-
Notifications
You must be signed in to change notification settings - Fork 0
/
my_model.py
204 lines (150 loc) · 8.95 KB
/
my_model.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
from keras.layers import Input, Lambda, Activation, concatenate, add
from keras.layers.convolutional import MaxPooling2D, Conv3D, Conv2D, Conv2DTranspose
from keras.layers.normalization import BatchNormalization
from keras.layers.core import Permute
from keras.layers.wrappers import TimeDistributed
from keras import backend as K
from keras.losses import categorical_crossentropy
K.set_image_dim_ordering("th")
channel_axis_5d = 2
def timeDist_DeepEM3D_Net(input, scale=True, size=256):
# Input Shape is 256 x 256 x 3 (tf) or 3 x 256 x 256 (th)
# make sure image size is multiple of 16
size_m = K.cast(K.round(size / 16 + 0.5) * 16, 'int32')
input = Permute([1, 3, 4, 2])(input)
input = TimeDistributed(Lambda(lambda image:
K.tf.image.resize_images(image, (size_m, size_m))))(input)
input = Permute([1, 4, 2, 3])(input)
x1, z1 = time_inception_resnet_stem_3Dconv(input)
x1 = time_inception_resnet_v2_A(x1, scale_residual=scale)
x2 = time_reduction_A(x1, k=256, l=256, m=384, n=384)
x2 = time_inception_resnet_v2_B(x2, scale_residual=scale)
x3 = time_reduction_resnet_v2_B(x2)
x3 = time_inception_resnet_v2_C(x3, scale_residual=scale)
u1 = TimeDistributed(Conv2DTranspose(2, (5, 5), strides=(2, 2), padding='same'))(z1)
u1 = BatchNormalization(axis=2)(u1)
u2 = TimeDistributed(Conv2DTranspose(2, (9, 9), strides=(4, 4), padding='same'))(x1)
u2 = BatchNormalization(axis=2)(u2)
u3 = TimeDistributed(Conv2DTranspose(2, (17, 17), strides=(8, 8), padding='same'))(x2)
u3 = BatchNormalization(axis=2)(u3)
u4 = TimeDistributed(Conv2DTranspose(2, (33, 33), strides=(16, 16), padding='same'))(x3)
u4 = BatchNormalization(axis=2)(u4)
merged = add([u1, u2, u3, u4])
out = Lambda(time_dist_softmax, output_shape=time_dist_softmax_out_shape)(merged)
out = Permute([1, 3, 4, 2])(out)
out = TimeDistributed(Lambda(lambda image:
K.tf.image.resize_images(image, (size, size))))(out)
out = Permute([1, 4, 2, 3])(out)
return out
def time_inception_resnet_stem_3Dconv(input):
# Input Shape is 320 x 320 x 3 (tf) or 3 x 320 x 320 (th)
# for original our biomedical paper, we used 2 layer of 3D convolution followed by 2D conv
# here we do 2 layer 3D conv, but keep the times (original #slices in Z direction) unchanged.
# First, permute time 5d time to the order : [ch,height,width,time(Z)] for 3D convolution
c = Permute([2, 3, 4, 1])(input)
# [h, w, time, c] for tensorflow
# c = Permute([3, 4, 1, 2])(input)
c = Conv3D(32, (3, 3, 3), activation='relu', strides=(2, 2, 1), padding='same')(c)
# next_size= 160#input[1]/2
c = Conv3D(64, (3, 3, 3), activation='relu', strides=(1, 1, 1), padding='same')(c)
# Now, for rest of time distributed 2D convs, we need Permute data back to dimension order [time,channel,height,
# width]
c = Permute([4, 1, 2, 3])(c)
c = TimeDistributed(Conv2D(64, (3, 3), activation='relu', padding='same'))(c)
# -------------- left branch -------------------------------------------
c1 = TimeDistributed(Conv2D(64, (1, 1), activation='relu', padding='same'))(c)
c1 = TimeDistributed(Conv2D(96, (3, 3), activation='relu', padding='same'))(c1)
# --------------- right branch ------------------------------------------
c2 = TimeDistributed(Conv2D(64, (1, 1), activation='relu', padding='same'))(c)
c2 = TimeDistributed(Conv2D(64, (7, 1), activation='relu', padding='same'))(c2)
c2 = TimeDistributed(Conv2D(64, (1, 7), activation='relu', padding='same'))(c2)
c2 = TimeDistributed(Conv2D(96, (3, 3), activation='relu', padding='same'))(c2)
m1 = concatenate([c1, c2], axis=channel_axis_5d)
# m_pad =ZeroPadding2D((1,1))(m1)
p1 = TimeDistributed(MaxPooling2D((3, 3), strides=(2, 2), padding='same'))(m1)
p2 = TimeDistributed(Conv2D(192, (3, 3), activation='relu', strides=(2, 2),
padding='same'))(m1)
m2 = concatenate([p1, p2], axis=channel_axis_5d)
m2 = BatchNormalization(axis=2)(m2) # channel axis
m2 = Activation('relu')(m2)
return m2, m1
def time_inception_resnet_v2_A(input, scale_residual=True):
# Input is relu activation
init = input
ir1 = TimeDistributed(Conv2D(32, (1, 1), activation='relu', padding='same'))(input)
ir2 = TimeDistributed(Conv2D(32, (1, 1), activation='relu', padding='same'))(input)
ir2 = TimeDistributed(Conv2D(32, (3, 3), activation='relu', padding='same'))(ir2)
ir3 = TimeDistributed(Conv2D(32, (1, 1), activation='relu', padding='same'))(input)
ir3 = TimeDistributed(Conv2D(48, (3, 3), activation='relu', padding='same'))(ir3)
ir3 = TimeDistributed(Conv2D(64, (3, 3), activation='relu', padding='same'))(ir3)
ir_merge = concatenate([ir1, ir2, ir3], axis=channel_axis_5d)
ir_conv = TimeDistributed(Conv2D(384, (1, 1), activation='linear', padding='same'))(ir_merge)
if scale_residual: ir_conv = Lambda(lambda x: x * 0.1)(ir_conv)
out = add([init, ir_conv])
out = BatchNormalization(axis=2)(out)
out = Activation("relu")(out)
return out
def time_reduction_A(input, k=192, l=224, m=256, n=384):
# r1 = TimeDistributed(Conv2D(384, 3, 3, activation='relu', strides=(2,2),padding='same'))(input)
r1 = TimeDistributed(MaxPooling2D((3, 3), strides=(2, 2), padding='same'))(input)
r2 = TimeDistributed(Conv2D(n, (3, 3), activation='relu', strides=(2, 2), padding='same'))(input)
r3 = TimeDistributed(Conv2D(k, (1, 1), activation='relu', padding='same'))(input)
r3 = TimeDistributed(Conv2D(l, (3, 3), activation='relu', padding='same'))(r3)
# m_pad =ZeroPadding2D((1,1))(r3)
r3 = TimeDistributed(Conv2D(m, (3, 3), activation='relu', strides=(2, 2), padding='same'))(r3)
m = concatenate([r1, r2, r3], axis=channel_axis_5d)
m = BatchNormalization(axis=2)(m)
m = Activation('relu')(m)
return m
def time_inception_resnet_v2_B(input, scale_residual=True):
# Input is relu activation
init = input
ir1 = TimeDistributed(Conv2D(192, (1, 1), activation='relu', padding='same'))(input)
ir2 = TimeDistributed(Conv2D(128, (1, 1), activation='relu', padding='same'))(input)
ir2 = TimeDistributed(Conv2D(160, (1, 7), activation='relu', padding='same'))(ir2)
ir2 = TimeDistributed(Conv2D(192, (7, 1), activation='relu', padding='same'))(ir2)
ir_merge = concatenate([ir1, ir2], axis=channel_axis_5d)
ir_conv = TimeDistributed(Conv2D(1152, (1, 1), activation='linear', padding='same'))(ir_merge)
if scale_residual: ir_conv = Lambda(lambda x: x * 0.1)(ir_conv)
out = add([init, ir_conv])
out = BatchNormalization(axis=2)(out)
out = Activation("relu")(out)
return out
def time_reduction_resnet_v2_B(input):
r1 = TimeDistributed(MaxPooling2D((3, 3), strides=(2, 2), padding='same'))(input)
# r1 = TimeDistributed(Conv2D(1152, 3, 3, activation='relu', strides=(2,2),padding='same'))(input)
r2 = TimeDistributed(Conv2D(256, (1, 1), activation='relu', padding='same'))(input)
r2 = TimeDistributed(Conv2D(384, (3, 3), activation='relu', strides=(2, 2), padding='same'))(r2)
r3 = TimeDistributed(Conv2D(256, (1, 1), activation='relu', padding='same'))(input)
r3 = TimeDistributed(Conv2D(288, (3, 3), activation='relu', strides=(2, 2), padding='same'))(r3)
r4 = TimeDistributed(Conv2D(256, (1, 1), activation='relu', padding='same'))(input)
r4 = TimeDistributed(Conv2D(288, (3, 3), activation='relu', padding='same'))(r4)
r4 = TimeDistributed(Conv2D(320, (3, 3), activation='relu', strides=(2, 2), padding='same'))(r4)
m = concatenate([r1, r2, r3, r4], axis=channel_axis_5d)
m = BatchNormalization(axis=2)(m)
m = Activation('relu')(m)
return m
def time_inception_resnet_v2_C(input, scale_residual=True):
# Input is relu activation
init = input
ir1 = TimeDistributed(Conv2D(192, (1, 1), activation='relu', padding='same'))(input)
ir2 = TimeDistributed(Conv2D(192, (1, 1), activation='relu', padding='same'))(input)
ir2 = TimeDistributed(Conv2D(224, (1, 3), activation='relu', padding='same'))(ir2)
ir2 = TimeDistributed(Conv2D(256, (3, 1), activation='relu', padding='same'))(ir2)
ir_merge = concatenate([ir1, ir2], axis=channel_axis_5d)
ir_conv = TimeDistributed(Conv2D(2144, (1, 1), activation='linear', padding='same'))(ir_merge)
if scale_residual: ir_conv = Lambda(lambda x: x * 0.1)(ir_conv)
out = add([init, ir_conv])
out = BatchNormalization(axis=2)(out)
out = Activation("relu")(out)
return out
def time_dist_softmax(x):
assert K.ndim(x) == 5 # K as backend
# e = K.exp(x - K.max(x, axis=2, keepdims=True))
e = K.exp(x)
# it seems here already consider neuron-boundary ratio
s = K.sum(e, axis=2, keepdims=True)
return e / s
def time_dist_softmax_out_shape(input_shape):
shape = list(input_shape)
return tuple(shape)