forked from Ghadeer-Ghosheh/IGNITE
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathnew_networks_DA.py
317 lines (236 loc) · 16.6 KB
/
new_networks_DA.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
# -*- coding: utf-8 -*-
"""
Created on Wed Jun 21 14:37:23 2023
@author: gghos
"""
import tensorflow.compat.v1 as tf
tf.enable_eager_execution()
tf.config.run_functions_eagerly(True)
import os
import tensorflow as tf2
import warnings
warnings.filterwarnings("ignore")
tf2.random.set_seed(42)
seed = 42
import keras.backend as K
os.environ['PYTHONHASHSEED'] = '1'
os.environ['TF_DETERMINISTIC_OPS'] = '1'
from tensorflow.keras.layers import (LSTM, Permute)
from tensorflow.python.keras.regularizers import L2 as l2_regularizer
tf.compat.v1.disable_v2_behavior()
class observed_only_vae(object):
def __init__(self,
time_steps,
dim, z_dim,keep_prob,l2scale,
enc_size, dec_size,
conditional=True):
self.time_steps = time_steps
self.dim = dim
self.z_dim = z_dim
self.enc_size = enc_size
self.dec_size = dec_size
self.keep_prob = keep_prob
self.l2scale = l2scale
self.conditional = conditional
def build_vae(self, input_data, conditions=True):
tf2.random.set_seed(seed)
self.cell_enc = LSTM(self.enc_size, dropout =self.keep_prob, return_state=True,recurrent_dropout = 0.0)
self.cell_dec = LSTM(self.dec_size,dropout =self.keep_prob, return_state=True,recurrent_dropout = 0.0)
batch_size = K.shape(input_data)[0]
self.e =tf2.random.normal((batch_size, self.z_dim))
self.c, mu, logsigma, sigma, z, Alpha_t = [0] * self.time_steps, [0] * self.time_steps, [0] * self.time_steps, \
[0] * self.time_steps, [0] * self.time_steps , [0] * self.time_steps
self.w_mu, self.b_mu, self.w_sigma, self.b_sigma, self.w_h_dec, self.b_h_dec = self.buildSampling()
self.w_W_e,self.b_W_e,self.w_U_e,self.b_U_e, self.w_W_d,self.b_W_d,self.w_U_d,self.b_U_d, self.w_v_d,self.b_v_d, self.w_v_e,self.b_v_e= self.Dual_Attention()
self.enc_state = tf.zeros((batch_size, self.enc_size))
self.cell_state_enc = tf.zeros((batch_size, self.enc_size))
for t in range(self.time_steps):
with tf.compat.v1.variable_scope('observed_only_VAE_Feature_Attention', reuse=tf.compat.v1.AUTO_REUSE):
Alpha_t[t] = self.FeatureAttention(self.enc_state, self.cell_state_enc,input_data)
X_tilde_t = tf.multiply( Alpha_t[t], input_data[:, None, t, :])
if self.conditional:
X_tilde_t = tf.concat([X_tilde_t, conditions[:, None, t, :]], axis=-1)
with tf.compat.v1.variable_scope('observed_only_VAE_Encoder', regularizer=l2_regularizer(self.l2scale), reuse=tf.compat.v1.AUTO_REUSE):
tf2.random.set_seed(seed)
self.enc_state, _, self.cell_state_enc = self.cell_enc(X_tilde_t, initial_state=[self.enc_state, self.cell_state_enc])
mu[t] = tf.matmul( self.enc_state, self.w_mu) + self.b_mu
logsigma[t] = tf.matmul( self.enc_state, self.w_sigma) + self.b_sigma
sigma[t] = tf.exp(logsigma[t])
z[t] = mu[t] + sigma[t] * self.e
self.cell_state_dec = tf.zeros((batch_size, self.dec_size))
self.dec_state = tf.zeros((batch_size, self.dec_size))
x_encoded = tf.stack(z, axis=1)
print(x_encoded.shape)
for t in range(self.time_steps):
with tf.compat.v1.variable_scope('observed_only_VAE_Temporal_Attention', reuse=tf.compat.v1.AUTO_REUSE):
Beta_t = self.TemporalAttention( self.dec_state,self.cell_state_dec, x_encoded)
context_vector= tf.multiply(Beta_t[:,None, t,:], x_encoded[:, None, t, :])
if self.conditional:
context_vector = tf.concat([context_vector, conditions[:, None, t, :]], axis=-1)
with tf.compat.v1.variable_scope('observed_only_VAE_Decoder', regularizer=l2_regularizer(self.l2scale), reuse=tf.compat.v1.AUTO_REUSE):
tf2.random.set_seed(seed)
self.dec_state, _, self.cell_state_dec = self.cell_dec(context_vector, initial_state=[self.dec_state, self.cell_state_dec])
activation= tf.keras.layers.Activation("sigmoid")
self.c[t] = activation(tf.matmul(self.dec_state, self.w_h_dec) + self.b_h_dec)
self.decoded = tf.stack(self.c, axis=1)
return self.decoded, sigma, mu, logsigma, z
def buildSampling(self):
w_mu = self.variable([self.enc_size, self.z_dim], scope_name='Sampling_layer/observed_only_VAE', name='w_mu')
b_mu = self.variable([self.z_dim], scope_name='Sampling_layer/observed_only_VAE', name='b_mu')
w_sigma = self.variable([self.enc_size, self.z_dim], scope_name='Sampling_layer/observed_only_VAE', name='w_sigma')
b_sigma = self.variable([self.z_dim], scope_name='Sampling_layer/observed_only_VAE', name='b_sigma')
w_h_dec = self.variable([self.dec_size, self.dim], scope_name='Decoder/Linear/observed_only_VAE', name='w_h_dec')
b_h_dec = self.variable([self.dim], scope_name='Decoder/Linear/observed_only_VAE', name='b_h_dec')
return w_mu, b_mu, w_sigma, b_sigma, w_h_dec, b_h_dec
def Dual_Attention(self):
w_W_e = self.variable([self.enc_size*2,self.time_steps], scope_name='observed_only_VAE/FeatureAttention', name='w_W_e')
b_W_e = self.variable([self.time_steps], scope_name='observed_only_VAE/FeatureAttention', name='b_W_e')
w_U_e = self.variable([self.time_steps,self.time_steps], scope_name='observed_only_VAE/FeatureAttention', name='w_U_e')
b_U_e = self.variable([self.time_steps], scope_name='observed_only_VAE/FeatureAttention', name='b_U_e')
w_v_e = self.variable([self.time_steps*2,1], scope_name='observed_only_VAE/FeatureAttention', name='w_v_e')
b_v_e = self.variable([1], scope_name='observed_only_VAE/FeatureAttention', name='b_v_e')
w_W_d = self.variable([self.dec_size*2,self.dec_size], scope_name='observed_only_VAE/TemporalAttention', name='w_W_d')
b_W_d= self.variable([self.dec_size], scope_name='observed_only_VAE/TemporalAttention', name='b_W_d')
w_U_d = self.variable([self.z_dim,self.dec_size], scope_name='observed_only_VAE/TemporalAttention', name='w_U_d')
b_U_d = self.variable([self.dec_size], scope_name='observed_only_VAE/TemporalAttention', name='b_U_d')
w_v_d = self.variable([self.enc_size*2,1], scope_name='observed_only_VAE/TemporalAttention', name='w_v_d')
b_v_d = self.variable([1], scope_name='observed_only_VAE/TemporalAttention', name='b_v_d')
return w_W_e,b_W_e,w_U_e,b_U_e, w_W_d,b_W_d,w_U_d,b_U_d, w_v_d, b_v_d, w_v_e,b_v_e
def variable(self, shape, scope_name, name):
with tf.compat.v1.variable_scope(scope_name, reuse=tf.compat.v1.AUTO_REUSE):
tf2.random.set_seed(seed)
v = tf.compat.v1.get_variable(name=name, shape=shape, initializer=tf2.keras.initializers.GlorotUniform(seed = seed))
return v
def Discriminator(self,input_discriminator,IMM):
batch_size = K.shape(input_discriminator)[0]
inputs= tf.concat([input_discriminator, IMM], axis = -1)
w_d = self.variable([self.dec_size, self.dim], scope_name='observed_only_VAE/Discriminator_oo', name='w_d')
b_d = self.variable([self.dim], scope_name='observed_only_VAE/Discriminator_oo', name='b_d')
self.discrim = LSTM(self.enc_size, return_sequences=True,recurrent_dropout = 0.0)
enc_state = tf.zeros((batch_size, self.time_steps, self.enc_size))
with tf.variable_scope("observed_only_VAE/Discriminator_oo", regularizer=l2_regularizer(self.l2scale), reuse=tf.compat.v1.AUTO_REUSE):
enc_state = self.discrim(inputs)
x= tf.matmul(enc_state, w_d) + b_d
layer = tf.keras.layers.Activation('sigmoid')
out= layer(x)
return(out)
def FeatureAttention(self, hidden_state, cell_state, input_data):
n = input_data.shape[2]
hs = K.repeat(
tf.concat([hidden_state, cell_state], axis=-1), n)
W_e= tf.matmul(hs, self.w_W_e) + self.b_W_e
U_e= tf.matmul(Permute((2, 1))(input_data), self.w_U_e) + self.b_U_e
tanh = tf.math.tanh(tf.concat([W_e,U_e ], axis=-1))
v_e= tf.matmul(tanh, self.w_v_e) + self.b_v_e
return tf.nn.softmax(Permute((2, 1))(v_e))
def TemporalAttention(self,hidden_state,cell_state,X_encoded):
W_d= tf.matmul( K.repeat(tf.concat([hidden_state, cell_state], axis=-1), X_encoded.shape[1]), self.w_W_d) + self.b_W_d
U_d= tf.matmul(X_encoded, self.w_U_d) + self.b_U_d
l= tf.matmul(tf.math.tanh(tf.concat([W_d ,U_d], axis=-1)), self.w_v_d) + self.b_v_d
return tf.nn.softmax(l, axis=1)
class IMM_vae(object):
def __init__(self,
time_steps,dim, z_dim,keep_prob,l2scale,
enc_size, dec_size,
conditional=True):
self.time_steps = time_steps
self.dim = dim
self.z_dim = z_dim
self.enc_size = enc_size
self.dec_size = dec_size
self.keep_prob = keep_prob
self.l2scale = l2scale
self.conditional = conditional
def build_vae(self, input_data, conditions=True):
tf2.random.set_seed(seed)
batch_size = K.shape(input_data)[0]
self.cell_enc = LSTM(self.enc_size, dropout =self.keep_prob, return_state=True,recurrent_dropout = 0.0)
self.cell_dec = LSTM(self.dec_size,dropout =self.keep_prob, return_state=True,recurrent_dropout = 0.0)
self.e =tf2.random.normal((batch_size, self.z_dim))
self.c, mu, logsigma, sigma, z, Alpha_t = [0] * self.time_steps, [0] * self.time_steps, [0] * self.time_steps, \
[0] * self.time_steps, [0] * self.time_steps , [0] * self.time_steps
self.w_mu, self.b_mu, self.w_sigma, self.b_sigma, self.w_h_dec, self.b_h_dec = self.buildSampling()
self.w_W_e,self.b_W_e,self.w_U_e,self.b_U_e, self.w_W_d,self.b_W_d,self.w_U_d,self.b_U_d, self.w_v_d,self.b_v_d, self.w_v_e,self.b_v_e= self.Dual_Attention()
self.enc_state = tf.zeros((batch_size, self.enc_size))
self.cell_state_enc = tf.zeros((batch_size, self.enc_size))
for t in range(self.time_steps):
Alpha_t[t] = self.FeatureAttention(self.enc_state, self.cell_state_enc,input_data)
X_tilde_t = tf.multiply(Alpha_t[t], input_data[:, None, t, :])
if self.conditional:
X_tilde_t = tf.concat([X_tilde_t, conditions[:, None, t, :]], axis=-1)
with tf.compat.v1.variable_scope('IMM_VAE_Encoder', regularizer=l2_regularizer(self.l2scale), reuse=tf.compat.v1.AUTO_REUSE):
tf2.random.set_seed(seed)
self.enc_state, _, self.cell_state_enc = self.cell_enc(X_tilde_t, initial_state=[self.enc_state, self.cell_state_enc])
mu[t] = tf.matmul( self.enc_state, self.w_mu) + self.b_mu
logsigma[t] = tf.matmul( self.enc_state, self.w_sigma) + self.b_sigma
sigma[t] = tf.exp(logsigma[t])
z[t] = mu[t] + sigma[t] * self.e
self.cell_state_dec = tf.zeros((batch_size, self.dec_size))
self.dec_state = tf.zeros((batch_size, self.dec_size))
x_encoded = tf.stack(z, axis=1)
for t in range(self.time_steps):
Beta_t = self.TemporalAttention( self.dec_state,self.cell_state_dec, x_encoded)
context_vector= tf.multiply(Beta_t[:,None, t,:], x_encoded[:, None, t, :])
if self.conditional:
context_vector = tf.concat([context_vector, conditions[:, None, t, :]], axis=-1)
with tf.compat.v1.variable_scope('IMM_VAE_Decoder', regularizer=l2_regularizer(self.l2scale), reuse=tf.compat.v1.AUTO_REUSE):
tf2.random.set_seed(seed)
self.dec_state, _, self.cell_state_dec = self.cell_dec(context_vector, initial_state=[self.dec_state, self.cell_state_dec])
activation= tf.keras.layers.Activation("sigmoid")
self.c[t] = activation(tf.matmul(self.dec_state, self.w_h_dec) + self.b_h_dec)
self.decoded = tf.stack(self.c, axis=1)
return self.decoded, sigma, mu, logsigma, z
def buildSampling(self):
w_mu = self.variable([self.enc_size, self.z_dim], scope_name='Sampling_layer/IMM_VAE', name='w_mu')
b_mu = self.variable([self.z_dim], scope_name='Sampling_layer/IMM_VAE', name='b_mu')
w_sigma = self.variable([self.enc_size, self.z_dim], scope_name='Sampling_layer/IMM_VAE', name='w_sigma')
b_sigma = self.variable([self.z_dim], scope_name='Sampling_layer/IMM_VAE', name='b_sigma')
w_h_dec = self.variable([self.dec_size, self.dim], scope_name='Decoder/Linear/IMM_VAE', name='w_h_dec')
b_h_dec = self.variable([self.dim], scope_name='Decoder/Linear/IMM_VAE', name='b_h_dec')
return w_mu, b_mu, w_sigma, b_sigma, w_h_dec, b_h_dec
def variable(self, shape, scope_name, name):
with tf.compat.v1.variable_scope(scope_name, reuse=tf.compat.v1.AUTO_REUSE):
tf2.random.set_seed(seed)
v = tf.compat.v1.get_variable(name=name, shape=shape, initializer=tf2.keras.initializers.GlorotUniform(seed = seed))
return v
def Discriminator(self,input_discriminator, IMM):
batch_size = K.shape(input_discriminator)[0]
inputs= tf.concat([input_discriminator, IMM], axis = -1)
w_d = self.variable([self.dec_size, self.dim], scope_name='IMM_VAE/Discriminator_IMM', name='w_d')
b_d = self.variable([self.dim], scope_name='IMM_VAE/Discriminator_IMM', name='b_d')
self.discrim = LSTM(self.enc_size, return_sequences=True,recurrent_dropout = 0.0)
enc_state = tf.zeros((batch_size, self.time_steps, self.enc_size))
with tf.variable_scope("IMM_VAE/Discriminator_IMM", regularizer=l2_regularizer(self.l2scale), reuse=tf.compat.v1.AUTO_REUSE):
enc_state = self.discrim(inputs)
x= tf.matmul(enc_state, w_d) + b_d
layer = tf.keras.layers.Activation('sigmoid')
out= layer(x)
return(out)
def Dual_Attention(self):
w_W_e = self.variable([self.enc_size*2,self.time_steps], scope_name='IMM_VAE/FeatureAttention', name='w_W_e')
b_W_e = self.variable([self.time_steps], scope_name='IMM_VAE/FeatureAttention', name='b_W_e')
w_U_e = self.variable([self.time_steps,self.time_steps], scope_name='IMM_VAE/FeatureAttention', name='w_U_e')
b_U_e = self.variable([self.time_steps], scope_name='IMM_VAE/FeatureAttention', name='b_U_e')
w_v_e = self.variable([self.time_steps*2,1], scope_name='IMM_VAE/FeatureAttention', name='w_v_e')
b_v_e = self.variable([1], scope_name='IMM_VAE/FeatureAttention', name='b_v_e')
w_W_d = self.variable([self.dec_size*2,self.dec_size], scope_name='IMM_VAE/TemporalAttention', name='w_W_d')
b_W_d= self.variable([self.dec_size], scope_name='IMM_VAE/TemporalAttention', name='b_W_d')
w_U_d = self.variable([self.z_dim,self.dec_size], scope_name='IMM_VAE/TemporalAttention', name='w_U_d')
b_U_d = self.variable([self.dec_size], scope_name='IMM_VAE/TemporalAttention', name='b_U_d')
w_v_d = self.variable([self.enc_size*2,1], scope_name='IMM_VAE/TemporalAttention', name='w_v_d')
b_v_d = self.variable([1], scope_name='IMM_VAE/TemporalAttention', name='b_v_d')
return w_W_e,b_W_e,w_U_e,b_U_e, w_W_d,b_W_d,w_U_d,b_U_d, w_v_d, b_v_d, w_v_e,b_v_e
def FeatureAttention(self, hidden_state, cell_state, input_data):
n = input_data.shape[2]
hs = K.repeat(
tf.concat([hidden_state, cell_state], axis=-1), n)
W_e= tf.matmul(hs, self.w_W_e) + self.b_W_e
U_e= tf.matmul(Permute((2, 1))(input_data), self.w_U_e) + self.b_U_e
tanh = tf.math.tanh(tf.concat([W_e,U_e ], axis=-1))
v_e= tf.matmul(tanh, self.w_v_e) + self.b_v_e
return tf.nn.softmax(Permute((2, 1))(v_e))
def TemporalAttention(self,hidden_state,cell_state,X_encoded):
W_d= tf.matmul( K.repeat(tf.concat([hidden_state, cell_state], axis=-1), X_encoded.shape[1]), self.w_W_d) + self.b_W_d
U_d= tf.matmul(X_encoded, self.w_U_d) + self.b_U_d
l= tf.matmul(tf.math.tanh(tf.concat([W_d ,U_d], axis=-1)), self.w_v_d) + self.b_v_d
return tf.nn.softmax(l, axis=1)