def channel_attention(x, outdim, kernel_size, strides, name):
with tf.variable_scope("CA-%s" % name):
skip_conn = x
# skip_conn = tf.identity(x, name='identity')
x = tfu.adaptive_global_average_pool_2d(x)
x = tf.layers.conv2d(x,
outdim // 16,
kernel_size=kernel_size,
strides=strides,
padding='SAME',
name="conv_1",
activation=None)
x = tf.nn.leaky_relu(x, alpha=0.2)
x = tf.layers.conv2d(x,
outdim,
kernel_size=kernel_size,
strides=strides,
padding='SAME',
name="conv_2",
activation=None)
x = tf.nn.sigmoid(x)
x = tf.multiply(skip_conn, x)
return x
def channel_attention(self, x, f, reduction, name):
"""
Channel Attention (CA) Layer
:param x: input layer
:param f: conv2d filter size
:param reduction: conv2d filter reduction rate
:param name: scope name
:return: output layer
"""
with tf.variable_scope("CA-%s" % name):
skip_conn = tf.identity(x, name='identity')
x = tfu.adaptive_global_average_pool_2d(x)
x = tfu.conv2d(x, f=f // reduction, k=1, name="conv2d-1")
x = self.act(x)
x = tfu.conv2d(x, f=f, k=1, name="conv2d-2")
x = tf.nn.sigmoid(x)
return tf.multiply(skip_conn, x)
self.opt = tf.train.AdamOptimizer(learning_rate=self.lr,
beta1=self.beta1, beta2=self.beta2, epsilon=self.opt_eps)
elif self.optimizer == 'sgd': # sgd + m with nesterov
elif self.optimizer == 'sgd': # sgd + m with nestrov
self.opt = tf.train.MomentumOptimizer(learning_rate=self.lr, momentum=self.momentum, use_nesterov=True)
else:
raise NotImplementedError("[-] Not supported optimizer (%s)" % self.optimizer)
raise NotImplementedError("[-] Not supported optimizer {}".format(self.optimizer))
def image_processing(self, x, sign, name):
with tf.variable_scope(name):
@@ -143,24 +144,24 @@ def channel_attention(self, x, f, reduction, name):
with tf.variable_scope("CA-%s" % name):
skip_conn = tf.identity(x, name='identity')
x = tfutil.adaptive_global_average_pool_2d(x)
x = tfu.adaptive_global_average_pool_2d(x)
x = tfutil.conv2d(x, f=f // reduction, k=1, name="conv2d-1")
x = tfu.conv2d(x, f=f // reduction, k=1, name="conv2d-1")
x = self.act(x)
x = tfutil.conv2d(x, f=f, k=1, name="conv2d-2")
x = tfu.conv2d(x, f=f, k=1, name="conv2d-2")
x = tf.nn.sigmoid(x)
return tf.multiply(skip_conn, x)
def residual_channel_attention_block(self, x, f, kernel_size, reduction, use_bn, name):
with tf.variable_scope("RCAB-%s" % name):
skip_conn = tf.identity(x, name='identity')