def snconv(self, x, filters, kernel_size, strides=1, padding='same', use_bias=True,
kernel_initializer=tc.layers.xavier_initializer(), name=None):
""" Spectral normalized convolutional layer """
name = self.get_name(name, 'snconv')
if isinstance(kernel_size, list):
assert_colorize(len(kernel_size) == 2)
H, W = kernel_size
else:
assert_colorize(isinstance(kernel_size, int))
H = W = kernel_size
with tf.variable_scope(name):
if padding.lower() != 'same' and padding.lower() != 'valid':
x = tf_utils.padding(x, kernel_size, strides, mode=padding)
padding = 'valid'
w = tf.get_variable('weight', shape=[H, W, x.shape[-1], filters],
initializer=kernel_initializer,
regularizer=self.l2_regularizer)
w = tf_utils.spectral_norm(w)
x = tf.nn.conv2d(x, w, strides=(1, strides, strides, 1), padding=padding.upper())
if use_bias:
b = tf.get_variable('bias', [filters], initializer=tf.zeros_initializer())
x = tf.nn.bias_add(x, b)
return x
def snconvtrans(self, x, filters, kernel_size, strides,
padding='same', use_bias=True,
kernel_initializer=tc.layers.xavier_initializer(), name=None):
name = self.get_name(name, 'snconvtrans')
if isinstance(kernel_size, list):
assert_colorize(len(kernel_size) == 2)
k_h, k_w = kernel_size
else:
assert_colorize(isinstance(kernel_size, int))
k_h = k_w = kernel_size
B, H, W, _ = x.shape.as_list()
# Compute output shape
if padding.lower() == 'valid':
output_shape = [B, (H-1) * strides + k_h, (W-1) * strides + k_w, filters]
else:
output_shape = [B, H * strides, W * strides, filters]
padding = 'SAME' # treat all other forms padding as same
with tf.variable_scope(name):
w = tf.get_variable('weight', shape=[k_h, k_w, filters, x.shape[-1]],
initializer=kernel_initializer,
regularizer=self.l2_regularizer)
w = tf_utils.spectral_norm(w)
x = tf.nn.conv2d_transpose(x, w,
output_shape=output_shape,
strides=[1, strides, strides, 1],
padding=padding.upper())
if use_bias:
b = tf.get_variable('bias', [filters], initializer=tf.zeros_initializer())
x = tf.nn.bias_add(x, b)
return x
def embedding(self, x, n_classes, embedding_size, sn, name='embedding'):
with tf.variable_scope(name):
embedding_map = tf.get_variable(name='embedding_map',
shape=[n_classes, embedding_size],
initializer=tc.layers.xavier_initializer())
if sn:
embedding_map_trans = tf_utils.spectral_norm(tf.transpose(embedding_map))
embedding_map = tf.transpose(embedding_map_trans)
return tf.nn.embedding_lookup(embedding_map, x)
def sndense(self, x, units, use_bias=True, kernel_initializer=tc.layers.xavier_initializer(), name=None):
name = self.get_name(name, 'sndense')
with tf.variable_scope(name):
w = tf.get_variable('weight', shape=[x.shape[-1], units],
initializer=kernel_initializer,
regularizer=self.l2_regularizer)
w = tf_utils.spectral_norm(w)
x = tf.matmul(x, w)
if use_bias:
b = tf.get_variable('bias', [units], initializer=tf.zeros_initializer())
x = x + b
return x
def call(self, x):
w = tf_utils.spectral_norm(self.kernel, self.u, self.iterations)
x = tf.nn.conv2d(x, w, strides=self.strides, padding=self.padding)
if self.use_bias:
x = tf.nn.bias_add(x, self.bias)
return x
def call(self, x):
w = tf_utils.spectral_norm(self.kernel, self.u, self.iterations)
x = tf.matmul(x, w)
if self.use_bias:
x = tf.nn.bias_add(x, self.bias)
return x