def Q(x, reuse=False):
with tf.variable_scope('infer', reuse=reuse):
feamap0 = lrelu(
conv2d(x, channelne[1], k=kersizee[0], s=1, name="conv2d_0"))
feamap1 = block_down(feamap0,
channelne[1],
k=kersizee[1],
s=1,
name='block1') # 32x32
feamap2 = resblock_down(feamap1,
channelne[2],
k=kersizee[2],
s=1,
name='block2') # 16x16
feamap3 = block_down(feamap2,
channelne[3],
k=kersizee[3],
s=1,
name='block3') # 8x8
feamap4 = block_down(feamap3,
channelne[4],
k=kersizee[4],
s=1,
name='block4') # 4x4
#feamap5 = lrelu(conv2d(feamap4, channelne[5],k=kersizee[4], s=1,name="conv2d_5")) # 4x4
#feamap5 = lrelu(conv2d(feamap5, channelne[5],k=4, s=1,padding='VALID',name="conv2d_6")) # 1x1
feamap5 = tf.reshape(feamap4,
[BATCH_SIZE, channelne[5] * fme[5] * fme[5]])
za = tf.nn.tanh(dense(feamap5, z_dim, gain=1, name='fully_f1'))
return za
def P(z, class_info, reuse=False, layer_condition=False):
with tf.variable_scope('gen', reuse=reuse):
#output_shape_da_conv5 = tf.stack([BATCH_SIZE, fmd[5], fmd[5], channelnd[5]])
z_generation = tf.concat([z, class_info], 1) # [z_dim + class_num, ]
ch = fmd[0] * fmd[0] * channelnd[0]
hf = dense(z_generation, ch, gain=np.sqrt(2) / 4, name='fully_f')
hc = tf.reshape(hf, [-1, fmd[0], fmd[0], channelnd[0]])
hc = tf.nn.relu(hc)
hc = tf.nn.relu(conv2d(hc, nzatte, k=kersized[0], s=1,
name="conv2d_0")) # 4x4
zattg = tf.reshape(hc, [BATCH_SIZE, fmd[0] * fmd[0], nzatte])
zattg = sparsech(zattg, int(nzatte * 0.1))
fmd1 = tf.zeros((BATCH_SIZE, fmd[0], fmd[0],
fmd[0] * fmd[0])) #?????????????????????????? ZEROS??
fmsparsel1 = sparseopattributeq(fmd1, zattg)
fmsparsel1 = tf.reshape(
fmsparsel1, [BATCH_SIZE, fmd[0], fmd[0], nzatte * fmd[0] * fmd[0]])
fmd2 = block_up_no(fmsparsel1,
channelnd[1],
k=kersized[0],
s=1,
name='block1',
class_info=class_info,
layer_condition=layer_condition) #8x8
fmsparsel2 = sparseop(fmd2, int(1 / 4 * fmd[1] * fmd[1]), 1)
fmd3 = block_up_no(fmsparsel2,
channelnd[2],
k=kersized[1],
s=1,
name='block2',
class_info=class_info,
layer_condition=layer_condition) #16x16
fmsparsel3 = sparseop(fmd3, int(1 / 4 * fmd[2] * fmd[2]), 1)
fmd4 = resblock_up_no(fmsparsel3,
channelnd[3],
k=kersized[2],
s=1,
name='block3',
class_info=class_info,
layer_condition=layer_condition) #32x32
fmsparsel4 = sparseop(fmd4, int(1 / 3 * fmd[3] * fmd[3]), 1)
fmd5 = block_up_no(fmsparsel4,
channelnd[4],
k=kersized[3],
s=1,
name='block4',
class_info=class_info,
layer_condition=layer_condition) #64x64
gx = tf.nn.tanh(
conv2d(fmd5, channelnd[5], k=kersized[4], s=1,
name="conv2d_5")) #64x64
return gx
def descriptor(x, reuse=False):
with tf.variable_scope('des', reuse=reuse):
feamap0 = lrelu(
conv2d(x, channelndes[1], k=kersizedes[0], s=1, name="conv2d_0"))
feamap1 = block_down(feamap0,
channelndes[1],
k=kersizedes[1],
s=1,
name='block1') # 32x32
feamap2 = resblock_down(feamap1,
channelndes[2],
k=kersizedes[2],
s=1,
name='block2') # 16x16
feamap3 = block_down(feamap2,
channelndes[3],
k=kersizedes[3],
s=1,
name='block3') # 8x8
feamap4 = block_down(feamap3,
channelndes[4],
k=kersizedes[4],
s=1,
name='block4') # 4x4
feamap4 = minibatch_stddev_layer(feamap4, 10)
feamap5 = lrelu(
conv2d(feamap4,
channelndes[5],
k=kersizedes[4],
s=1,
name="conv2d_5")) # 4x4
#feamap5 = lrelu(conv2d(feamap5, channelndes[5],k=4, s=1,padding='VALID',name="conv2d_6")) # 1x1
feamap5 = tf.reshape(
feamap5, [BATCH_SIZE, channelndes[5] * fmdes[5] * fmdes[5]])
out = dense(feamap5, 1, gain=1, name='fully_f')
return out
