-
Notifications
You must be signed in to change notification settings - Fork 1
/
M3SPADA_model.py
174 lines (141 loc) · 6.71 KB
/
M3SPADA_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
# -*- coding: utf-8 -*-
import tensorflow as tf
from tensorflow import keras
from keras import layers
from tensorflow.keras.layers import Layer, Dense, Activation, Dropout, Conv1D
from tensorflow.keras.layers import BatchNormalization, Conv2D
from tensorflow.keras.layers import Concatenate, GlobalAveragePooling2D
class Conv1dBloc(Layer):
"""Elementary 1D-convolution block for TempCNN encoder"""
def __init__(self, filters_nb, kernel_size, drop_val, **kwargs):
super(Conv1dBloc, self).__init__(**kwargs)
self.conv1D = Conv1D(filters_nb, kernel_size,
padding="same", kernel_initializer='he_normal')
self.batch_norm = BatchNormalization()
self.act = Activation('relu')
self.output_ = Dropout(drop_val)
@tf.function(experimental_relax_shapes=True)
def call(self, inputs, training=False):
conv1d = self.conv1D(inputs)
batch_norm = self.batch_norm(conv1d, training=training)
act = self.act(batch_norm)
return self.output_(act, training=training)
class TempCnnEncoder(Layer):
""""Encoder of SITS on temporal dimension"""
def __init__(self, drop_val=0.5, **kwargs):
super(TempCnnEncoder, self).__init__(**kwargs)
self.conv_bloc1 = Conv1dBloc(64, 5, drop_val)
self.conv_bloc2 = Conv1dBloc(64, 5, drop_val)
self.conv_bloc3 = Conv1dBloc(64, 5, drop_val)
self.flatten = layers.Flatten()
@tf.function(experimental_relax_shapes=True)
def call(self, inputs, training=False):
conv1 = self.conv_bloc1(inputs, training=training)
conv2 = self.conv_bloc2(conv1, training=training)
conv3 = self.conv_bloc3(conv2, training=training)
flatten = self.flatten(conv3)
return flatten
class PatchEncoder(Layer):
""""Encoder of VHR on temporal dimension"""
def __init__(self, drop_val=0.5, **kwargs):
super(PatchEncoder, self).__init__(**kwargs)
self.conv2D_1 = Conv2D(128, 5, strides=(1, 1), padding="same",
kernel_initializer='he_normal')
self.batch_norm_1 = BatchNormalization()
self.act_1 = Activation('relu')
self.drop_out_1 = Dropout(drop_val)
self.conv2D_2 = Conv2D(128, 5, strides=(1, 1), padding="same",
kernel_initializer='he_normal')
self.batch_norm_2 = BatchNormalization()
self.act_2 = Activation('relu')
self.drop_out_2 = Dropout(drop_val)
self.conv2D_3 = Conv2D(128, 5, strides=(1, 1), padding="same",
kernel_initializer='he_normal')
self.batch_norm_3 = BatchNormalization()
self.act_3 = Activation('relu')
self.drop_out_3 = Dropout(drop_val)
self.global_avg_pooling = GlobalAveragePooling2D()
def call(self, inputs, training=False):
conv2d_1 = self.conv2D_1(inputs)
bn_1 = self.batch_norm_1(conv2d_1, training=training)
act_1 = self.act_1(bn_1)
do_1 = self.drop_out_1(act_1, training=training)
conv2d_2 = self.conv2d_2(do_1)
bn_2 = self.batch_norm_2(conv2d_2, training=training)
act_2 = self.act_2(bn_2)
do_2 = self.drop_out_2(act_2, training=training)
conv2d_3 = self.conv2d_3(do_2)
bn_3 = self.batch_norm_3(conv2d_3, training=training)
act_3 = self.act_3(bn_3)
do_3 = self.drop_out_3(act_3, training=training)
return self.global_avg_pooling(do_3)
class Classifier(Layer):
""""Generic classifier for concatenated features"""
def __init__(self, nb_class, nb_units, drop_val=0.5, **kwargs):
super(Classifier, self).__init__(**kwargs)
self.dense = Dense(nb_units)
self.batch_norm = BatchNormalization()
self.act = Activation('relu')
self.dropout = Dropout(drop_val)
self.output_ = Dense(nb_class, activation="softmax")
@tf.function(experimental_relax_shapes=True)
def call(self, inputs, training=False):
dense = self.dense(inputs)
batch_norm = self.batch_norm(dense, training=training)
act = self.act(batch_norm)
dropout = self.dropout(act, training=training)
return self.output_(dropout)
class AuxiliaryClassifier(Layer):
""""Auxiliary classifier for each branch encoder"""
def __init__(self, nb_class, **kwargs):
super(AuxiliaryClassifier, self).__init__(**kwargs)
self.output_ = Dense(nb_class, activation="softmax")
# @tf.function(experimental_relax_shapes=True)
def call(self, inputs):
return self.output_(inputs)
@tf.custom_gradient
def gradient_reverse(x, lamb_da=1.0):
y = tf.identity(x)
def custom_grad(dy):
return lamb_da * -dy, None
return y, custom_grad
class GradReverse(Layer):
""""Gradient reversal layer (GRL)"""
def __init__(self):
super().__init__()
def call(self, x, lamb_da=1.0):
return gradient_reverse(x, lamb_da)
class M3SPADAModel(keras.Model):
""""M3SPADAModel model is composed of
a feature extractor for times series: ts_encoder
a feature extractor for VHR images: vhr_encoder
a label predictor/classifier for concatenated features: fusion_classifier
a label predictor/classifier for ts features: ts_classifier
a label predictor/classifier for vhr features: vhr_classifier
a domain predictor/classifier: domain_classifier
a GRL to connect concatenated features and domain predictor/classifier
"""
def __init__(self, nb_class, nb_units, drop_val=0.5, **kwargs):
super(M3SPADAModel, self).__init__(**kwargs)
self.ts_encoder = TempCnnEncoder()
self.ts_classifier = AuxiliaryClassifier(nb_class)
self.vhr_encoder = PatchEncoder(drop_val)
self.vhr_classifier = AuxiliaryClassifier(nb_class)
self.concat = Concatenate()
self.fusion_classifier = Classifier(nb_class, nb_units)
self.grl = GradReverse()
self.domain_classifier = Classifier(2, 256)
@tf.function(experimental_relax_shapes=True)
def call(self, inputs, training=False, lamb_da=1.0):
ts_input, vhr_input = inputs
ts_enc_out = self.ts_encoder(ts_input, training=training)
ts_classifier_out = self.ts_classifier(ts_enc_out)
vhr_enc_out = self.vhr_encoder(vhr_input, training=training)
vhr_classifier_out = self.vhr_classifier(vhr_enc_out)
concat_out = self.concat([ts_enc_out, vhr_enc_out])
fusion_classifier_out = self.fusion_classifier(concat_out,
training=training)
grl = self.grl(concat_out, lamb_da)
domain_classifier_out = self.domain_classifier(grl, training=training)
return concat_out, ts_classifier_out, vhr_classifier_out, \
fusion_classifier_out, domain_classifier_out