def create_discriminator(Xin, is_training, reuse=False, networktype='ganD'):
with tf.variable_scope(networktype, reuse=reuse):
Xout = conv(Xin,
is_training,
kernel_w=4,
stride=2,
pad=1,
Cout=128,
act='lrelu',
norm=None,
name='conv1') # 14
Xout = conv(Xout,
is_training,
kernel_w=4,
stride=2,
pad=1,
Cout=256,
act='lrelu',
norm='batchnorm',
name='conv2') # 7
Xout = conv(Xout,
is_training,
kernel_w=3,
stride=1,
pad=None,
Cout=1,
act=None,
norm='batchnorm',
name='conv4') # 5
Xout = tf.nn.sigmoid(Xout)
return Xout
def create_decoder(Xin,
is_training,
latentD,
Cout=1,
reuse=False,
networktype='cdaeD'):
with tf.variable_scope(networktype, reuse=reuse):
Xout = dense(Xin,
is_training,
Cout=7 * 7 * 256,
act='reLu',
norm='batchnorm',
name='dense1')
Xout = tf.reshape(Xout, shape=[-1, 7, 7, 256]) # 7
Xout = deconv(Xout,
is_training,
kernel_w=4,
stride=2,
Cout=256,
epf=2,
act='reLu',
norm='batchnorm',
name='deconv1') # 14
Xout = deconv(Xout,
is_training,
kernel_w=4,
stride=2,
Cout=Cout,
epf=2,
act=None,
norm=None,
name='deconv2') # 28
Xout = tf.nn.sigmoid(Xout)
return Xout
def create_generator(Xin,
is_training,
Cout=1,
reuse=False,
networktype='ganG'):
'''input : batchsize * latentD
output: batchsize * 28 * 28 * 1'''
with tf.variable_scope(networktype, reuse=reuse):
Xout = dense(Xin,
is_training,
Cout=7 * 7 * 256,
act='reLu',
norm='batchnorm',
name='dense1')
Xout = tf.reshape(Xout, shape=[-1, 7, 7, 256]) # 7
Xout = deconv(Xout,
is_training,
kernel_w=4,
stride=2,
epf=2,
Cout=128,
act='reLu',
norm='batchnorm',
name='deconv1') # 14
Xout = deconv(Xout,
is_training,
kernel_w=4,
stride=2,
epf=2,
Cout=Cout,
act=None,
norm=None,
name='deconv2') # 28
Xout = tf.nn.sigmoid(Xout)
return Xout
def create_encoder(Xin,
is_training,
latentD,
reuse=False,
networktype='cdaeE'):
'''Xin: batchsize * H * W * Cin
output1-2: batchsize * Cout'''
with tf.variable_scope(networktype, reuse=reuse):
Xout = conv(Xin,
is_training,
kernel_w=4,
stride=2,
Cout=64,
pad=1,
act='reLu',
norm='batchnorm',
name='conv1') # 14*14
Xout = conv(Xout,
is_training,
kernel_w=4,
stride=2,
Cout=128,
pad=1,
act='reLu',
norm='batchnorm',
name='conv2') # 7*7
Xout = dense(Xout,
is_training,
Cout=latentD,
act=None,
norm=None,
name='dense_mean')
return Xout
def create_gan_G(GE0, is_training, Cout=3, trainable=True, reuse=False, networktype='ganG'):
with tf.variable_scope(networktype, reuse=reuse):
GE1 = conv(GE0, is_training, kernel_w=4, stride=2, Cout=64 , pad=1, trainable=trainable, act='lreLu', norm=None, name='ENconv1') # 128
GE2 = conv(GE1, is_training, kernel_w=4, stride=2, Cout=128, pad=1, trainable=trainable, act='lreLu', norm='instance', name='ENconv2') # 64
GE3 = conv(GE2, is_training, kernel_w=4, stride=2, Cout=256, pad=1, trainable=trainable, act='lreLu', norm='instance', name='ENconv3') # 32
GE4 = conv(GE3, is_training, kernel_w=4, stride=2, Cout=512, pad=1, trainable=trainable, act='lreLu', norm='instance', name='ENconv4') # 16
GE5 = conv(GE4, is_training, kernel_w=4, stride=2, Cout=512, pad=1, trainable=trainable, act='lreLu', norm='instance', name='ENconv5') # 8
GE6 = conv(GE5, is_training, kernel_w=4, stride=2, Cout=512, pad=1, trainable=trainable, act='lreLu', norm='instance', name='ENconv6') # 4
GE7 = conv(GE6, is_training, kernel_w=4, stride=2, Cout=512, pad=1, trainable=trainable, act='lreLu', norm='instance', name='ENconv7') # 2
GBNeck = conv(GE7, is_training, kernel_w=4, stride=2, Cout=512, pad=1, trainable=trainable, act='lreLu', norm='instance', name='GBNeck') # 1 - Bottleneck
GD7 = deconv(GBNeck, is_training, kernel_w=4, stride=2, Cout=512, epf=2, trainable=trainable, act='reLu', norm='instance', name='DEdeconv1');GD7 = dropout(GD7, is_training, p=0.5); # 2
GD6 = deconv(conch(GD7, GE7), is_training, kernel_w=4, stride=2, Cout=512, epf=2, trainable=trainable, act='reLu', norm='instance', name='DEdeconv2');GD6 = dropout(GD6, is_training, p=0.5) # 4
GD5 = deconv(conch(GD6, GE6), is_training, kernel_w=4, stride=2, Cout=512, epf=2, trainable=trainable, act='reLu', norm='instance', name='DEdeconv3');GD5 = dropout(GD5, is_training, p=0.5) # 8
GD4 = deconv(conch(GD5, GE5), is_training, kernel_w=4, stride=2, Cout=512, epf=2, trainable=trainable, act='reLu', norm='instance', name='DEdeconv4') # 16
GD3 = deconv(conch(GD4, GE4), is_training, kernel_w=4, stride=2, Cout=512, epf=2, trainable=trainable, act='reLu', norm='instance', name='DEdeconv5') # 32
GD2 = deconv(conch(GD3, GE3), is_training, kernel_w=4, stride=2, Cout=256, epf=2, trainable=trainable, act='reLu', norm='instance', name='DEdeconv6') # 64
GD1 = deconv(conch(GD2, GE2), is_training, kernel_w=4, stride=2, Cout=128, epf=2, trainable=trainable, act='reLu', norm='instance', name='DEdeconv7') # 128
GD0 = deconv(conch(GD1, GE1), is_training, kernel_w=4, stride=2, Cout=Cout, epf=2, trainable=trainable, act=None, norm='instance', name='DEdeconv8') # 256
Xout = tf.nn.tanh(GD0)
return Xout
def create_gan_D(inSource, inTarget, is_training, trainable=True, reuse=False, networktype='ganD'):
with tf.variable_scope(networktype, reuse=reuse):
inSource = conch(inSource, inTarget)
Dxz = conv(inSource, is_training, kernel_w=4, stride=2, Cout=64, trainable=trainable, act='lrelu', norm=None, name='conv1') # 128
Dxz = conv(Dxz, is_training, kernel_w=4, stride=2, Cout=128, trainable=trainable, act='lrelu', norm='instance', name='conv2') # 64
Dxz = conv(Dxz, is_training, kernel_w=4, stride=2, Cout=256, trainable=trainable, act='lrelu', norm='instance', name='conv3') # 32
Dxz = conv(Dxz, is_training, kernel_w=1, stride=1, Cout=1, trainable=trainable, act='lrelu', norm='instance', name='conv4') # 32
Dxz = tf.nn.sigmoid(Dxz)
return Dxz
def create_gan_G(z,
is_training,
Cout=1,
trainable=True,
reuse=False,
networktype='ganG'):
'''input : batchsize * latentDim
output: batchsize * 28 * 28 * 1'''
with tf.variable_scope(networktype, reuse=reuse):
Gout = dense(z,
is_training,
Cout=7 * 7 * 256,
trainable=trainable,
act='reLu',
norm='batchnorm',
name='dense1')
Gout = tf.reshape(Gout, shape=[-1, 7, 7, 256]) # 7
Gout = deconv(Gout,
is_training,
kernel_w=4,
stride=2,
epf=2,
Cout=128,
trainable=trainable,
act='reLu',
norm='batchnorm',
name='deconv1') # 14
Gout = deconv(Gout,
is_training,
kernel_w=4,
stride=2,
epf=2,
Cout=Cout,
trainable=trainable,
act=None,
norm=None,
name='deconv2') # 28
Gout = tf.nn.sigmoid(Gout)
return Gout
def create_gan_D(xz,
is_training,
trainable=True,
reuse=False,
networktype='ganD'):
with tf.variable_scope(networktype, reuse=reuse):
Dout = conv(xz,
is_training,
kernel_w=4,
stride=2,
pad=1,
Cout=128,
trainable=trainable,
act='lrelu',
norm=None,
name='conv1') # 14
Dout = conv(Dout,
is_training,
kernel_w=4,
stride=2,
pad=1,
Cout=256,
trainable=trainable,
act='lrelu',
norm='batchnorm',
name='conv2') # 7
Dout = conv(Dout,
is_training,
kernel_w=3,
stride=1,
pad=None,
Cout=1,
trainable=trainable,
act=None,
norm='batchnorm',
name='conv4') # 5
Dout = tf.nn.sigmoid(Dout)
return Dout
def create_discriminator(Xin, is_training, reuse=False, networktype='ganD'):
with tf.variable_scope(networktype, reuse=reuse):
Xout = dense(Xin,
is_training,
Cout=7 * 7 * 256,
act='reLu',
norm='batchnorm',
name='dense1')
Xout = tf.reshape(Xout, shape=[-1, 7, 7, 256]) # 7
Xout = conv(Xout,
is_training,
kernel_w=3,
stride=1,
pad=1,
Cout=128,
act='lrelu',
norm='batchnorm',
name='conv1') # 7
Xout = conv(Xout,
is_training,
kernel_w=3,
stride=1,
pad=1,
Cout=256,
act='lrelu',
norm='batchnorm',
name='conv2') # 7
Xout = conv(Xout,
is_training,
kernel_w=3,
stride=1,
pad=None,
Cout=1,
act=None,
norm='batchnorm',
name='conv3') # 5
Xout = tf.nn.sigmoid(Xout)
return Xout
def conch(A, B):
'''Concatenate channelwise'''
with tf.variable_scope("deconv"):
X = tf.concat([A, B], axis=3)
return X