def create_layer(self, input, include_w_input=None, is_training=True, center=False, dp_rate=0.0, **kwargs):
print("name: {}".format(self.name))
if self.add_to_input:
input = tf.add(input, include_w_input)
if self.concat_to_input:
input = tf.concat([input, include_w_input],axis=-1)
self.input_shape = get_incoming_shape(input)
print(self.input_shape)
number_of_input_channels = self.input_shape[3]
self.number_of_input_channels = number_of_input_channels
with tf.variable_scope('conv', reuse=False):
initializer = None
if self.weight_init == 'He':
initializer = tf.contrib.layers.variance_scaling_initializer(factor=2.0, mode='FAN_IN', uniform=False)
elif self.weight_init == 'Xnormal':
initializer = tf.contrib.layers.xavier_initializer(uniform=False, seed=None)
W = tf.get_variable(('W{}'.format(self.name)), shape=(self.kernel_size, self.kernel_size,
number_of_input_channels, self.output_channels),
initializer=initializer)
b = tf.Variable(tf.zeros([self.output_channels]))
tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES, W)
output = tf.nn.atrous_conv2d(input, W, rate=self.dilation, padding='SAME')
output = self.apply_dropout(output, dp_rate, is_training)
# apply batch-norm
if self.batch_norm:
print("apply batch norm")
output = tf.contrib.layers.batch_norm(output, is_training=is_training)
output = tf.add(tf.contrib.layers.batch_norm(output), b)
output = self.get_act_values(output)
output = self.zero_center_output(output, center)
return output
def create_layer(self, input, **kwargs):
print("name: {}".format(self.name))
self.input_shape = get_incoming_shape(input)
print(self.input_shape)
number_of_input_channels = self.input_shape[3]
self.number_of_input_channels = number_of_input_channels
with tf.variable_scope('dfconv', reuse=False):
if self.weight_init == 'He':
self.weight_init = tf.contrib.layers.variance_scaling_initializer(factor=2.0, mode='FAN_IN', uniform=False)
elif self.weight_init == 'Xnormal':
self.weight_init = tf.contrib.layers.xavier_initializer(uniform=False, seed=None)
#tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES, W)
output = tl.layers.DeformableConv2d(net, self.prev_layer, self.offset_layer, self.n_filter, self.filter_size, self.act_fn, self.name, self.weight_init)
output = self.get_act_values(output)
return output
def create_layer(self, input):
self.input_shape = get_incoming_shape(input)
print(self.input_shape)
number_of_input_channels = self.input_shape[3]
self.number_of_input_channels = number_of_input_channels
with tf.variable_scope('conv', reuse=False):
W = tf.get_variable(('W{}'.format(self.name[-3:])),shape=(self.kernel_size, self.kernel_size,
number_of_input_channels,
self.output_channels))
b = tf.Variable(tf.zeros([self.output_channels]))
#self.encoder_matrix = W
Conv2d.layer_index += 1
output = tf.nn.atrous_conv2d(input, W, rate=self.dilation, padding='SAME')
output = lrelu(tf.add(tf.contrib.layers.batch_norm(output), b))
return output
def create_layer(self,
input,
include_w_input=None,
pooling_method="MAX",
unpooling_method="nearest_neighbor",
center=False,
**kwargs):
print("name: {}".format(self.name))
self.input_shape = get_incoming_shape(input)
print(self.input_shape)
if self.add_to_input:
input = tf.add(input, include_w_input)
if self.concat_to_input:
input = tf.concat([input, include_w_input], axis=-1)
output = self.apply_pool(input,
self.kernel_size,
pooling_method=pooling_method,
unpooling_method=unpooling_method)
output = self.zero_center_output(output, center)
return output
def create_layer(self, input):
self.input_shape = get_incoming_shape(input)
print(self.input_shape)
return max_pool_2d(input, self.kernel_size)