def block_e(x,res):
with tf.variable_scope('Gen_Enc%dx%d' % (2**res, 2**res)):
with tf.variable_scope('Conv'):
x = leaky_relu(batchnorm(apply_bias(conv2d(x, fmaps = nf(res-2), kernel = 3,
cf = self.channel_first), cf = self.channel_first), cf = self.channel_first))
x = downscale2d(x, cf = self.channel_first)
return x
def block_d(x,res):
layers = []
with tf.variable_scope('Gen_Dec_%dx%d' % (2**res, 2**res)):
x = upscale2d(x, cf = self.channel_first)
with tf.variable_scope('Conv'):
x = leaky_relu(batchnorm(apply_bias(conv2d(x, fmaps = nf(res-1), kernel = 3, cf = self.channel_first),
cf = self.channel_first), cf = self.channel_first))
return x
def fromrgb(x, res):
with tf.variable_scope('Gen_Enc_FromRGB_lod%d' %
(self.resolution_log2 - res)):
return leaky_relu(
apply_bias(conv2d(x,
fmaps=nf(res - 1),
kernel=1,
cf=self.channel_first),
cf=self.channel_first))
def block(x,res):
layers = []
with tf.variable_scope('Disc_%dx%d' % (2**res, 2**res)):
if res > 4:
with tf.variable_scope('Conv'):
x = leaky_relu(batchnorm(apply_bias(conv2d(x, fmaps = nf(res-2), kernel = 3,
cf = self.channel_first), cf = self.channel_first), cf = self.channel_first))
x = downscale2d(x, cf = self.channel_first)
else:
with tf.variable_scope('Patch'):
x = tf.sigmoid(apply_bias(conv2d(x, fmaps = 1, kernel = 3,
cf = self.channel_first), cf = self.channel_first))
return x