vae/vae_celeb.py at master · lchingen/vae · GitHub

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
import numpy as np
import tensorflow as tf
from tensorflow.layers import conv2d, conv2d_transpose
from tensorflow.layers import batch_normalization as BN
from tensorflow.layers import flatten
from tensorflow.layers import dense
from tensorflow.nn import relu, leaky_relu, sigmoid, tanh

from config import *

def ACT(inputs, act_fn):
    if act_fn == 'relu':
        act = relu(inputs)
    elif act_fn == 'lrelu':
        act = leaky_relu(inputs)
    elif act_fn == 'sigmoid':
        act = sigmoid(inputs)
    elif act_fn == 'tanh':
        act = tanh(inputs)
    else:
        act = inputs
    return act


def CONV(inputs, filters, kernel_size, strides, padding, is_transpose):
    if is_transpose:
        conv = conv2d_transpose(inputs=inputs,
                                filters=filters,
                                kernel_size=kernel_size,
                                strides=strides,
                                padding=padding)
    else:
        conv = conv2d(inputs=inputs,
                      filters=filters,
                      kernel_size=kernel_size,
                      strides=strides,
                      padding=padding)
    return conv


def CONV_BN_ACT(inputs, filters, kernel_size, strides, padding, act_fn, is_training, is_transpose):
    conv = CONV(inputs=inputs,
                filters=filters,
                kernel_size=kernel_size,
                strides=strides,
                padding=padding,
                is_transpose=is_transpose)

    norm = BN(inputs=conv, training=is_training)
    act  = ACT(inputs=norm, act_fn=act_fn)
    return act


def Encoder(inputs, is_training):
    conv_0 = CONV_BN_ACT(inputs=inputs,
                         filters=16,
                         kernel_size=[3,3],
                         strides=[1,1],
                         padding='same',
                         act_fn='lrelu',
                         is_training=is_training,
                         is_transpose=False)

    conv_1 = CONV_BN_ACT(inputs=conv_0,
                         filters=32,
                         kernel_size=[3,3],
                         strides=[2,2],
                         padding='same',
                         act_fn='lrelu',
                         is_training=is_training,
                         is_transpose=False)

    conv_2 = CONV_BN_ACT(inputs=conv_1,
                         filters=64,
                         kernel_size=[3,3],
                         strides=[2,2],
                         padding='same',
                         act_fn='lrelu',
                         is_training=is_training,
                         is_transpose=False)

    conv_3 = CONV_BN_ACT(inputs=conv_2,
                         filters=128,
                         kernel_size=[3,3],
                         strides=[2,2],
                         padding='same',
                         act_fn='lrelu',
                         is_training=is_training,
                         is_transpose=False)


    flat = flatten(conv_3)

    z_mean = dense(inputs=flat, units=z_dim)
    z_log_var = dense(inputs=flat, units=z_dim)
    return z_mean, z_log_var


def Decoder(z, is_training):
    upsample = dense(inputs=z,
                     units=input_dim[0]/8 * input_dim[1]/8 * 128,
                     activation=relu)

    reshaped = tf.reshape(upsample, [-1,8,8,128])

    tconv_0 = CONV_BN_ACT(inputs=reshaped,
                          filters=128,
                          kernel_size=[3,3],
                          strides=[2,2],
                          padding='same',
                          act_fn='relu',
                          is_training=is_training,
                          is_transpose=True)

    tconv_1 = CONV_BN_ACT(inputs=tconv_0,
                          filters=64,
                          kernel_size=[3,3],
                          strides=[2,2],
                          padding='same',
                          act_fn='relu',
                          is_training=is_training,
                          is_transpose=True)

    tconv_2 = CONV_BN_ACT(inputs=tconv_1,
                          filters=32,
                          kernel_size=[3,3],
                          strides=[2,2],
                          padding='same',
                          act_fn='relu',
                          is_training=is_training,
                          is_transpose=True)

    tconv_3 = CONV_BN_ACT(inputs=tconv_2,
                          filters=16,
                          kernel_size=[3,3],
                          strides=[1,1],
                          padding='same',
                          act_fn='relu',
                          is_training=is_training,
                          is_transpose=True)

    tconv_4 = CONV(inputs=tconv_3,
                   filters=3,
                   kernel_size=[3,3],
                   strides=[1,1],
                   padding='same',
                   is_transpose=True)

    act_0   = ACT(inputs=tconv_4,
                  act_fn='sigmoid')
    return act_0


def Vae(x, z, mode):
    # Training flag for BN
    if mode == 'TRAIN':
        is_training = True
    else:
        is_training = False

    # Encode
    z_mean, z_log_var = Encoder(x, is_training)

    # Sample (skip if only testing decoder)
    epsilon = tf.random_normal(tf.shape(z_mean))
    if mode == 'TRAIN' or mode =='TEST':
        z = z_mean + tf.exp(z_log_var) * epsilon
    elif mode == 'TEST':
        z = z_mean

    # Decode
    y = Decoder(z, is_training)
    return y, z_mean, z_log_var