-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathDACINModel.py
171 lines (148 loc) · 8.51 KB
/
DACINModel.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
import tensorflow as tf
from tensorflow.keras import layers
import os
class DACIN(object):
def __init__(self, network_architecture, learning_rate, batch_size, label_dim, model_path=None):
self.network_architecture = network_architecture
self.learning_rate = learning_rate
self.batch_size = batch_size
self.label_dim = label_dim
self.model_path = model_path
self.gen_mvs = self.generator_mvs()
self.dis_mvs = self.discriminator_mvs()
self.enc_clu = self.encoder_cluster()
self.dec_clu = self.decoder_cluster()
self.se_clu = self.self_expression()
self.var_se_list = []
for var in self.enc_clu.trainable_variables:
self.var_se_list.append(var)
for var in self.dec_clu.trainable_variables:
self.var_se_list.append(var)
for var in self.se_clu.trainable_variables:
self.var_se_list.append(var)
self.optimizer = tf.keras.optimizers.legacy.Adam(self.learning_rate)
# Generator for mvs
def generator_mvs(self):
# Functional model
inputs = tf.keras.Input(self.network_architecture['n_input']*2+self.label_dim)
G_h1 = layers.Dense(self.network_architecture['n_gen_1'], activation='relu')(inputs)
G_h2 = layers.Dense(self.network_architecture['n_gen_2'], activation='relu')(G_h1)
G_prob = layers.Dense(self.network_architecture['n_input'], activation='sigmoid')(G_h2)
model = tf.keras.Model(inputs=inputs, outputs=G_prob)
return model
# Discriminator for mvs
def discriminator_mvs(self):
# Functional model
inputs = tf.keras.Input(self.network_architecture['n_input']*2)
D_h1 = layers.Dense(self.network_architecture['n_dis_1'], activation='relu')(inputs)
D_h2 = layers.Dense(self.network_architecture['n_dis_2'], activation='relu')(D_h1)
D_prob = layers.Dense(self.network_architecture['n_input'], activation='sigmoid')(D_h2)
model = tf.keras.Model(inputs=inputs, outputs=D_prob)
return model
def encoder_cluster(self):
inputs = tf.keras.Input(self.network_architecture['n_input'])
E_h1 = layers.Dense(self.network_architecture['n_enc_1'], activation='relu')(inputs)
E_h2 = layers.Dense(self.network_architecture['n_enc_2'], activation='relu')(E_h1)
z = layers.Dense(self.network_architecture['n_z'], activation=None)(E_h2)
model = tf.keras.Model(inputs=inputs, outputs=z, name="enc_clu")
return model
def self_expression(self):
inputs = tf.keras.Input(self.batch_size)
outputs = layers.Dense(self.batch_size, activation=None)(inputs)
model = tf.keras.Model(inputs=inputs, outputs=outputs, name="se_clu")
return model
def decoder_cluster(self):
inputs = tf.keras.Input(self.network_architecture['n_z'])
D_h1 = layers.Dense(self.network_architecture['n_dec_1'], activation='relu')(inputs)
D_h2 = layers.Dense(self.network_architecture['n_dec_2'], activation='relu')(D_h1)
x_rec = layers.Dense(self.network_architecture['n_input'], activation='sigmoid')(D_h2)
model = tf.keras.Model(inputs=inputs, outputs=x_rec, name="dec_clu")
return model
def whole_network(self, X, M, H, C, is_training=False):
# generator
self.g_inputs = tf.concat(values=[X, M, C], axis=1) # concatenate by colume, size becomes n * (dim*2)
self.g_X = self.gen_mvs(self.g_inputs, training=is_training)
# Combine with observed data_optdigits
self.Hat_X = X * M + self.g_X * (1 - M)
# discriminator
self.d_inputs = tf.concat(values=[self.Hat_X, H], axis=1) # concatenate by colume, size becomes n * (dim*2)
self.D_prob = self.dis_mvs(self.d_inputs, training=is_training)
# encoder for clustering
self.z = self.enc_clu(self.Hat_X)
self.z_c = self.se_clu(tf.transpose(self.z))
self.z_c = tf.transpose(self.z_c)
self.dec_x = self.dec_clu(self.z_c)
def loss_optimizer(self, X, M):
# generator loss
self.g_loss_adversarial = -tf.reduce_mean((1 - M) * tf.math.log(self.D_prob + 1e-8))
self.g_loss_cmp = tf.reduce_sum((M * X - M * self.g_X) ** 2) / tf.reduce_sum(M)
self.g_loss_mvs = tf.reduce_sum(((1-M) * X - (1-M) * self.g_X) ** 2) / tf.reduce_sum(1-M)
#self.g_loss = self.g_loss_adversarial + 100*self.g_loss_cmp + 10*self.g_loss_mvs
self.g_loss = self.g_loss_adversarial + 100 * self.g_loss_cmp
# discriminator loss
self.d_loss = -tf.reduce_mean(M * tf.math.log(self.D_prob + 1e-8) + (1 - M) * tf.math.log(1. - self.D_prob + 1e-8))
# autoencoder loss
self.ae_loss_cmp = tf.reduce_sum((M * self.Hat_X - M * self.dec_x) ** 2) / tf.reduce_sum(M)
self.ae_loss_mvs = tf.reduce_sum(((1-M) * self.Hat_X - (1-M) * self.dec_x) ** 2) / tf.reduce_sum(1-M)
self.se_loss = tf.reduce_mean((self.z - self.z_c)**2)
self.se_coef = self.se_clu.trainable_variables[0]
self.se_coef = tf.convert_to_tensor(self.se_coef)
self.se_loss_coef = tf.reduce_mean(self.se_coef ** 2)
self.cluster_loss = self.ae_loss_cmp + self.ae_loss_mvs + self.se_loss + 100*self.se_loss_coef
def train_step(self, X, M, H, C, is_training):
with tf.GradientTape() as gen_tape, tf.GradientTape(persistent=True) as dis_tape, tf.GradientTape(persistent=True) as cluster_tape:
self.whole_network(X, M, H, C, is_training)
self.loss_optimizer(X, M)
for i in range(5):
self.gradients_of_dis = dis_tape.gradient(self.d_loss, self.dis_mvs.trainable_variables)
self.optimizer.apply_gradients(zip(self.gradients_of_dis, self.dis_mvs.trainable_variables))
self.gradients_of_gen = gen_tape.gradient(self.g_loss, self.gen_mvs.trainable_variables)
self.optimizer.apply_gradients(zip(self.gradients_of_gen, self.gen_mvs.trainable_variables))
self.gradients_of_clu = cluster_tape.gradient(self.cluster_loss, self.var_se_list)
self.optimizer.apply_gradients(zip(self.gradients_of_clu, self.var_se_list))
def save_model(self):
self.gen_mvs.save(self.model_path + "/gen_mvs", save_format='tf')
self.dis_mvs.save(self.model_path + "/dis_mvs", save_format='tf')
self.enc_clu.save(self.model_path + "/enc_clu", save_format='tf')
self.dec_clu.save(self.model_path + "/dec_clu", save_format='tf')
self.se_clu.save(self.model_path + "/se_clu", save_format='tf')
self.se_clu.save_weights(self.model_path + "/se_weights", save_format='tf')
def train_model(DACIN, miss_data_x, data_m, data_h, data_cluster, training_epochs):
# show the model structure
# DACINLabelOneHot.gen_mvs.summary()
# DACINLabelOneHot.dis_mvs.summary()
# DACINLabelOneHot.enc_clu.summary()
# DACINLabelOneHot.dec_clu.summary()
# DACINLabelOneHot.se_clu.summary()
# logs
logdir = os.path.join("./logs/")
summary_writer = tf.summary.create_file_writer(logdir)
tf.summary.trace_on(graph=True, profiler=True)
for epoch in range(training_epochs):
DACIN.train_step(miss_data_x, data_m, data_h, data_cluster, is_training=True)
display_step = 100
save_step = 99
if epoch % display_step == 0:
print("Epoch:", '%04d' % (epoch + 1), "g_loss=", "{:.9f}".format(DACIN.g_loss),
"d_loss=", "{:.9f}".format(DACIN.d_loss),
"cluster_loss=", "{:.9f}".format(DACIN.cluster_loss),
"g_loss_adversarial=", "{:.9f}".format(DACIN.g_loss_adversarial),
"g_loss_cmp=", "{:.9f}".format(DACIN.g_loss_cmp),
"g_loss_mvs=", "{:.9f}".format(DACIN.g_loss_mvs),
"ae_loss_cmp=", "{:.9f}".format(DACIN.ae_loss_cmp),
"ae_loss_mvs=", "{:.9f}".format(DACIN.ae_loss_mvs),
"se_loss=", "{:.9f}".format(DACIN.se_loss),
"se_loss_coef=", "{:.9f}".format(DACIN.se_loss_coef))
if epoch % save_step == 0:
DACIN.save_model()
def get_se_coef(model_path):
se_clu = tf.keras.models.load_model(model_path + "/se_clu", compile=True)
se_coef = se_clu.trainable_variables[0]
se_coef = se_coef.numpy()
return se_coef
def imp_res_get(miss_data_x, data_m, data_cluster, model_path):
gen_mvs = tf.keras.models.load_model(model_path + "/gen_mvs", compile=False)
g_inputs = tf.concat(values=[miss_data_x, data_m, data_cluster], axis=1)
g_X = gen_mvs(g_inputs, training=False)
imp_x = miss_data_x * data_m + g_X * (1 - data_m)
return imp_x, g_X