def conv_bn_rectify(net, num_filters, wd, name, is_training, reuse):
with tf.variable_scope(name):
net = layers.conv_2d_layer(net, [3,3], net.get_shape()[3], num_filters, nonlinearity=None, wd=wd,
padding='SAME', name='conv', with_biases=False)
biases = layers._variable_on_cpu('biases', net.get_shape()[3], tf.constant_initializer(0.0), dtype=tf.float32)
net = tf.nn.bias_add(net, biases)
net = tf.contrib.layers.batch_norm(net, scope=tf.get_variable_scope(), decay=0.9, reuse=reuse, is_training=is_training)
net = tf.nn.relu(net)
return net
def net_vgglike(images, nclass, scale, is_training, reuse, threshold):
net = conv_bn_rectify(images, int(64 * scale), 0.0, 'conv_1', is_training,
reuse, threshold)
net = conv_bn_rectify(net, int(64 * scale), 0.0, 'conv_2', is_training,
reuse, threshold)
net = tf.nn.max_pool(net,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID')
net = conv_bn_rectify(net, int(128 * scale), 0.0, 'conv_3', is_training,
reuse, threshold)
net = conv_bn_rectify(net, int(128 * scale), 0.0, 'conv_4', is_training,
reuse, threshold)
net = tf.nn.max_pool(net,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID')
net = conv_bn_rectify(net, int(256 * scale), 0.0, 'conv_5', is_training,
reuse, threshold)
net = conv_bn_rectify(net, int(256 * scale), 0.0, 'conv_6', is_training,
reuse, threshold)
net = conv_bn_rectify(net, int(256 * scale), 0.0, 'conv_7', is_training,
reuse, threshold)
net = tf.nn.max_pool(net,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID')
net = conv_bn_rectify(net, int(512 * scale), 0.0, 'conv_8', is_training,
reuse, threshold)
net = conv_bn_rectify(net, int(512 * scale), 0.0, 'conv_9', is_training,
reuse, threshold)
net = conv_bn_rectify(net, int(512 * scale), 0.0, 'conv_10', is_training,
reuse, threshold)
net = tf.nn.max_pool(net,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID')
net = conv_bn_rectify(net, int(512 * scale), 0.0, 'conv_11', is_training,
reuse, threshold)
net = conv_bn_rectify(net, int(512 * scale), 0.0, 'conv_12', is_training,
reuse, threshold)
net = conv_bn_rectify(net, int(512 * scale), 0.0, 'conv_13', is_training,
reuse, threshold)
net = tf.nn.max_pool(net,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID')
net = tf.reshape(net, [
-1,
(net.get_shape()[1] * net.get_shape()[2] * net.get_shape()[3]).value
])
net = layers.sbp_dropout(net, 1, is_training, 'sbp_dense_1', reuse,
threshold)
net = layers.dense_layer(net,
net.get_shape()[1],
int(512 * scale),
nonlinearity=None,
wd=0.0,
name='dense_1',
with_biases=False)
biases = layers._variable_on_cpu('biases_dense_1',
net.get_shape()[1],
tf.constant_initializer(0.0),
dtype=tf.float32)
net = tf.nn.bias_add(net, biases)
net = tf.contrib.layers.batch_norm(net,
scope=tf.get_variable_scope(),
reuse=reuse,
is_training=False,
center=True,
scale=True)
net = tf.nn.relu(net)
net = layers.sbp_dropout(net, 1, is_training, 'sbp_dense_2', reuse,
threshold)
net = layers.dense_layer(net,
net.get_shape()[1],
nclass,
nonlinearity=None,
wd=0.0,
name='dense_2',
with_biases=False)
biases = layers._variable_on_cpu('biases_dense_2',
net.get_shape()[1],
tf.constant_initializer(0.0),
dtype=tf.float32)
net = tf.nn.bias_add(net, biases)
return net