def deconv2D(inputs, kernel, output_channel,strides, shape, spectral_normed=False, weight_normed=False, batch_normed=False,
lr_normed=False, layer_normed=False, use_bias=False, instance_normed=False, name='conv'):
s = [1, strides, strides, 1]
with tf.variable_scope(name) as scope:
if scope_has_variables(scope):
scope.reuse_variables()
if not weight_normed:
w = tf.get_variable("w", [kernel, kernel, output_channel, inputs.get_shape()[-1]],
initializer=tf.contrib.layers.variance_scaling_initializer())
if spectral_normed:
w = spectral_norm(w)
if weight_normed:
deconv = WeightNorm(tf.keras.layers.Conv2DTranspose(output_channel, kernel, use_bias=use_bias, strides=strides, padding='SAME'))(inputs)
elif batch_normed:
deconv = tf.keras.layers.BatchNormalization(tf.keras.layers.Conv2DTranspose(output_channel, kernel, use_bias=use_bias, strides=strides, padding='SAME'))(inputs)
else:
deconv = tf.nn.conv2d_transpose(inputs, w, shape, strides=s, padding='SAME')
if layer_normed:
deconv = tf.contrib.layers.layer_norm(deconv)
if batch_normed:
deconv = tf.layers.batch_normalization(deconv)
if instance_normed:
deconv = instance_norm(deconv)
if lr_normed:
deconv = tf.nn.lrn(deconv, bias=0.00005)
return deconv
def conv2D(inputs, kernel, output_channel, stride, use_bias=False, name='conv_1', spectral_normed=True, weight_normed=False, batch_normed=False,
layer_normed=False, instance_normed=False, lr_normed=False, stddev=0.02, padding="SAME"):
with tf.variable_scope(name) as scope:
if scope_has_variables(scope):
scope.reuse_variables()
if not weight_normed:
w = tf.get_variable("w", [kernel, kernel, inputs.get_shape()[-1], output_channel],
initializer=tf.contrib.layers.variance_scaling_initializer())
if spectral_normed:
w = spectral_norm(w)
if weight_normed:
conv = WeightNorm(tf.keras.layers.Conv2D(output_channel, kernel, use_bias=use_bias, strides=stride, padding=padding))(inputs)
elif batch_normed:
conv = tf.keras.layers.BatchNormalization(tf.keras.layers.Conv2D(output_channel, kernel, use_bias=use_bias, strides=stride, padding=padding))(inputs)
else:
conv = tf.nn.conv2d(inputs, w, strides=[1, stride, stride, 1], padding=padding)
if layer_normed:
conv = tf.contrib.layers.layer_norm(conv)
if batch_normed:
conv = tf.layers.batch_normalization(conv)
if instance_normed:
conv = instance_norm(conv)
if lr_normed:
conv = tf.nn.lrn(conv, bias=0.00005)
return conv
def conv3_sn(inputs,
kernel,
output_channel,
stride,
use_bias=False,
name='conv_1',
spectral_normed=True,
stddev=0.02,
padding="SAME"):
with tf.variable_scope(name) as scope:
if scope_has_variables(scope):
scope.reuse_variables()
w = tf.get_variable(
"w",
[kernel, kernel, kernel,
inputs.get_shape()[-1], output_channel],
initializer=tf.contrib.layers.xavier_initializer())
if spectral_normed:
conv = tf.nn.conv3d(inputs,
spectral_norm(w),
strides=[1, stride, stride, stride, 1],
padding=padding)
else:
conv = tf.nn.conv3d(inputs,
w,
strides=[1, stride, stride, stride, 1],
padding=padding)
return conv
def deconv1D(inputs,
kernel,
output_channel,
stride,
shape,
name='conv',
spectral_normed=False,
weight_normed=False,
batch_normed=False,
batch_renormed=False,
lr_normed=False,
layer_normed=False,
use_bias=False,
instance_normed=False,
is_training=False):
with tf.variable_scope(name) as scope:
if scope_has_variables(scope):
scope.reuse_variables()
w = tf.get_variable("w",
[kernel, output_channel,
inputs.get_shape()[-1]],
initializer=tf.contrib.layers.xavier_initializer())
if spectral_normed:
w = spectral_norm(w)
deconv = tf.contrib.nn.conv1d_transpose(inputs,
w,
shape,
stride=stride,
padding='SAME')
if batch_renormed:
clip = {
'rmax': tf.constant(3, dtype=tf.float32),
'rmin': tf.constant(1 / 3, dtype=tf.float32),
'dmax': tf.constant(5, dtype=tf.float32)
}
deconv = tf.layers.batch_normalization(deconv,
training=is_training,
renorm=True,
renorm_clipping=clip)
if batch_normed:
deconv = tf.layers.batch_normalization(deconv,
training=is_training)
if layer_normed:
deconv = tf.contrib.layers.layer_norm(deconv)
if instance_normed:
deconv = instance_norm(deconv)
if lr_normed:
deconv = tf.nn.lrn(deconv, bias=0.00005)
return deconv
def conv1D(inputs,
kernel,
output_channel,
stride,
use_bias=False,
name='conv_1',
stddev=0.02,
padding="SAME",
spectral_normed=True,
batch_normed=False,
depthwise=False,
batch_renormed=False,
layer_normed=False,
instance_normed=False,
lr_normed=False,
is_training=False):
with tf.variable_scope(name) as scope:
if scope_has_variables(scope):
scope.reuse_variables()
w = tf.get_variable(
"w", [kernel, inputs.get_shape()[-1], output_channel],
initializer=tf.contrib.layers.xavier_initializer())
if spectral_normed:
w = spectral_norm(w)
conv = tf.nn.conv1d(inputs, w, stride=stride, padding=padding)
if batch_renormed:
clip = {
'rmax': tf.constant(3, dtype=tf.float32),
'rmin': tf.constant(1 / 3, dtype=tf.float32),
'dmax': tf.constant(5, dtype=tf.float32)
}
conv = tf.layers.batch_normalization(conv,
training=is_training,
renorm=True,
renorm_clipping=clip)
#conv = tf.layers.batch_normalization(conv)
if batch_normed:
conv = tf.layers.batch_normalization(conv, training=is_training)
if layer_normed:
conv = tf.contrib.layers.layer_norm(conv)
if instance_normed:
conv = instance_norm(conv)
if lr_normed:
conv = tf.nn.lrn(conv, bias=0.00005)
return conv
def linear(input_, output_size, use_bias=False, bias_start=0.0, spectral_normed=False, batch_normed=False, weight_normed=False, lr_normed=False, layer_normed=False, instance_normed=False, name="linear"):
shape = input_.get_shape().as_list()
with tf.variable_scope(name) as scope:
if scope_has_variables(scope):
scope.reuse_variables()
if not weight_normed:
weight = tf.get_variable("w", [shape[1], output_size], tf.float32, tf.contrib.layers.variance_scaling_initializer())
if spectral_normed:
weight = spectral_norm(weight)
#print(name) #debug use only
if weight_normed:
mul = WeightNorm(tf.keras.layers.Dense(output_size, use_bias=use_bias))(input_)
elif batch_normed:
mul = tf.keras.layers.BatchNormalization(tf.keras.layers.Dense(output_size, use_bias=use_bias))(input_)
else:
mul = tf.matmul(input_, weight)
if use_bias and not weight_normed:
bias = tf.get_variable(name="bias", shape=[output_size],
initializer=tf.constant_initializer(bias_start))
mul += bias
#if layer_normed:
#mul = tf.contrib.layers.layer_norm(mul)
if batch_normed:
mul = tf.layers.batch_normalization(mul)
#if instance_normed:
#mul = instance_norm(mul)
return mul
def conv2D(inputs,
kernel,
output_channel,
stride,
use_bias=False,
name='conv_1',
spectral_normed=True,
weight_normed=False,
batch_normed=False,
depthwise=False,
batch_renormed=False,
layer_normed=False,
instance_normed=False,
lr_normed=False,
is_training=False,
stddev=0.02,
padding="SAME"):
with tf.variable_scope(name) as scope:
if scope_has_variables(scope):
scope.reuse_variables()
if not weight_normed:
w = tf.get_variable(
"w", [kernel, kernel,
inputs.get_shape()[-1], output_channel],
initializer=tf.contrib.layers.variance_scaling_initializer())
if spectral_normed:
w = spectral_norm(w)
if depthwise:
conv = tf.keras.layers.SeparableConv2D(output_channel,
kernel,
use_bias=use_bias,
strides=stride,
padding=padding)(inputs)
if weight_normed:
conv = WeightNorm(
tf.keras.layers.Conv2D(output_channel,
kernel,
use_bias=use_bias,
strides=stride,
padding=padding))(inputs)
else:
conv = tf.nn.conv2d(inputs,
w,
strides=[1, stride, stride, 1],
padding=padding)
if batch_renormed:
clip = {
'rmax': tf.constant(3, dtype=tf.float32),
'rmin': tf.constant(1 / 3, dtype=tf.float32),
'dmax': tf.constant(5, dtype=tf.float32)
}
conv = tf.layers.batch_normalization(conv,
training=is_training,
renorm=True,
renorm_clipping=clip)
#conv = tf.layers.batch_normalization(conv)
if batch_normed:
conv = tf.layers.batch_normalization(conv, training=is_training)
if layer_normed:
conv = tf.contrib.layers.layer_norm(conv)
if instance_normed:
conv = instance_norm(conv)
if lr_normed:
conv = tf.nn.lrn(conv, bias=0.00005)
return conv
