def FCblock(self, name, x, chout, is_train, is_last=False):
with tf.variable_scope(name, reuse=self._reuse):
C = ops.linear(x, chout, 'linear')
if not is_last:
C = ops.norms(C, is_train, 'norm', self._norm)
C = ops.acts(C, self._activation)
return C
def __call__(self, z, y=None, is_train=True):
with tf.variable_scope(self.name, reuse=self._reuse):
if y is not None:
z = tf.concat([z, y], 1)
y = tf.reshape(y, [self._batch_size, 1, 1, y.shape[-1].value])
# 16*16*512
G = ops.linear(z,
16 * 16 * self._nf * 8,
'linear1',
use_sn=self._sn)
# 16*16*512
G = tf.reshape(G, [-1, 16, 16, self._nf * 8])
G = ops.norms(G, is_train, 'norm1', self._norm)
G = ops.acts(G, self._activation)
# 32 32 256
if y is not None:
G = ops.conv_cond_concat(G, y)
G = self.block('block1', G, self._nf * 4, 32, is_train)
# 64 64 128
if y is not None:
G = ops.conv_cond_concat(G, y)
G = self.block('block2', G, self._nf * 2, 64, is_train)
# 128 128 64
if y is not None:
G = ops.conv_cond_concat(G, y)
G = self.block('block3', G, self._nf, 128, is_train)
# 128 128 3
G = self.block('block4', G, 3, 128, is_train, True)
self._reuse = True
self.var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
self.name)
return G
def block(self, name, x, chout, is_train, is_norm=True, is_act=True):
with tf.variable_scope(name, reuse=self._reuse):
M = ops.linear(x, chout, 'linear')
if is_norm:
M = ops.norms(M, is_train, 'norm', self._norm)
if is_act:
M = ops.acts(M, self._activation)
return M
def __call__(self, x, is_train=True):
with tf.variable_scope(self.name, reuse=self._reuse):
# 14 14 128
C = self.block('block1', x, self._nf, is_train, False)
# 7 7 256
C = self.block('block2', C, self._nf * 2, is_train)
# 4 4 512
C = Cn = self.block('block3', C, self._nf * 4, is_train)
# 4*4*512
C = tf.reshape(C, [self._batch_size, -1])
# 10
C = ops.linear(C, self._num_labels, 'linear')
self._reuse = True
self.var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
self.name)
return C, Cn
def __call__(self, x, is_train=True):
with tf.variable_scope(self.name, reuse=self._reuse):
# 128 128 64 -> 64 64 128
D = self.block('block1', x, self._nf)
# 64 64 128 -> 32 32 256
D = self.block('block2', D, self._nf * 2)
# 32 32 256 -> 16 16 512
D = self.block('block3', D, self._nf * 4)
# 16 16 512
D = Cn = self.block('block4', D, self._nf * 8, True)
# 512 / GAP
D = tf.reduce_mean(D, axis=[1, 2])
# 1
D = ops.linear(D, 1, 'linear', use_sn=self._sn)
self._reuse = True
self.var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
self.name)
return D, Cn
def __call__(self, z, y=None, is_train=True):
with tf.variable_scope(self.name, reuse=self._reuse):
if y is not None:
z = tf.concat([z, y], 1)
# 4*4*512
G = ops.linear(z, 4 * 4 * self._nf * 4, 'linear')
# 4 4 512
G = tf.reshape(G, [-1, 4, 4, self._nf * 4])
G = ops.norms(G, is_train, 'norm', self._norm)
G = ops.acts(G, self._activation)
# 7 7 256
G = self.block('block1', G, self._nf * 2, 7, is_train)
# 14 14 128
G = self.block('block2', G, self._nf, 14, is_train)
# 28 28 1
G = self.block('block3', G, 1, 28, is_train, True)
self._reuse = True
self.var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
self.name)
return G