GraphConvolutionLayer

import tensorflow as tf
class GraphConvolutionLayer(tf.keras.layers.Layer):
  def __init__(self, input_shape, adjacency_matrices, units=16, activation_function=tf.nn.relu, regularizer=None, constraint=None):
    super(GraphConvolutionLayer, self).__init__()
    self.adjacency_matrices = adjacency_matrices
    self.activation_function = activation_function
    self.max_hop = len(adjacency_matrices)
    self.units = units
    self.w = dict()
    for k in range(max(1,self.max_hop)):
      self.w[k] = self.add_weight( 
                             shape=(input_shape[-1], units)
                             , initializer='glorot_normal'
                             , trainable=True
                             , regularizer=regularizer
                             , constraint=constraint 
                             , dtype='float32'
                            )
    self.b = self.add_weight( 
                             shape=(self.units,)
                             , initializer='zeros'
                             , trainable=True
                             , dtype='float32'
                            )
  def call(self, inputs):
    return self.activation_function(self.b + tf.add_n([tf.matmul(tf.matmul(self.adjacency_matrices[k], inputs), self.w[k]) for k in range(max(1, self.max_hop))]))