
self._translation_range = tuple(map(lambda x : x * 2, translation_range))
self._target_shape = np.array(target_shape)
self._orig_shape = np.array(orig_shape)
self._batch_size = batch_size
self._unit = np.flip(self._target_shape - 1, axis=0).reshape([2,1]) / np.flip(self._orig_shape - 1, axis=0).reshape([1,2])
self._relative = relative_angle is not None
if self._relative:
self._relative_scale = relative_scale
self._relative_angle = tuple(map(lambda x : x / 180 * math.pi * relative_angle, angle_range) )
self._relative_translation = tuple(map(lambda x: x * relative_translation, self._translation_range)) if relative_translation is not None else None
self._aspect_range = aspect_range
def _get_relative_transform(self, F):
aspect_ratio = (self._target_shape[0] - 1) / (self._target_shape[1] - 1)
rotation = F.random.uniform(*self._relative_angle, shape=(self._batch_size))
scale = F.random.uniform(*self._relative_scale, shape=(self._batch_size))
affine_params = [scale * rotation.cos(), scale * -rotation.sin() * aspect_ratio, F.zeros_like(scale),
scale * rotation.sin() / aspect_ratio, scale * rotation.cos(), F.zeros_like(scale),
F.zeros_like(scale), F.zeros_like(scale), F.ones_like(scale)]
affine = F.reshape(F.stack(*affine_params, axis=1), [0, 3, 3])
inverse = F.stack(
rotation.cos() / scale,
rotation.sin() / scale,
-rotation.sin() / scale,
rotation.cos() / scale,
axis=1
)
inverse = F.reshape(inverse, [0, 2, 2])
return affine, inverse
def hybrid_forward(self, F, img1, img2, flow, mask):
rotation = F.random.uniform(*self._angle_range, shape=(self._batch_size))
aspect_ratio = F.random.uniform(*self._aspect_range, shape=(self._batch_size)) if self._aspect_range is not None else 1
scale = F.random.uniform(*self._scale_range, shape=(self._batch_size))
os = (self._orig_shape[0] - 1, self._orig_shape[1] - 1)
ts = (self._target_shape[0] - 1, self._target_shape[1] - 1)
abs_rotation = F.abs(rotation)
scale = F.minimum(scale, os[1] / (aspect_ratio * (ts[0] * F.sin(abs_rotation) + ts[1] * F.cos(abs_rotation))))
scale = F.minimum(scale, os[0] / (ts[0] * F.cos(abs_rotation) + ts[1] * F.sin(abs_rotation)))
pad_x, pad_y = 1 - scale * self._unit[0, 0], 1 - scale * self._unit[1, 1]
translation_x = F.random.uniform(-1, 1, shape=(self._batch_size,)) * pad_x + F.random.uniform(*self._translation_range, shape=(self._batch_size))
translation_y = F.random.uniform(-1, 1, shape=(self._batch_size,)) * pad_y + F.random.uniform(*self._translation_range, shape=(self._batch_size))
affine_params = [scale * aspect_ratio * rotation.cos() * self._unit[0, 0], scale * aspect_ratio * -rotation.sin() * self._unit[1, 0], translation_x,
scale * rotation.sin() * self._unit[0, 1], scale * rotation.cos() * self._unit[1, 1], translation_y]
affine_params = F.stack(*affine_params, axis=1)
affine_inverse = F.stack(
rotation.cos() / (scale * aspect_ratio),
rotation.sin() / (scale * aspect_ratio),
-rotation.sin() / scale,
rotation.cos() / scale,
axis=1
)
linv = F.reshape(affine_inverse, [0, 2, 2])
mask = mask.broadcast_like(flow.slice_axis(axis = 1, begin = 0, end = 1))
rel_affine, rel_inverse = self._get_relative_transform(F)
affine_2 = F.reshape(F.batch_dot(F.reshape(affine_params, [0, 2, 3]), rel_affine), [0, 6])
if self._relative_translation is not None:
rel_translation = F.random.uniform(*self._relative_translation, shape=(self._batch_size, 2, 1, 1))
rel_scale = F.concat(F.ones([self._batch_size, 1, 1, 1]) * (self._orig_shape[1] - 1) / 2,
F.ones([self._batch_size, 1, 1, 1]) * (self._orig_shape[0] - 1) / 2, dim=1)
flow = F.broadcast_minus(flow, rel_translation * rel_scale)
concat_img = F.concat(img1, mask, F.broadcast_mul(flow, mask), dim=1)
grid = F.GridGenerator(data=affine_params, transform_type='affine', target_shape=list(self._target_shape))
force_translation = F.maximum(grid.max(axis=(2, 3), keepdims=True) - 1, 0) + F.minimum(grid.min(axis=(2, 3), keepdims=True) + 1, 0)
grid = F.broadcast_minus(grid, force_translation)
grid = grid.clip(-1, 1)
concat_img = F.BilinearSampler(data=concat_img, grid=grid)
img1 = F.slice_axis(concat_img, axis=1, begin=0, end=3)
mask = F.slice_axis(concat_img, axis=1, begin=3, end=4)
flow = F.slice_axis(concat_img, axis=1, begin=4, end=6)
flow = F.broadcast_div(flow, F.maximum(mask, 1e-8))
# relative
grid_2 = F.GridGenerator(data=affine_2, transform_type='affine', target_shape=list(self._target_shape))
grid_2 = F.broadcast_minus(grid_2, force_translation)
if self._relative_translation is not None:
grid_2 = F.broadcast_add(grid_2, rel_translation)
img2 = F.BilinearSampler(data=img2, grid=grid_2)
inverse_2 = F.batch_dot(rel_inverse, linv)
flow = F.reshape_like(F.batch_dot(inverse_2, F.reshape(flow, (0, 0, -3))), flow)
scale = F.stack(F.ones([self._batch_size]) * (self._target_shape[1] - 1) / 2,
F.zeros([self._batch_size]),
F.zeros([self._batch_size]),
F.ones([self._batch_size]) * (self._target_shape[0] - 1) / 2,
axis=1)
scale = F.reshape(scale, [0, 2, 2])
I = F.reshape(F.one_hot(F.arange(0, 2), depth=2), [1, 2, 2])
grid = F.GridGenerator(data=F.reshape(F.one_hot(F.arange(0, 2), depth=3), [1, 6]),
transform_type='affine',
target_shape=list(self._target_shape))
grid = F.reshape(F.repeat(grid, axis=0, repeats=self._batch_size), [0, 0, -3])
factor = F.batch_dot(F.broadcast_minus(rel_inverse, I), scale)
flow = flow + F.reshape_like(F.batch_dot(factor, grid), flow)
return img1, img2, flow, mask