-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathFlyVisNetH_model.py
229 lines (192 loc) · 13.4 KB
/
FlyVisNetH_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
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
# FlyVisNetH
# Angel Canelo 2024.08.02
###### import ######################
import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.keras.models import Model
from tensorflow.keras.regularizers import l1, l2
from tensorflow.keras.initializers import RandomUniform, RandomNormal
from tensorflow.keras.constraints import Constraint
##################################
########### CNN Model ############
class filter_cons(Constraint):
def __init__(self, min_value, max_value):
self.min_value = min_value
self.max_value = max_value
def __call__(self, w):
return tf.clip_by_value(w, self.min_value, self.max_value)
class FlyVisNetH:
def FlyVisNet_model(self, HEIGHT, WIDTH, classes):
pos_constraint = filter_cons(min_value=0.01, max_value=0.5)
neg_constraint = filter_cons(min_value=-0.5, max_value=-0.01)
## Filter values of for Retina, Lamina, and Medulla layers
filter_exc = RandomUniform(minval=0.2, maxval=0.3)
filter_inh = RandomUniform(minval=-0.2, maxval=-0.1)
filter_inh2 = RandomUniform(minval=-0.2, maxval=-0.1)
## Filter values of for Retina, Lamina, and Medulla layers (Option 2)
# filter_exc = RandomUniform(minval=0.1, maxval=0.2)
# filter_inh = RandomUniform(minval=-0.25, maxval=-0.2)
# filter_inh2 = RandomUniform(minval=-0.3, maxval=-0.25)
filter_T2 = RandomUniform(minval=0.01, maxval=0.15)
filter_T3 = RandomUniform(minval=-0.15, maxval=-0.01)
filter_T4 = RandomUniform(minval=0.01, maxval=0.1)
filter_T5 = RandomUniform(minval=0.01, maxval=0.15)
filter_LC11 = RandomUniform(minval=0.01, maxval=0.08)
filter_LC15 = RandomUniform(minval=0.01, maxval=0.04)
filter_LPLC2 = RandomUniform(minval=0.25, maxval=0.5)
filter_TmY9 = RandomUniform(minval=0.01, maxval=0.2)
filter_TmY5 = RandomUniform(minval=0.01, maxval=0.2)
filter_TmY4 = RandomUniform(minval=0.01, maxval=0.2)
filter_Li = RandomUniform(minval=-0.15, maxval=-0.01)
filter_LPi = RandomUniform(minval=-0.1, maxval=-0.05)
ON_act = 'linear'
OFF_act = 'linear'
L_act = 'linear'
Ts = 'relu'
T2T3_act = 'linear'
inter_act = 'linear'
act = 'relu'
train = True
inputs = layers.Input(shape=[HEIGHT, WIDTH, 1]) # (height, width, channels)
# RETINA
R16 = layers.Conv2D(6, 1, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_inh,
activation=OFF_act, padding='same', name='R16', trainable=train, kernel_constraint=neg_constraint)(inputs)
# LAMINA
L1 = layers.Conv2D(1, 1, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_inh,
activation=OFF_act, padding='same', name='L1_kernel', trainable=train, kernel_constraint=neg_constraint)(R16)
L1 = layers.MaxPooling2D(pool_size=(2, 2), name='L1')(L1)
L1b = layers.MaxPooling2D(pool_size=(2, 2))(L1)
L2 = layers.Conv2D(1, 1, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_exc,
activation=L_act, padding='same', name='L2_kernel', trainable=train, kernel_constraint=pos_constraint)(R16)
L2 = layers.MaxPooling2D(pool_size=(2, 2), name='L2')(L2)
L2b = layers.MaxPooling2D(pool_size=(2, 2))(L2)
L3 = layers.Conv2D(1, 1, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_exc,
activation=L_act, padding='same', name='L3_kernel', trainable=train, kernel_constraint=pos_constraint)(R16)
L3 = layers.MaxPooling2D(pool_size=(2, 2), name='L3')(L3)
# MEDULLA
Mi1 = layers.Conv2D(1, 4, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_exc,
activation=ON_act, padding='same', name='Mi1_kernel', trainable=train, kernel_constraint=pos_constraint)(L1)
Mi1 = layers.MaxPooling2D(pool_size=(2, 2), name='Mi1')(Mi1)
Tm3 = layers.Conv2D(1, 4, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_exc,
activation=ON_act, padding='same', name='Tm3_kernel', trainable=train, kernel_constraint=pos_constraint)(L1)
Tm3 = layers.MaxPooling2D(pool_size=(2, 2), name='Tm3')(Tm3)
C3 = layers.concatenate([L1, L3])
C3 = layers.Conv2D(1, 4, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_inh2,
activation=OFF_act, padding='same', name='C3_kernel', trainable=train, kernel_constraint=neg_constraint)(C3)
C3 = layers.MaxPooling2D(pool_size=(2, 2), name='C3')(C3)
C3 = layers.Lambda(lambda x: tf.roll(x, shift=1, axis=0))(C3)
Mi4 = layers.concatenate([L1, L3])
Mi4 = layers.Conv2D(1, 4, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_inh2,
activation=OFF_act, padding='same', name='Mi4_kernel', trainable=train, kernel_constraint=neg_constraint)(Mi4)
Mi4 = layers.MaxPooling2D(pool_size=(2, 2), name='Mi4')(Mi4)
Mi4 = layers.Lambda(lambda x: tf.roll(x, shift=1, axis=0))(Mi4)
Tm1 = layers.Conv2D(1, 4, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_exc,
activation=ON_act, padding='same', name='Tm1_kernel', trainable=train, kernel_constraint=pos_constraint)(L2)
Tm1 = layers.MaxPooling2D(pool_size=(2, 2), name='Tm1')(Tm1)
Tm2 = layers.Conv2D(1, 4, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_exc,
activation=ON_act, padding='same', name='Tm2_kernel', trainable=train, kernel_constraint=pos_constraint)(L2)
Tm2 = layers.MaxPooling2D(pool_size=(2, 2), name='Tm2')(Tm2)
Tm4 = layers.Conv2D(1, 4, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_exc,
activation=ON_act, padding='same', name='Tm4_kernel', trainable=train, kernel_constraint=pos_constraint)(L2)
Tm4 = layers.MaxPooling2D(pool_size=(2, 2), name='Tm4')(Tm4)
Mi9 = layers.Conv2D(1, 4, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_inh2,
activation=OFF_act, padding='same', name='Mi9_kernel', trainable=train, kernel_constraint=neg_constraint)(L3)
Mi9 = layers.MaxPooling2D(pool_size=(2, 2), name='Mi9')(Mi9)
Mi9 = layers.Lambda(lambda x: tf.roll(x, shift=1, axis=0))(Mi9)
Tm9 = layers.Conv2D(1, 4, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_exc,
activation=ON_act, padding='same', name='Tm9_kernel', trainable=train, kernel_constraint=pos_constraint)(L3)
Tm9 = layers.MaxPooling2D(pool_size=(2, 2), name='Tm9')(Tm9)
Tm9 = layers.Lambda(lambda x: tf.roll(x, shift=1, axis=0))(Tm9)
CT1 = layers.concatenate([L3, L2])
CT1 = layers.Conv2D(1, 4, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_inh,
activation=OFF_act, padding='same', name='CT1_kernel', trainable=train, kernel_constraint=neg_constraint)(CT1)
CT1 = layers.MaxPooling2D(pool_size=(2, 2), name='CT1')(CT1)
CT1 = layers.Lambda(lambda x: tf.roll(x, shift=1, axis=0))(CT1)
TmY9 = layers.concatenate([Tm2, Mi1, C3, Mi4, Tm1, Mi9])
TmY9 = layers.Conv2D(1, 4, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_TmY9,
activation=OFF_act, padding='same', name='TmY9_kernel', trainable=train, kernel_constraint=pos_constraint)(TmY9)
TmY9 = layers.MaxPooling2D(pool_size=(2, 2), name='TmY9')(TmY9)
###
TmY4 = layers.concatenate([Tm2, Mi1, C3, Mi4, Tm1, Mi9])
TmY4 = layers.Conv2D(1, 4, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_TmY4,
activation=ON_act, padding='same', name='TmY4_kernel', trainable=train, kernel_constraint=pos_constraint)(TmY4)
TmY4 = layers.MaxPooling2D(pool_size=(2, 2), name='TmY4')(TmY4)
###
TmY5 = layers.concatenate([Mi4, Mi9])
TmY5 = layers.Conv2D(1, 4, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_TmY5,
activation=ON_act, padding='same', name='TmY5_kernel', trainable=train, kernel_constraint=pos_constraint)(TmY5)
TmY5b = layers.MaxPooling2D(pool_size=(2, 2), name='TmY5')(TmY5)
T2 = layers.concatenate([L1b, Tm2, Mi1, Tm3, C3, Mi4, Tm9, CT1])
T2 = layers.Conv2D(1, 3, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_T2,
activation=T2T3_act, padding='same', name='T2_kernel', trainable=train, kernel_constraint=pos_constraint)(
T2)
T2 = layers.MaxPooling2D(pool_size=(2, 2), name='T2')(T2)
T3 = layers.concatenate([L2b, Tm1, Tm2, Tm3, Mi1, Mi9, C3, Tm9, CT1])
T3 = layers.Conv2D(1, 3, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_T3,
activation=OFF_act, padding='same', name='T3_kernel', trainable=train, kernel_constraint=neg_constraint)(
T3)
T3 = layers.MaxPooling2D(pool_size=(2, 2), name='T3')(T3)
T4 = layers.concatenate([Mi1, Tm3, C3, Mi4, Mi9])
T4 = layers.Conv2D(1, 3, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_T4,
activation=Ts, padding='same', name='T4_kernel', trainable=train, kernel_constraint=pos_constraint)(T4)
T4 = layers.MaxPooling2D(pool_size=(2, 2), name='T4')(T4)
# LOBULA
T5 = layers.concatenate([Tm1, Tm2, Tm4, Tm9, CT1])
T5 = layers.Conv2D(1, 3, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_T5,
activation=Ts, padding='same', name='T5_kernel', trainable=train, kernel_constraint=pos_constraint)(T5)
T5 = layers.MaxPooling2D(pool_size=(2, 2), name='T5')(T5)
Li = layers.concatenate([T2, T3])
Li = layers.Conv2D(1, 5, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_Li,
activation=inter_act, padding='same', trainable=train, kernel_constraint=neg_constraint)(Li)
LPi = layers.concatenate([T4, T5])
LPi = layers.Conv2D(1, 5, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_LPi,
activation=inter_act, padding='same', trainable=True, kernel_constraint=neg_constraint)(LPi)
LC11 = layers.concatenate([T2, T3])
LC11 = layers.Conv2D(1, 3, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_LC11,
activation=act, padding='same', name='LC11_kernel', trainable=train, kernel_constraint=pos_constraint)(
LC11)
LC11 = layers.MaxPooling2D(pool_size=(2, 2), name='LC11')(LC11)
LC15 = layers.concatenate([Li, TmY9, TmY4])
LC15 = layers.Conv2D(1, 3, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_LC15,
activation=act, padding='same', name='LC15_kernel', trainable=train, kernel_constraint=pos_constraint)(
LC15)
LC15 = layers.MaxPooling2D(pool_size=(2, 2), name='LC15')(LC15)
LPLC2 = layers.concatenate([T4, T5, LPi, TmY5b])
LPLC2 = layers.Conv2D(1, 3, data_format="channels_last", kernel_regularizer=l2(1e-3),
kernel_initializer=filter_LPLC2,
activation=act, padding='same', name='LPLC2_kernel', trainable=train, kernel_constraint=pos_constraint)(
LPLC2)
LPLC2 = layers.MaxPooling2D(pool_size=(2, 2), name='LPLC2')(LPLC2)
# OPTIC GLOMERULI
CB = layers.Concatenate()([LC11, LC15, LPLC2])
CB = layers.Flatten()(CB)
CB = layers.Dense(128, kernel_initializer='glorot_uniform', kernel_regularizer=l2(1e-3),
activation='relu')(CB)
CB = layers.Dense(classes, kernel_initializer='glorot_uniform', kernel_regularizer=l2(1e-3),
activation='softmax', name='classification')(CB)
return Model(inputs=inputs, outputs=CB, name='FlyVisNetH')