def build_single_net(x, is_training, FLAGS):
n = FLAGS.blocks
# shape = x.get_shape().as_list()
with tf.variable_scope('pre'):
pre = layers.conv(x, num_outputs=16, kernel_size = [3, 3], scope='conv', b_norm=True, is_training=is_training,
weight_decay=FLAGS.weight_decay)
# pre = layers.max_pool2d(pre, [2, 2], padding='SAME', scope='pool')
h = pre
for i in range(1, n + 1):
h = block(h, 16, FLAGS.weight_decay, '16_block{}'.format(i), is_training)
h = block(h, 32, FLAGS.weight_decay, '32_block1', is_training, True)
for i in range(2, n + 1):
h = block(h, 32, FLAGS.weight_decay, '32_block{}'.format(i), is_training)
h = block(h, 64, FLAGS.weight_decay, '64_block1', is_training, True)
for i in range(2, n + 1):
h = block(h, 64, FLAGS.weight_decay, '64_block{}'.format(i), is_training)
shape = h.get_shape().as_list()
h = tf.contrib.layers.avg_pool2d(h, [shape[1], shape[2]], scope='global_pool')
res = layers.conv(h, num_outputs=FLAGS.num_classes, kernel_size=[1, 1], scope='fc1', padding='VALID',
b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay, activation_fn=None)
return h, tf.reshape(res, [-1, FLAGS.num_classes])
def build_net(x, is_training, FLAGS):
with tf.variable_scope('branch1'):
y1 = build_single_net(x, is_training, FLAGS)
y1 = p_relu(y1)
with tf.variable_scope('branch2'):
y2 = build_single_net(x, is_training, FLAGS)
y2 = p_relu(y2)
with tf.variable_scope('branch3'):
y3 = build_single_net(x, is_training, FLAGS)
y3 = p_relu(y3)
with tf.variable_scope('branch4'):
y4 = build_single_net(x, is_training, FLAGS)
con = tf.concat([y1, y2, y3], axis=3) #192
con = tf.reshape(con, [-1, 1, 192])
w = layers.conv(y4,
num_outputs=FLAGS.num_classes * 192,
kernel_size=[1, 1],
scope='fc2',
padding='VALID',
b_norm=True,
is_training=is_training,
weight_decay=FLAGS.weight_decay,
activation_fn=None)
w = tf.reshape(w, [-1, 192, FLAGS.num_classes])
res = tf.matmul(con, w)
res = layers.batch_norm(res, is_training=is_training, scope='bn')
return tf.reshape(res, [-1, FLAGS.num_classes])
def block(inputs, num_outputs, weight_decay, scope, is_training, down_sample = False, trainable=True):
with tf.variable_scope(scope):
num_inputs = inputs.get_shape().as_list()[3]
res = layers.conv(inputs, num_outputs = num_outputs, kernel_size=[3, 3],
strides=[1, 2, 2, 1] if down_sample else [1, 1, 1, 1],
scope='conv1', b_norm=True, is_training=is_training, weight_decay=weight_decay, trainable=trainable)
res = layers.conv(res, num_outputs=num_outputs, kernel_size=[3, 3], activation_fn=None,
scope='conv2', b_norm=True, is_training=is_training, weight_decay=weight_decay, trainable=trainable)
if num_inputs != num_outputs:
inputs = layers.conv(inputs, num_outputs=num_outputs, kernel_size=[1, 1], activation_fn=None,
scope='short_cut', strides=[1, 2, 2,1 ], b_norm=True, is_training=is_training,
weight_decay=weight_decay, trainable=trainable)
res = tf.nn.relu(res + inputs)
return res
def build_net(x, is_training, FLAGS):
with tf.variable_scope('compress70'):
y1 = build_single_net(x, is_training, FLAGS, False)
with tf.variable_scope('compress90'):
y2 = build_single_net(x, is_training, FLAGS, False)
with tf.variable_scope('gamma0_8'):
y3 = build_single_net(x, is_training, FLAGS, False)
with tf.variable_scope('resize0_5'):
y4 = build_single_net(x, is_training, FLAGS, False)
with tf.variable_scope('resize0_8'):
y5 = build_single_net(x, is_training, FLAGS, False)
with tf.variable_scope('resize1_5'):
y6 = build_single_net(x, is_training, FLAGS, False)
with tf.variable_scope('resize2_0'):
y7 = build_single_net(x, is_training, FLAGS, False)
with tf.variable_scope('cluster'):
y8 = build_single_net(x, is_training, FLAGS)
con = tf.concat([y1, y2, y3, y4, y5, y6, y7], axis=3) #192
con = tf.reshape(con, [-1, 1, 448])
w = layers.conv(y8, num_outputs=FLAGS.num_classes*448, kernel_size=[1, 1], scope='fc2', padding='VALID',
b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay, activation_fn=None)
w = tf.reshape(w, [-1, 448, FLAGS.num_classes])
res = tf.matmul(con, w)
res = layers.batch_norm(res, is_training=is_training, scope='bn')
return tf.reshape(res, [-1, FLAGS.num_classes])
def build_net(x, is_training, FLAGS):
confs = []
ys = []
with tf.variable_scope('compress70'):
conf1, y1 = build_single_net(x, is_training, FLAGS)
conf1 = tf.sigmoid(layers.conv(conf1, num_outputs=1, kernel_size=[1, 1], scope='fc2', padding='VALID',
b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay, activation_fn=None))
confs.append(conf1)
ys.append(y1)
with tf.variable_scope('compress90'):
conf2, y2 = build_single_net(x, is_training, FLAGS)
conf2 = tf.sigmoid(layers.conv(conf2, num_outputs=1, kernel_size=[1, 1], scope='fc2', padding='VALID',
b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay,
activation_fn=None))
confs.append(conf2)
ys.append(y2)
with tf.variable_scope('gamma0_8'):
conf3, y3 = build_single_net(x, is_training, FLAGS)
conf3 = tf.sigmoid(layers.conv(conf3, num_outputs=1, kernel_size=[1, 1], scope='fc2', padding='VALID',
b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay,
activation_fn=None))
confs.append(conf3)
ys.append(y3)
with tf.variable_scope('gamma1_2'):
conf4, y4 = build_single_net(x, is_training, FLAGS)
conf4 = tf.sigmoid(layers.conv(conf4, num_outputs=1, kernel_size=[1, 1], scope='fc2', padding='VALID',
b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay,
activation_fn=None))
confs.append(conf4)
ys.append(y4)
with tf.variable_scope('resize0_5'):
conf5, y5 = build_single_net(x, is_training, FLAGS)
conf5 = tf.sigmoid(layers.conv(conf5, num_outputs=1, kernel_size=[1, 1], scope='fc2', padding='VALID',
b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay,
activation_fn=None))
confs.append(conf5)
ys.append(y5)
with tf.variable_scope('resize0_8'):
conf6, y6 = build_single_net(x, is_training, FLAGS)
conf6 = tf.sigmoid(layers.conv(conf6, num_outputs=1, kernel_size=[1, 1], scope='fc2', padding='VALID',
b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay,
activation_fn=None))
confs.append(conf6)
ys.append(y6)
with tf.variable_scope('resize1_5'):
conf7, y7 = build_single_net(x, is_training, FLAGS)
conf7 = tf.sigmoid(layers.conv(conf7, num_outputs=1, kernel_size=[1, 1], scope='fc2', padding='VALID',
b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay,
activation_fn=None))
confs.append(conf7)
ys.append(y7)
with tf.variable_scope('resize2_0'):
conf8, y8 = build_single_net(x, is_training, FLAGS)
conf8 = tf.sigmoid(layers.conv(conf8, num_outputs=1, kernel_size=[1, 1], scope='fc2', padding='VALID',
b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay,
activation_fn=None))
confs.append(conf8)
ys.append(y8)
return confs, ys