def linear(input_,
output_size,
name="Linear",
stddev=0.01,
scale=1.0,
with_learnable_sn_scale=False,
with_sn=False,
bias_start=0.0,
with_w=False,
update_collection=None,
with_singular_values=False):
shape = input_.get_shape().as_list()
with tf.variable_scope(name):
scope_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
tf.get_variable_scope().name)
has_summary = any([('Matrix' in v.op.name) for v in scope_vars])
matrix = tf.get_variable("Matrix", [shape[1], output_size], tf.float32,
tf.random_normal_initializer(stddev=stddev))
mul = tf.matmul(input_, matrix)
bias = tf.get_variable("bias", [output_size],
initializer=tf.constant_initializer(bias_start))
if not has_summary:
variable_summaries({'b': bias})
variable_summaries({'W': matrix},
with_singular_values=with_singular_values)
if with_w:
return mul + bias, matrix, bias
else:
return mul + bias
def linear_one_hot(input_,
output_size,
num_classes,
name="Linear_one_hot",
stddev=0.01,
scale=1.0,
with_learnable_sn_scale=False,
with_sn=False,
bias_start=0.0,
with_w=False,
update_collection=None,
with_singular_values=False):
with tf.variable_scope(name):
scope_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
tf.get_variable_scope().name)
has_summary = any([('Matrix' in v.op.name) for v in scope_vars])
matrix = tf.get_variable("Matrix", [num_classes, output_size],
tf.float32,
tf.random_normal_initializer(stddev=stddev))
embed = tf.nn.embedding_lookup(matrix, input_)
if not has_summary:
variable_summaries({'W': matrix},
with_singular_values=with_singular_values)
if with_w:
return embed, matrix
else:
return embed
def __init__(self,
epsilon=1e-5,
momentum=0.9,
name="batch_norm",
format='NCHW'):
with tf.variable_scope(name):
self.epsilon = epsilon
self.momentum = momentum
self.name = name
if format == 'NCHW':
self.axis = 1
elif format == 'NHWC':
self.axis = 3
def deconv2d(input_,
output_shape,
k_h=5,
k_w=5,
d_h=2,
d_w=2,
stddev=0.02,
scale=1.0,
with_learnable_sn_scale=False,
with_sn=False,
name="deconv2d",
with_w=False,
update_collection=None,
data_format='NCHW',
with_singular_values=False):
with tf.variable_scope(name):
scope_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
tf.get_variable_scope().name)
has_summary = any([('w' in v.op.name) for v in scope_vars])
out_channel, in_channel = get_in_out_shape(
output_shape,
input_.get_shape().as_list(), data_format)
strides = get_strides(d_h, d_w, data_format)
# filter : [height, width, output_channels, in_channels]
w = tf.get_variable(
'w', [k_h, k_w, out_channel, in_channel],
initializer=tf.random_normal_initializer(stddev=stddev))
deconv = tf.nn.conv2d_transpose(input_,
w,
output_shape=output_shape,
strides=strides,
data_format=data_format)
biases = tf.get_variable('biases', [out_channel],
initializer=tf.constant_initializer(0.0))
deconv = tf.reshape(
tf.nn.bias_add(deconv, biases, data_format=data_format),
deconv.get_shape())
if not has_summary:
variable_summaries({'b': biases})
variable_summaries({'W': w},
with_singular_values=with_singular_values)
if with_w:
return deconv, w, biases
else:
return deconv
def conv2d(input_,
output_dim,
k_h=5,
k_w=5,
d_h=2,
d_w=2,
stddev=0.02,
scale=1.0,
with_learnable_sn_scale=False,
with_sn=False,
name="snconv2d",
update_collection=None,
data_format='NCHW',
with_singular_values=False):
with tf.variable_scope(name):
scope_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
tf.get_variable_scope().name)
has_summary = any([('w' in v.op.name) for v in scope_vars])
out_channel, in_channel = get_in_out_shape(
[output_dim],
input_.get_shape().as_list(), data_format)
strides = get_strides(d_h, d_w, data_format)
w = tf.get_variable(
'w', [k_h, k_w, in_channel, out_channel],
initializer=tf.truncated_normal_initializer(stddev=stddev))
conv = tf.nn.conv2d(input_,
w,
strides=strides,
padding='SAME',
data_format=data_format)
biases = tf.get_variable('biases', [output_dim],
initializer=tf.constant_initializer(0.0))
conv = tf.reshape(
tf.nn.bias_add(conv, biases, data_format=data_format),
conv.get_shape())
if not has_summary:
variable_summaries({'b': biases})
variable_summaries({'W': w},
with_singular_values=with_singular_values)
return conv