-
Notifications
You must be signed in to change notification settings - Fork 2
/
VAE.py
175 lines (139 loc) · 7.82 KB
/
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
"""
Variational Autoencoder (VAE) implementation
"""
# MIT License
#
# Copyright (c) 2019 David Sandberg
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
import tensorflow as tf
import numpy as np
import math
class VAE(object):
def __init__(self, input_dims, learning_rate, warmup_temp_ph, nrof_stochastic_units, nrof_mlp_units):
assert len(nrof_stochastic_units)==len(nrof_mlp_units)
self.nrof_stochastic_units = nrof_stochastic_units
self.nrof_mlp_units = nrof_mlp_units
self.input_dims = input_dims
self.learning_rate = learning_rate
self.warmup_temp_ph = warmup_temp_ph
def softlimit(self, x, limit=0.1):
return tf.math.log(tf.math.exp(x) + 1.0 + limit)
def dense(self, x, nrof_units, activation=None, training=True, use_batch_norm=True):
x = tf.compat.v1.layers.Flatten()(x)
x = tf.compat.v1.layers.Dense(units=nrof_units)(x)
if use_batch_norm:
x = tf.compat.v1.layers.BatchNormalization()(x, training=training)
x = x if activation is None else activation(x)
return x
def mlp(self, x, nrof_units, activation, nrof_layers=1, training=True):
for _ in range(nrof_layers):
x = self.dense(x, nrof_units=nrof_units, activation=activation, training=training)
return x
def sample(self, mu, sigma):
epsilon = tf.random.normal(tf.shape(mu), mean=0.0, stddev=1.0)
return mu + sigma*epsilon
def decoder_sample(self, z, mu, sigma, draw_dec_sample=False):
if draw_dec_sample:
return self.sample(mu, sigma)
return z
def log_normal2(self, x, mean, log_var, eps=1e-5):
c = - 0.5 * math.log(2*math.pi)
return c - log_var/2 - (x - mean)**2 / (2 * tf.math.exp(log_var) + eps)
def build_graph(self, x, is_training):
o = dict()
dbg = dict()
reuse = None if is_training else True
with tf.compat.v1.variable_scope('model', reuse=reuse):
nrof_layers = len(self.nrof_mlp_units)
p_mu, p_sigma = [None]*nrof_layers, [None]*nrof_layers
q_mu, q_sigma = [None]*nrof_layers, [None]*nrof_layers
z = [None]*nrof_layers
# Create encoder
h = x
for l in range(nrof_layers):
dbg['enc_prevh_%d' % l ] = h
h = self.mlp(h, self.nrof_mlp_units[l], activation=tf.nn.leaky_relu, nrof_layers=nrof_layers, training=is_training)
dbg['h_%d' % l] = h
""" prepare for bidirectional inference """
q_mu[l] = self.dense(h, self.nrof_stochastic_units[l], training=is_training)
q_sigma[l] = self.dense(h, self.nrof_stochastic_units[l], activation=self.softlimit, training=is_training)
z[l] = self.sample(q_mu[l], q_sigma[l])
h = z[l]
dbg['z_%d' % l] = z[l]
dbg['q_mu_%d' % l] = q_mu[l]
dbg['q_sigma_%d' % l] = q_sigma[l]
# Create decoder
for l in range(nrof_layers-1, -1, -1):
if l == nrof_layers-1:
""" At the highest latent layer, mu & sigma are identical to those output from encoder.
And making actual z is not necessary for the highest layer."""
mu, sigma = q_mu[l], q_sigma[l]
p_mu[l], p_sigma[l] = tf.zeros_like(mu), tf.ones_like(sigma)
else:
""" prior is developed from z of the above layer """
p_mu[l] = self.dense(h, self.nrof_stochastic_units[l], training=is_training)
p_sigma[l] = self.dense(h, self.nrof_stochastic_units[l], activation=self.softlimit, training=is_training)
dbg['p_mu_%d' % l] = p_mu[l]
dbg['p_sigma_%d' % l] = p_sigma[l]
dbg['z_%d' % l] = z[l]
h = self.mlp(z[l], self.nrof_mlp_units[l], activation=tf.nn.leaky_relu, nrof_layers=nrof_layers, training=is_training)
dbg['h_%d' % l] = h
nrof_features = np.prod(self.input_dims)
x_reconst = self.dense(h, nrof_features, activation=tf.nn.sigmoid, training=is_training, use_batch_norm=False) # Reconstruction
if len(self.input_dims)==3:
x_reconst = tf.reshape(x_reconst, (-1, self.input_dims[0], self.input_dims[1], self.input_dims[2]))
dbg['x_reconst'] = x_reconst
dbg['x_orig'] = x
# Calculate reconstruction loss
eps = 1e-5
x_reconst_clip = tf.clip_by_value(x_reconst, eps, 1-eps)
log_pxz = -tf.reduce_sum(tf.keras.losses.binary_crossentropy(x, x_reconst_clip), axis=[1,2])
o['log_px'] = tf.reduce_mean(log_pxz)
# Calculate ELBO, i.e. log P(X|Z) + temp*(log(P(Z|X)) - log(q(Z)))
log_qz_list = []
log_pz_list = []
for l in range(nrof_layers):
log_qz_l = self.log_normal2(z[l], q_mu[l], tf.math.log(q_sigma[l])*2)
log_pz_l = self.log_normal2(z[l], p_mu[l], tf.math.log(p_sigma[l])*2)
dbg['log_pz_l_%d' % l] = log_pz_l
dbg['log_qz_l_%d' % l] = log_qz_l
log_pz_list += [ tf.reduce_sum(log_pz_l, axis=1) ] # Sum over log probabilities
log_qz_list += [ tf.reduce_sum(log_qz_l, axis=1) ] # Sum over log probabilities
log_pz = tf.add_n(log_pz_list)
log_qz = tf.add_n(log_qz_list)
elbo = log_pxz + self.warmup_temp_ph * (log_pz - log_qz)
o['log_pz'] = [ tf.reduce_mean(w) for w in log_pz_list ] # Average over the batch
o['log_qz'] = [ tf.reduce_mean(w) for w in log_qz_list ] # Average over the batch
o['kl_tot'] = -tf.reduce_mean(log_pz - log_qz) # Average over the batch
o['kl'] = [ -tf.reduce_mean(lpz-lqz) for (lpz, lqz) in zip(log_pz_list, log_qz_list) ]
""" set losses """
o['elbo'] = tf.reduce_mean(elbo) # Average over the batch
loss = -o['elbo']
train_op = None
if is_training:
optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=self.learning_rate, beta1=0.5)
grads = optimizer.compute_gradients(loss) # Minimize -elbo
for i,(g,v) in enumerate(grads):
if g is not None:
grads[i] = (tf.clip_by_norm(g,5),v) # clip gradients
update_ops = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer.apply_gradients(grads)
return train_op, o, dbg