def gan(self, x_image, reuse=False, features=False):
with tf.variable_scope('gan') as scope:
if reuse:
scope.reuse_variables()
c_i = x_image
c_i = lin_bn_lrelu(c_i, 1000, self.bn_settings, name="gan_dens_%d" % 0)
if features:
return lin(c_i, 1000, self.bn_settings, name="gan_dens_%d" % 1)
c_i = lin_bn_lrelu(c_i, 1000, self.bn_settings, name="gan_dens_%d" % 1)
y = lin(c_i, 1, "gan_out")
return y
def decoder(self, z, reuse=False):
with tf.variable_scope('decoder') as scope:
if reuse:
scope.reuse_variables()
z_dim = self.z_dim
if self.y_dim:
z = tf.concat(1, [z, self.y_labels])
z_dim = self.z_dim + self.y_dim
c_i = z
c_i = lin(c_i, 784, name='dec_lin_1')
# Now c_i has shape batch_size x 784
c_i = tf.maximum(0., c_i)
c_i = tf.reshape(c_i, [self.batch_size, 28, 28, 1])
# Now c_i has shape batch_size x 28 x 28 x 1
c_i = conv(c_i, 5, 1, 16, name="dec_conv_1")
c_i = tf.maximum(0., c_i)
# Now c_i has shape batch_size x 28 x 28 x 16
c_i = conv(c_i, 5, 1, 1, name="dec_conv_2")
# Now c_i has shape batch_size x 28 x 28 x 1
c_i = tf.reshape(c_i, shape=[self.batch_size, 784])
# Now c_i has shape batch_size x 784
y_image = tf.nn.sigmoid(c_i)
return y_image
def encoder(self):
c_i = self.x_image
c_i = lin_lrelu(c_i, 2048, name="enc_dens_%d" % 0)
c_i = lin_bn_lrelu(c_i, 2048, self.bn_settings, name="enc_dens_%d" % 1)
z = lin(c_i, self.z_dim, name="enc_out")
return z
def encoder(self):
c_i = self.x_image
c_i = tf.reshape(c_i, shape=[self.batch_size, 28, 28, 1])
# Now c_i has shape batch_size x 28 x 28 x 1
c_i = conv(c_i, 5, 1, 16, name="enc_conv_1")
c_i = tf.maximum(0., c_i)
# Now c_i has shape batch_size x 28 x 28 x 16
c_i = conv(c_i, 5, 1, 16, name="enc_conv_2")
c_i = tf.maximum(0., c_i)
# Now c_i has shape batch_size x 28 x 28 x 16
c_i = tf.reshape(c_i, shape=[self.batch_size, 16 * 784])
# Now c_i has shape batch_size x 784
c_i = lin(c_i, 400, name="enc_lin_1")
c_i = tf.maximum(0., c_i)
# Now c_i has shape batch_size x 400
z = lin(c_i, self.z_dim, name="enc_lin_2")
return z
def decoder(self, z, reuse=False):
with tf.variable_scope('decoder') as scope:
c_i = z
if self.y_dim:
c_i = tf.concat(1, [z, self.y_labels])
if reuse:
scope.reuse_variables()
c_i = lin_lrelu(c_i, 2048, name="dec_dens_%d" % 0)
c_i = lin_bn_lrelu(c_i, 2048, self.bn_settings, name="dec_dens_%d" % 1)
c_i = lin(c_i, self.input_dim, name="dec_out")
x_reconstructed = tf.nn.sigmoid(c_i)
return x_reconstructed
def discriminator(self, z, reuse=False):
with tf.variable_scope('discriminator') as scope:
if reuse:
scope.reuse_variables()
c_i = z
if self.y_dim:
c_i = tf.concat(1, [z, self.y_labels])
c_i = lin_lrelu(c_i, 500, name="disc_dens_%d" % 0)
c_i = lin_lrelu(c_i, 500, name="disc_dens_%d" % 1)
y = lin(c_i, 1, name="disc_out")
return y