-
Notifications
You must be signed in to change notification settings - Fork 0
/
latent_model_vae.py
241 lines (175 loc) · 9.93 KB
/
latent_model_vae.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
# -*- coding: utf-8 -*-
"""
Created on Sat Oct 9 14:00:42 2021
@author: durmaz
"""
from __future__ import absolute_import, division, print_function, unicode_literals
import numpy as np
from sklearn.preprocessing import scale
import tensorflow as tf
def cust_loss(y_true, y_pred):
y_pred = tf.clip_by_value(y_pred, tf.keras.backend.epsilon(), 1.0-tf.keras.backend.epsilon())
log_loss = tf.math.add(tf.math.multiply(y_true, tf.math.log(y_pred)), tf.math.multiply(tf.math.multiply(1.0-y_true, tf.math.log(1.0-y_pred)), 0.1))
return tf.math.reduce_sum(tf.math.reduce_mean(-1.0*log_loss, axis=-1), axis=-1)
def cust_loss_mse(y_true, y_pred):
mse_err = tf.keras.losses.mean_squared_error(y_true, y_pred)
return tf.math.reduce_sum(mse_err, axis=-1)
class Sampling(tf.keras.layers.Layer):
def call(self, inputs):
z_mean, z_log_var = inputs
batch = tf.shape(z_mean)[0]
dim = tf.shape(z_mean)[1]
epsilon = tf.keras.backend.random_normal(shape=(batch, dim))
return z_mean + tf.exp(0.5 * z_log_var) * epsilon
class VariationalAutoEncoder(tf.keras.Model):
def __init__(self, enc, dec, name="autoencoder"):
super(VariationalAutoEncoder, self).__init__(name=name)
self.encoder = enc
self.decoder = dec
self.custom_weight = tf.Variable(name='kl_reg', initial_value=0.0, trainable=False, dtype="float32")
def call(self, inputs):
# Reconstruction Loss
z_mean, z_log_var, z, lat_expr_sec, sd_expr_sec, z_expr, lat_mut_sec, sd_mut_sec, z_mut, lat_psi_sec, sd_psi_sec, z_psi = self.encoder(inputs)
reconstructed = self.decoder([z, z_expr, z_mut, z_psi])
# Add KL divergence regularization loss.
kl_loss = -0.5 * tf.reduce_mean(
z_log_var - tf.square(z_mean) - tf.exp(z_log_var) + 1.0
)
kl_loss_expr = -0.5 * tf.reduce_mean(
sd_expr_sec - tf.square(lat_expr_sec) - tf.exp(sd_expr_sec) + 1.0
)
kl_loss_mut = -0.5 * tf.reduce_mean(
sd_mut_sec - tf.square(lat_mut_sec) - tf.exp(sd_mut_sec) + 1.0
)
kl_loss_psi = -0.5 * tf.reduce_mean(
sd_psi_sec - tf.square(lat_psi_sec) - tf.exp(sd_psi_sec) + 1.0
)
kl_loss_comb = tf.math.reduce_sum([kl_loss, kl_loss_expr, kl_loss_mut, kl_loss_psi])
self.add_loss(self.custom_weight*kl_loss_comb)
self.add_metric(self.custom_weight*kl_loss_comb, name='kl_loss')
return reconstructed
class EncoderModel(tf.keras.Model):
def __init__(self, latent_dim, name="encoder"):
super(EncoderModel, self).__init__(name=name)
self.z_expr = tf.keras.layers.Dense(latent_dim, activation='linear', name='expr_lat')
self.z_mut = tf.keras.layers.Dense(latent_dim, activation='linear', name='mut_lat')
self.z_psi = tf.keras.layers.Dense(latent_dim, activation='linear', name='psi_lat')
self.z_expr_sd = tf.keras.layers.Dense(latent_dim, activation='linear', name='expr_sd')
self.z_mut_sd = tf.keras.layers.Dense(latent_dim, activation='linear', name='mut_sd')
self.z_psi_sd = tf.keras.layers.Dense(latent_dim, activation='linear', name='psi_sd')
self.z_expr_sec = tf.keras.layers.Dense(latent_dim, activation='linear', name='expr_lat')
self.z_mut_sec = tf.keras.layers.Dense(latent_dim, activation='linear', name='mut_lat')
self.z_psi_sec = tf.keras.layers.Dense(latent_dim, activation='linear', name='psi_lat')
self.z_expr_sd_sec = tf.keras.layers.Dense(latent_dim, activation='linear', name='expr_sd')
self.z_mut_sd_sec = tf.keras.layers.Dense(latent_dim, activation='linear', name='mut_sd')
self.z_psi_sd_sec = tf.keras.layers.Dense(latent_dim, activation='linear', name='psi_sd')
self.sampling = Sampling()
def call(self, inputs):
lat_expr = self.z_expr(inputs[0])
lat_mut = self.z_mut(inputs[1])
lat_psi = self.z_psi(inputs[2])
sd_expr = self.z_expr_sd(inputs[0])
sd_mut = self.z_mut_sd(inputs[1])
sd_psi = self.z_psi_sd(inputs[2])
lat_expr_sec = self.z_expr_sec(inputs[0])
lat_mut_sec = self.z_mut_sec(inputs[1])
lat_psi_sec = self.z_psi_sec(inputs[2])
sd_expr_sec = self.z_expr_sd_sec(inputs[0])
sd_mut_sec = self.z_mut_sd_sec(inputs[1])
sd_psi_sec = self.z_psi_sd_sec(inputs[2])
z_mean = tf.math.add(tf.math.add(lat_expr, lat_mut), lat_psi)
z_log_var = tf.math.add(tf.math.add(sd_expr, sd_mut), sd_psi)
z = self.sampling([z_mean, z_log_var])
z_expr = self.sampling([lat_expr_sec, sd_expr_sec])
z_mut = self.sampling([lat_mut_sec, sd_mut_sec])
z_psi = self.sampling([lat_psi_sec, sd_psi_sec])
return z_mean, z_log_var, z, lat_expr_sec, sd_expr_sec, z_expr, lat_mut_sec, sd_mut_sec, z_mut, lat_psi_sec, sd_psi_sec, z_psi
class DecoderModel(tf.keras.Model):
def __init__(self, EXPR_DIM, MUT_DIM, PSI_DIM, name="decoder"):
super(DecoderModel, self).__init__(name=name)
self.out_expr = tf.keras.layers.Dense(EXPR_DIM, activation='linear')
self.out_mut = tf.keras.layers.Dense(MUT_DIM, activation='linear')
self.out_psi = tf.keras.layers.Dense(PSI_DIM, activation='linear')
self.out_expr_sec = tf.keras.layers.Dense(EXPR_DIM, activation='linear')
self.out_mut_sec = tf.keras.layers.Dense(MUT_DIM, activation='linear')
self.out_psi_sec = tf.keras.layers.Dense(PSI_DIM, activation='linear')
def call(self, inputs):
out_expr = self.out_expr(inputs[0])
out_mut = self.out_mut(inputs[0])
out_psi = self.out_psi(inputs[0])
out_expr_sec = self.out_expr_sec(inputs[1])
out_mut_sec = self.out_mut_sec(inputs[2])
out_psi_sec = self.out_psi_sec(inputs[3])
d_expr = tf.math.add(out_expr, out_expr_sec)
d_mut = tf.math.sigmoid(tf.math.add(out_mut, out_mut_sec))
d_psi = tf.math.add(out_psi, out_psi_sec)
return d_expr, d_mut, d_psi
class ModifyKLWeight(tf.keras.callbacks.Callback):
def __init__(self):
super(ModifyKLWeight, self).__init__()
def on_epoch_end(self, epoch, logs=None):
if not hasattr(self.model, 'custom_weight'):
raise ValueError('Does not have weight attribute for KL')
custom_weight = tf.keras.backend.get_value(self.model.custom_weight)
if (epoch % 5 == 0) & (epoch > 1):
tf.keras.backend.set_value(self.model.custom_weight, 0.0)
else:
if(custom_weight < 1.0): # Keep in case threshold needed
tf.keras.backend.set_value(self.model.custom_weight, tf.keras.backend.get_value(self.model.custom_weight) + 0.1)
print("\nEpoch %d: KL Weight is %.4f." % (epoch, tf.keras.backend.get_value(self.model.custom_weight)))
def sp_ae(data=None, dim=None, doCV=False, n_epoch=None, m_save=False):
# Autoencoder
BATCH_SIZE = 64
if n_epoch is None:
NUM_EPOCHS = 10000
else:
NUM_EPOCHS = n_epoch
LAT_DIM = dim
N_DIM_EXP=data[0].shape[1]
N_DIM_MUT=data[1].shape[1]
N_DIM_PSI=data[2].shape[1]
# Model
enc = EncoderModel(LAT_DIM)
dec = DecoderModel(N_DIM_EXP, N_DIM_MUT, N_DIM_PSI)
vae = VariationalAutoEncoder(enc, dec)
if doCV is True:
# Callback
red_lr = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=0.1,patience=35,verbose=1,mode='auto', min_lr=1e-12)
cust_cback = ModifyKLWeight()
# Train
rand_idx = np.random.permutation(data[0].shape[0])
optimizer = tf.keras.optimizers.Adam(learning_rate=1e-4)
vae.compile(optimizer=optimizer, loss=[cust_loss_mse,cust_loss,cust_loss_mse])
hist = vae.fit([data[0][rand_idx,], data[1][rand_idx,], data[2][rand_idx,]], [data[0][rand_idx,], data[1][rand_idx,], data[2][rand_idx,]], epochs=NUM_EPOCHS, batch_size=BATCH_SIZE, shuffle=True, validation_split=0.2, callbacks=[cust_cback, red_lr])
else:
cust_cback = ModifyKLWeight()
rand_idx = np.random.permutation(data[0].shape[0])
optimizer = tf.keras.optimizers.Adam(learning_rate=1e-4)
vae.compile(optimizer=optimizer, loss=[cust_loss_mse,cust_loss,cust_loss_mse])
hist = vae.fit([data[0][rand_idx,], data[1][rand_idx,], data[2][rand_idx,]], [data[0][rand_idx,], data[1][rand_idx,], data[2][rand_idx,]], epochs=NUM_EPOCHS, batch_size=BATCH_SIZE, shuffle=True, callbacks=[cust_cback])
# Save
if m_save:
enc.save('/data/splicing_enc_model')
dec.save('/data/splicing_dec_model')
# Clear
tf.keras.backend.clear_session()
return hist.history
if __name__ == '__main__':
l_wd = 'Research/splicing'
# Load Data
psi = scale(np.loadtxt('{}/data/psi_mat.tsv.gz'.format(l_wd), delimiter='\t', dtype=np.float64))
gexp = scale(np.loadtxt('{}/data/expr_mat.tsv.gz'.format(l_wd), delimiter='\t', dtype=np.float64))
mut = np.loadtxt('{}/data/mut_mat.tsv.gz'.format(l_wd), delimiter='\t', dtype=np.float64)
# AE
test_dim = np.asarray([16, 32, 64, 128, 256, 512])
cv_res_psi = []
cv_res_expr = []
cv_res_mut = []
for dim in test_dim:
train_hist = sp_ae(data=[gexp, mut, psi], dim=dim, doCV=True, n_epoch=10)
cv_res_expr.append(train_hist['val_output_1_loss'])
cv_res_mut.append(train_hist['val_output_2_loss'])
cv_res_psi.append(train_hist['val_output_3_loss'])
#np.save('{}/data/CV_PSI_VAE.npy'.format(l_wd), arr=cv_res_psi)
#np.save('{}/data/CV_EXPR_VAE.npy'.format(l_wd), arr=cv_res_expr)
#np.save('{}/data/CV_MUT_VAE.npy'.format(l_wd), arr=cv_res_mut)