This is how FPN is built

def build_pyramid(net_name, end_points, bilinear=True):
  """build pyramid features from a typical network,
  assume each stage is 2 time larger than its top feature
  Returns:
    returns several endpoints
  """
  pyramid = {}
  if isinstance(net_name, str):
    pyramid_map = _networks_map[net_name]
  else:
    pyramid_map = net_name
  # pyramid['inputs'] = end_points['inputs']
  #arg_scope = _extra_conv_arg_scope()
  arg_scope = _extra_conv_arg_scope_with_bn()
  with tf.variable_scope('pyramid'):
    with slim.arg_scope(arg_scope):
      
      pyramid['P5'] = \
        slim.conv2d(end_points[pyramid_map['C5']], 256, [1, 1], stride=1, scope='C5')
      
      for c in range(4, 1, -1):
        s, s_ = pyramid['P%d'%(c+1)], end_points[pyramid_map['C%d' % (c)]]

        # s_ = slim.conv2d(s_, 256, [3, 3], stride=1, scope='C%d'%c)
        
        up_shape = tf.shape(s_)
        # out_shape = tf.stack((up_shape[1], up_shape[2]))
        # s = slim.conv2d(s, 256, [3, 3], stride=1, scope='C%d'%c)
        s = tf.image.resize_bilinear(s, [up_shape[1], up_shape[2]], name='C%d/upscale'%c)
        s_ = slim.conv2d(s_, 256, [1,1], stride=1, scope='C%d'%c)
        
        s = tf.add(s, s_, name='C%d/addition'%c)
        s = slim.conv2d(s, 256, [3,3], stride=1, scope='C%d/fusion'%c)
        
        pyramid['P%d'%(c)] = s
      
      return pyramid