def apool(self,
k_height,
k_width,
d_height=2,
d_width=2,
mode='VALID',
input_layer=None,
num_channels_in=None):
"""Construct an average pooling layer."""
if input_layer is None:
input_layer = self.top_layer
else:
self.top_size = num_channels_in
name = 'apool' + str(self.counts['apool'])
self.counts['apool'] += 1
pool = pooling_layers.average_pooling2d(
input_layer, [k_height, k_width], [d_height, d_width],
padding=mode,
data_format=self.channel_pos,
name=name)
self.top_layer = pool
return pool
def NonLocalBlock(input_x,
out_channels,
sub_sample=1,
nltype=0,
is_bn=False,
scope='NonLocalBlock'):
'''https://github.com/nnUyi/Non-Local_Nets-Tensorflow
'''
batchsize, height, width, in_channels = input_x.get_shape().as_list()
typedict = {
0: 'embedded_gaussian',
1: 'gaussian',
2: 'dot_product',
3: 'concat'
}
with tf.variable_scope(scope) as sc:
if nltype <= 2:
with tf.variable_scope('g') as scope:
g = conv2d(input_x,
out_channels,
1,
strides=1,
padding='same',
name='g')
if sub_sample > 1:
g = average_pooling2d(g,
pool_size=sub_sample,
strides=sub_sample,
name='g_pool')
with tf.variable_scope('phi') as scope:
if nltype == 0 or nltype == 2:
phi = conv2d(input_x,
out_channels,
1,
strides=1,
padding='same',
name='phi')
elif nltype == 1:
phi = input_x
if sub_sample > 1:
phi = average_pooling2d(phi,
pool_size=sub_sample,
strides=sub_sample,
name='phi_pool')
with tf.variable_scope('theta') as scope:
if nltype == 0 or nltype == 2:
theta = conv2d(input_x,
out_channels,
1,
strides=1,
padding='same',
name='theta')
elif nltype == 1:
theta = input_x
g_x = tf.reshape(g, [batchsize, -1, out_channels])
theta_x = tf.reshape(theta, [batchsize, -1, out_channels])
# theta_x = tf.reshape(theta, [batchsize, out_channels, -1])
# theta_x = tf.transpose(theta_x, [0,2,1])
phi_x = tf.reshape(phi, [batchsize, -1, out_channels])
phi_x = tf.transpose(phi_x, [0, 2, 1])
#phi_x = tf.reshape(phi_x, [batchsize, out_channels, -1])
f = tf.matmul(theta_x, phi_x)
# ???
if nltype <= 1:
# f_softmax = tf.nn.softmax(f, -1)
f = tf.exp(f)
f_softmax = f / tf.reduce_sum(f, axis=-1, keepdims=True)
elif nltype == 2:
f = tf.nn.relu(f) #/int(f.shape[-1])
f_mean = tf.reduce_sum(f, axis=[2], keepdims=True)
#print(f.shape,f_mean.shape)
f_softmax = f / f_mean
y = tf.matmul(f_softmax, g_x)
y = tf.reshape(y, [batchsize, height, width, out_channels])
with tf.variable_scope('w') as scope:
w_y = conv2d(y,
in_channels,
1,
strides=1,
padding='same',
name='w')
# if is_bn:
# w_y = slim.batch_norm(w_y)
z = w_y #input_x + w_y
return z
def _gap(self, bottom):
height = bottom.get_shape()[1]
width = bottom.get_shape()[2]
return average_pooling2d(bottom, [height,width], [height,width])
def _avgpool(self, bottom, stride):
return average_pooling2d(bottom, [stride,stride], [stride,stride])