-
Notifications
You must be signed in to change notification settings - Fork 0
/
model_components.py
313 lines (273 loc) · 24.7 KB
/
model_components.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
import tensorflow as tf
import custom as cm
from utils import image
class DiscriminatorComponentGAN:
def __init__ (self, device, scope, type, training):
self.device = device
self.scope = scope
self.type = type
self.training = training
def build(self, image_input, label_input, model_scope, reuse=tf.AUTO_REUSE):
x = image_input
y = label_input
with tf.device(self.device):
with tf.name_scope(model_scope):
with tf.variable_scope(self.scope, reuse=reuse):
_x = x # first layer initialized - this is for preventing unwanted modification of the raw input data and/or for preventing unwanted passing of raw input to intermediate layers
_y = y # first layer initialized
_xs = x.shape.as_list()[1:] # shape of x initialized [height, width, channel]
# VANILLA GAN
if self.type=='gan':
_x = tf.layers.conv2d(inputs=_x, filters=64, kernel_size=[4,4], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='conv0')
_x = tf.nn.leaky_relu(_x, name='leakyrelu0')
_x = tf.layers.conv2d(inputs=_x, filters=128, kernel_size=[4,4], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='conv1')
_x = tf.layers.batch_normalization(inputs=_x, training=self.training, name='batchnorm0')
_x = tf.nn.leaky_relu(_x, name='leakyrelu1')
_x = tf.layers.flatten(inputs=_x, name='flatten')
_x = tf.layers.dense(inputs=_x, units=1024, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), name='fullyconnected0')
_x = tf.layers.batch_normalization(inputs=_x, training=self.training, name='batchnorm1')
_x = tf.nn.leaky_relu(_x, name='leakyrelu2')
_x = tf.layers.dense(inputs=_x, units=1, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), name='fullyconnected1')
_xl = _x
_x = tf.nn.sigmoid(_x, name='sigmoid')
# DEEP CONVOLUTIONAL GAN
elif self.type=='dcgan':
_x = tf.layers.conv2d(inputs=_x, filters=64, kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='conv0')
_x = tf.nn.leaky_relu(_x, name='leakyrelu0')
_x = tf.layers.conv2d(inputs=_x, filters=128, kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='conv1')
_x = tf.layers.batch_normalization(inputs=_x, training=self.training, name='batchnorm0')
_x = tf.nn.leaky_relu(_x, name='leakyrelu1')
_x = tf.layers.conv2d(inputs=_x, filters=256, kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='conv2')
_x = tf.layers.batch_normalization(inputs=_x, training=self.training, name='batchnorm_1')
_x = tf.nn.leaky_relu(_x, name='leakyrelu2')
_x = tf.layers.conv2d(inputs=_x, filters=512, kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='conv3')
_x = tf.layers.batch_normalization(inputs=_x, training=self.training, name='batchnorm_2')
_x = tf.nn.leaky_relu(_x, name='leakyrelu3')
_x = tf.layers.flatten(inputs=_x, name='flatten')
_x = tf.layers.dense(inputs=_x, units=1, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), name='fullyconnected0')
_xl = _x
_x = tf.nn.sigmoid(_x, name='sigmoid')
# SELF-ATTENTION GAN
elif self.type=='sagan':
raise NotImplementedError('{} is to be updated'.format(type))
# CONDITIONAL GAN
elif self.type=='cgan':
_x = tf.concat([_x, tf.tile(_y[:, tf.newaxis, tf.newaxis, :], multiples=[1, _xs[0], _xs[1], 1], name=self.scope+'_convconcat_tile0')], axis=3, name=self.scope+'_convconcat_concat0')
_x = tf.layers.conv2d(inputs=_x, filters=64, kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='conv0')
_x = tf.nn.leaky_relu(_x, name='leakyrelu0')
_x = tf.concat([_x, tf.tile(_y[:, tf.newaxis, tf.newaxis, :], multiples=[1, _xs[0]//2, _xs[1]//2, 1], name=self.scope+'_convconcat_tile1')], axis=3, name=self.scope+'_convconcat_concat1')
_x = tf.layers.conv2d(inputs=_x, filters=128, kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='conv1')
_x = tf.layers.batch_normalization(inputs=_x, training=self.training, name='batchnorm1')
_x = tf.nn.leaky_relu(_x, name='leakyrelu1')
_x = tf.reshape(_x, shape=[-1, (_xs[0]//4)*(_xs[1]//4)*128], name='reshape0')
_x = tf.concat([_x, _y], axis=1, name='image_label_concat0')
_x = tf.layers.dense(inputs=_x, units=256, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), name='fullyconnected0')
_x = tf.layers.batch_normalization(inputs=_x, training=self.training, name='batchnorm_2')
_x = tf.nn.leaky_relu(_x, name='leakyrelu2')
_x = tf.concat([_x, _y], axis=1, name='image_label_concat1')
_x = tf.layers.dense(inputs=_x, units=1, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), name='fullyconnected1')
_xl = _x
_x = tf.nn.sigmoid(_x, name='sigmoid')
# AUXILIARY CLASSIFIER GAN
elif self.type=='acgan':
_x = tf.layers.conv2d(inputs=_x, filters=16, kernel_size=[3,3], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='conv0')
_x = tf.nn.leaky_relu(_x, name='leakyrelu0')
_x = tf.layers.conv2d(inputs=_x, filters=32, kernel_size=[3,3], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=1, padding='SAME', name='conv1')
_x = tf.layers.batch_normalization(inputs=_x, training=self.training, name='batchnorm0')
_x = tf.nn.leaky_relu(_x, name='leakyrelu1')
_x = tf.layers.conv2d(inputs=_x, filters=64, kernel_size=[3,3], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='conv2')
_x = tf.layers.batch_normalization(inputs=_x, training=self.training, name='batchnorm_1')
_x = tf.nn.leaky_relu(_x, name='leakyrelu2')
_x = tf.layers.conv2d(inputs=_x, filters=128, kernel_size=[3,3], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=1, padding='SAME', name='conv3')
_x = tf.layers.batch_normalization(inputs=_x, training=self.training, name='batchnorm_2')
_x = tf.nn.leaky_relu(_x, name='leakyrelu3')
_x = tf.layers.conv2d(inputs=_x, filters=256, kernel_size=[3,3], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='conv4')
_x = tf.layers.batch_normalization(inputs=_x, training=self.training, name='batchnorm_4')
_x = tf.nn.leaky_relu(_x, name='leakyrelu4')
_x = tf.layers.conv2d(inputs=_x, filters=512, kernel_size=[3,3], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=1, padding='SAME', name='conv5')
_x = tf.layers.batch_normalization(inputs=_x, training=self.training, name='batchnorm_5')
_x = tf.nn.leaky_relu(_x, name='leakyrelu5')
_x = tf.layers.flatten(inputs=_x, name='flatten')
_x = tf.layers.dense(inputs=_x, units=1, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), name='fullyconnected0')
_xl = _x
_x = tf.nn.sigmoid(_x, name='sigmoid')
# SPECTRAL NORMALIZATION GAN
elif self.type=='sngan':
def resblock(inputs, filters, kernel_size, downsample, name):
x = inputs
_x = inputs
with tf.name_scope(name):
_x = tf.nn.relu(_x)
_x = tf.layers.conv2d(inputs=_x, filters=filters, kernel_size=kernel_size, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=1, padding='SAME')
_x = tf.nn.relu(_x)
_x = tf.layers.conv2d(inputs=_x, filters=filters, kernel_size=kernel_size, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=1, padding='SAME')
if downsample:
_x = tf.layers.average_pooling2d(inputs=_x, pool_size=kernel_size, strides=2, padding='SAME')
x = tf.layers.conv2d(inputs=x, filters=filters, kernel_size=kernel_size, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=1, padding='SAME')
x = tf.layers.average_pooling2d(inputs=x, pool_size=kernel_size, strides=2, padding='SAME')
_resblock_output = _x + x
return _resblock_output
_x = resblock(inputs=_x, filters=32, kernel_size=[3,3], downsample=True, name='resblock0')
_x = resblock(inputs=_x, filters=64, kernel_size=[3,3], downsample=True, name='resblock1')
_x = resblock(inputs=_x, filters=128, kernel_size=[3,3], downsample=True, name='resblock2')
_x = tf.concat([_x, tf.tile(_y[:, tf.newaxis, tf.newaxis, :], multiples=[1, _xs[0]//8, _xs[1]//8, 1])], axis=3)
_x = resblock(inputs=_x, filters=256, kernel_size=[3,3], downsample=True, name='resblock3')
_x = resblock(inputs=_x, filters=512, kernel_size=[3,3], downsample=True, name='resblock4')
_x = resblock(inputs=_x, filters=512, kernel_size=[3,3], downsample=False, name='resblock5')
_x = tf.nn.relu(_x)
_x = tf.reduce_sum(_x, axis=[1,2], name='global_sum_pooling')
_x = tf.layers.dense(inputs=_x, units=1, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02))
_xl = _x
_x = tf.nn.sigmoid(_x)
else:
raise ValueError('unknown gan model type: {}'.format(type))
self.model_output = _x, _xl
return self.model_output
class GeneratorComponentGAN:
def __init__ (self, device, scope, type, training):
self.device = device
self.scope = scope
self.type = type
self.training = training
def build(self, noise_input, image_input, label_input, model_scope, reuse=tf.AUTO_REUSE):
x = image_input # only used for checking the shape
y = label_input
z = noise_input
with tf.device(self.device):
with tf.name_scope(model_scope):
with tf.variable_scope(self.scope, reuse=reuse):
_z = z # first layer initialized
_y = y # first layer initialized
_x = x # first layer initialized
_xs = x.shape.as_list()[1:] # shape of x initialized [height, width, channel]
# VANILLA GAN
if self.type=='gan':
_z = tf.layers.dense(inputs=_z, units=1024, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), name='fullyconnected0')
_z = tf.layers.batch_normalization(inputs=_z, training=self.training, name='batchnorm0')
_z = tf.nn.relu(_z, name='relu0')
_z = tf.layers.dense(inputs=_z, units=(_xs[0]//4)*(_xs[1]//4)*128, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), name='fullyconnected1')
_z = tf.layers.batch_normalization(inputs=_z, training=self.training, name='batchnorm1')
_z = tf.nn.relu(_z, name='relu1')
_z = tf.reshape(_z, shape=[-1, _xs[0]//4, _xs[1]//4, 128], name='reshape_0')
_z = tf.layers.conv2d_transpose(inputs=_z, filters=64, kernel_size=[4,4], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='deconv0')
_z = tf.layers.batch_normalization(inputs=_z, training=self.training, name='batchnorm2')
_z = tf.nn.relu(_z, name='relu2')
_z = tf.layers.conv2d_transpose(inputs=_z, filters=_xs[2], kernel_size=[4,4], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='deconv1')
_z = tf.nn.sigmoid(_z, name='sigmoid')
_z = image.scale_out(_z)
# DEEP CONVOLUTIONAL GAN
elif self.type=='dcgan':
_z = tf.layers.dense(inputs=_z, units=(_xs[0]//16)*(_xs[1]//16)*512, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), name='fullyconnected0')
_z = tf.reshape(_z, shape=[-1, _xs[0]//16, _xs[1]//16, 512], name='reshape_0')
_z = tf.layers.batch_normalization(inputs=_z, training=self.training, name='batchnorm0')
_z = tf.nn.relu(_z, name='relu0')
_z = tf.layers.conv2d_transpose(inputs=_z, filters=256, kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='deconv0')
_z = tf.layers.batch_normalization(inputs=_z, training=self.training, name='batchnorm1')
_z = tf.nn.relu(_z, name='relu1')
_z = tf.layers.conv2d_transpose(inputs=_z, filters=128, kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='deconv1')
_z = tf.layers.batch_normalization(inputs=_z, training=self.training, name='batchnorm2')
_z = tf.nn.relu(_z, name='relu2')
_z = tf.layers.conv2d_transpose(inputs=_z, filters=64, kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='deconv2')
_z = tf.layers.batch_normalization(inputs=_z, training=self.training, name='batchnorm3')
_z = tf.nn.relu(_z, name='relu3')
_z = tf.layers.conv2d_transpose(inputs=_z, filters=_xs[2], kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='deconv3')
_z = tf.nn.tanh(_z, name='tanh0')
# SELF-ATTENTION GAN
elif self.type=='sagan':
raise NotImplementedError('{} is to be updated'.format(type))
# CONDITIONAL GAN
elif self.type=='cgan':
_z = tf.concat([_z, _y], axis=1, name='noise_label_concat0')
_z = tf.layers.dense(inputs=_z, units=1024, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), name='fullyconnected0')
_z = tf.layers.batch_normalization(inputs=_z, training=self.training, name='batchnorm0')
_z = tf.nn.relu(_z, name='relu0')
_z = tf.concat([_z, _y], axis=1, name='noise_label_concat1')
_z = tf.layers.dense(inputs=_z, units=(_xs[0]//4)*(_xs[1]//4)*128, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), name='fullyconnected1')
_z = tf.layers.batch_normalization(inputs=_z, training=self.training, name='batchnorm1')
_z = tf.nn.relu(_z, name='relu1')
_z = tf.reshape(_z, shape=[-1, _xs[0]//4, _xs[1]//4, 128], name='reshape0')
_z = tf.concat([_z, tf.tile(_y[:, tf.newaxis, tf.newaxis, :], multiples=[1, _xs[0]//4, _xs[1]//4, 1], name=self.scope+'_convconcat_tile0')], axis=3, name=self.scope+'_convconcat_concat0')
_z = tf.layers.conv2d_transpose(inputs=_z, filters=128, kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='deconv0')
_z = tf.layers.batch_normalization(inputs=_z, training=self.training, name='batchnorm2')
_z = tf.nn.relu(_z, name='relu2')
_z = tf.concat([_z, tf.tile(_y[:, tf.newaxis, tf.newaxis, :], multiples=[1, _xs[0]//2, _xs[1]//2, 1], name=self.scope+'_convconcat_tile1')], axis=3, name=self.scope+'_convconcat_concat1')
_z = tf.layers.conv2d_transpose(inputs=_z, filters=_xs[2], kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='deconv1')
_z = tf.nn.sigmoid(_z, name='sigmoid')
_z = image.scale_out(_z)
# AUXILIARY CLASSIFIER GAN
elif self.type=='acgan':
_z = tf.concat([_z, _y], axis=1, name='noise_label_concat0')
_z = tf.layers.dense(inputs=_z, units=(_xs[0]//8)*(_xs[1]//8)*384, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), name='fullyconnected0')
_z = tf.reshape(_z, shape=[-1, _xs[0]//8, _xs[1]//8, 384], name='reshape0')
_z = tf.nn.relu(_z, name='relu0')
_z = tf.layers.conv2d_transpose(inputs=_z, filters=192, kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='deconv0')
_z = tf.layers.batch_normalization(inputs=_z, training=self.training, name='batchnorm1')
_z = tf.nn.relu(_z, name='relu1')
_z = tf.layers.conv2d_transpose(inputs=_z, filters=96, kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='deconv1')
_z = tf.layers.batch_normalization(inputs=_z, training=self.training, name='batchnorm2')
_z = tf.nn.relu(_z, name='relu2')
_z = tf.layers.conv2d_transpose(inputs=_z, filters=_xs[2], kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME', name='deconv2')
_z = tf.nn.tanh(_z, name='tanh')
# SPECTRAL NORMALIZATION GAN
elif self.type=='sngan':
def resblock(inputs, filters, kernel_size, upsample, name):
z = inputs
_z = inputs
with tf.name_scope(name):
_z = cm.layers.conditional_batch_normalization(inputs=_z, training=self.training, y1=label_input, y2=tf.zeros_like(label_input), name=name+'_cond_batchnorm0')
# _z = tf.layers.batch_normalization(inputs=_z, training=self.training, name=name+'_batchnorm0')
_z = tf.nn.relu(_z)
if upsample:
_z = tf.layers.conv2d_transpose(inputs=_z, filters=filters, kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME')
z = tf.layers.conv2d_transpose(inputs=z, filters=filters, kernel_size=[5,5], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=2, padding='SAME')
# _z = tf.image.resize_nearest_neighbor(_z, [_z.shape.as_list()[1]*2, _z.shape.as_list()[2]*2])
# z = tf.image.resize_nearest_neighbor(z, [z.shape.as_list()[1]*2, z.shape.as_list()[2]*2])
z = tf.layers.conv2d(inputs=z, filters=filters, kernel_size=1, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=1, padding='SAME')
_z = tf.layers.conv2d(inputs=_z, filters=filters, kernel_size=kernel_size, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=1, padding='SAME')
_z = cm.layers.conditional_batch_normalization(inputs=_z, training=self.training, y1=label_input, y2=tf.zeros_like(label_input), name=name+'_cond_batchnorm1')
# _z = tf.layers.batch_normalization(inputs=_z, training=self.training, name=name+'_batchnorm1')
_z = tf.nn.relu(_z)
_z = tf.layers.conv2d(inputs=_z, filters=filters, kernel_size=kernel_size, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=1, padding='SAME')
_resblock_output = _z + z
return _resblock_output
_z = tf.layers.dense(inputs=_z, units=(_xs[0]//32)*(_xs[1]//32)*1024, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02))
_z = tf.reshape(_z, [-1, _xs[0]//32, _xs[1]//32, 1024])
_z = resblock(inputs=_z, filters=512, kernel_size=[3,3], upsample=True, name='resblock0')
_z = resblock(inputs=_z, filters=256, kernel_size=[3,3], upsample=True, name='resblock1')
_z = resblock(inputs=_z, filters=128, kernel_size=[3,3], upsample=True, name='resblock2')
_z = resblock(inputs=_z, filters=64, kernel_size=[3,3], upsample=True, name='resblock3')
_z = resblock(inputs=_z, filters=32, kernel_size=[3,3], upsample=True, name='resblock4')
_z = cm.layers.conditional_batch_normalization(inputs=_z, training=self.training, y1=label_input, y2=tf.zeros_like(label_input), name='cond_batchnorm0')
# _z = tf.layers.batch_normalization(inputs=_z, training=self.training, name='batchnorm0')
_z = tf.nn.relu(_z)
_z = tf.layers.conv2d(inputs=_z, filters=3, kernel_size=[3,3], kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), strides=1, padding='SAME')
_z = tf.nn.tanh(_z)
else:
raise ValueError('unknown gan model type: {}'.format(type))
self.model_output = _z
return self.model_output
class ClassifierComponentGAN:
def __init__ (self, device, scope, type, training):
self.device = device
self.scope = scope
self.type = type
self.training = training
def build(self, image_input, label_input, model_scope, reuse=tf.AUTO_REUSE):
x = image_input
y = label_input
with tf.device(self.device):
with tf.name_scope(model_scope):
with tf.variable_scope(self.scope, reuse=reuse):
_x = x # first layer initialized
_ys = y.shape.as_list()[1]
# AUXILIARY CLASSIFIER GAN
if self.type=='acgan':
_x = tf.layers.dense(inputs=_x, units=128, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), name='fullyconnected0')
_x = tf.layers.batch_normalization(inputs=_x, training=self.training, name='batchnorm0')
_x = tf.nn.relu(_x, name='relu0')
_x = tf.layers.flatten(inputs=_x, name='flatten')
_x = tf.layers.dense(inputs=_x, units=_ys, kernel_initializer=tf.initializers.truncated_normal(stddev=0.02), name='fullyconnected1')
else:
raise ValueError('unknown gan model type: {}'.format(type))
self.model_output = _x
return self.model_output