def build_wgan_discriminator(self,
batch_local,
batch_global,
reuse=False,
training=True,
calc_perceptual_loss=False,
losses=None):
with tf.variable_scope('discriminator', reuse=reuse):
if (calc_perceptual_loss):
_, _, feats_local, _, dlocal = self.build_wgan_local_discriminator_verbose(
batch_local, reuse=reuse, training=training)
_, _, feats_global, _, dglobal = self.build_wgan_global_discriminator_verbose(
batch_global, reuse=reuse, training=training)
with tf.variable_scope('perceptual_loss', reuse=reuse):
feats_local_flat = flatten(feats_local, "flatten_local")
feats_global_flat = flatten(feats_global, "flatten_global")
fl_neg, fl_pos = tf.split(feats_local_flat, 2)
if ('perceptual_loss' not in losses):
losses['perceptual_loss'] = 0
losses['perceptual_loss'] += self.get_perceptual_loss(
fl_pos, fl_neg, name="loss_local")
fg_neg, fg_pos = tf.split(feats_global_flat, 2)
losses['perceptual_loss'] += self.get_perceptual_loss(
fg_pos, fg_neg, name="loss_global")
else:
dlocal = self.build_wgan_local_discriminator(batch_local,
reuse=reuse,
training=training)
dglobal = self.build_wgan_global_discriminator(
batch_global, reuse=reuse, training=training)
dout_local = tf.layers.dense(dlocal, 1, name='dout_local_fc')
dout_global = tf.layers.dense(dglobal, 1, name='dout_global_fc')
return dout_local, dout_global
def build_wgan_global_discriminator(self, x, reuse=False, training=True):
with tf.variable_scope('discriminator_global', reuse=reuse):
cnum = 64
x = dis_conv(x, cnum, name='conv1', training=training)
x = dis_conv(x, cnum*2, name='conv2', training=training)
x = dis_conv(x, cnum*4, name='conv3', training=training)
x = dis_conv(x, cnum*4, name='conv4', training=training)
x = flatten(x, name='flatten')
return x
def build_sn_patch_gan_discriminator(self, x, reuse=False, training=True):
with tf.compat.v1.variable_scope('sn_patch_gan', reuse=reuse):
cnum = 64
x = dis_conv(x, cnum, name='conv1', training=training)
x = dis_conv(x, cnum*2, name='conv2', training=training)
x = dis_conv(x, cnum*4, name='conv3', training=training)
x = dis_conv(x, cnum*4, name='conv4', training=training)
x = dis_conv(x, cnum*4, name='conv5', training=training)
x = dis_conv(x, cnum*4, name='conv6', training=training)
x = flatten(x, name='flatten')
return x
def build_wgan_global_discriminator_verbose(self,
x,
reuse=False,
training=True):
with tf.variable_scope('discriminator_global', reuse=reuse):
cnum = 64
x1 = dis_conv(x, cnum, name='conv1', training=training)
x2 = dis_conv(x1, cnum * 2, name='conv2', training=training)
x3 = dis_conv(x2, cnum * 4, name='conv3', training=training)
x4 = dis_conv(x3, cnum * 4, name='conv4', training=training)
x5 = flatten(x4, name='flatten')
return x1, x2, x3, x4, x5
