图注意力网络(GAT) TensorFlow实现

论文

图注意力网络来自 Graph Attention Networks，ICLR 2018. https://arxiv.org/abs/1710.10903

GAT层

输入

N为节点的个数，F为feature的个数，这表示输入为N个节点的每个节点的F个feature

输出

表示对这N个节点的 F’ 个输出，输出位N个节点的每个节点的F’个feature

注意力机制

GAT.py

import tensorflow as tf
from tensorflow import keras
from tensorflow.python.keras import activations
from tensorflow.python.keras import constraints
from tensorflow.python.keras import initializers
from tensorflow.python.keras import regularizers
 
 
class GraphAttentionLayer(keras.layers.Layer):
    def compute_output_signature(self, input_signature):
        pass
 
    def __init__(self,
                 input_dim,
                 output_dim,
                 adj,
                 nodes_num,
                 dropout_rate=0.0,
                 activation=None,
                 use_bias=True,
                 kernel_initializer='glorot_uniform',
                 bias_initializer='zeros',
                 kernel_regularizer=None,
                 bias_regularizer=None,
                 activity_regularizer=None,
                 kernel_constraint=None,
                 bias_constraint=None,
                 coef_dropout=0.0,
                 **kwargs):
        """
        :param input_dim: 输入的维度
        :param output_dim: 输出的维度，不等于input_dim
        :param adj: 具有自环的tuple类型的邻接表[coords, values, shape]， 可以采用sp.coo_matrix生成
        :param nodes_num: 点数量
        :param dropout_rate: 丢弃率，防过拟合，默认0.5
        :param activation: 激活函数
        :param use_bias: 偏移，默认True
        :param kernel_initializer: 权值初始化方法
        :param bias_initializer: 偏移初始化方法
        :param kernel_regularizer: 权值正则化
        :param bias_regularizer: 偏移正则化
        :param activity_regularizer: 输出正则化
        :param kernel_constraint: 权值约束
        :param bias_constraint: 偏移约束
        :param coef_dropout: 互相关系数丢弃，默认0.0
        :param kwargs:
        """
        super(GraphAttentionLayer, self).__init__()
        self.activation = activations.get(activation)
        self.use_bias = use_bias
        self.kernel_initializer = initializers.get(kernel_initializer)
        self.bias_initializer = initializers.get(bias_initializer)
        self.kernel_regularizer = regularizers.get(kernel_regularizer)
        self.bias_regularizer = regularizers.get(bias_regularizer)
        self.kernel_constraint = constraints.get(kernel_constraint)
        self.bias_constraint = constraints.get(bias_constraint)
        self.input_dim = input_dim
        self.output_dim = output_dim
        self.support = [tf.SparseTensor(indices=adj[0][0], values=adj[0][1], dense_shape=adj[0][2])]
        self.dropout_rate = dropout_rate
        self.coef_drop = coef_dropout
        self.nodes_num = nodes_num
        self.kernel = None
        self.mapping = None
        self.bias = None
 
    def build(self, input_shape):
        """
        只执行一次
        """
        self.kernel = self.add_weight(shape=(self.input_dim, self.output_dim),
                                      initializer=self.kernel_initializer,
                                      regularizer=self.kernel_regularizer,
                                      constraint=self.kernel_constraint,
                                      trainable=True)
 
        if self.use_bias:
            self.bias = self.add_weight(shape=(self.nodes_num, self.output_dim),
                                        initializer=self.kernel_initializer,
                                        regularizer=self.kernel_regularizer,
                                        constraint=self.kernel_constraint,
                                        trainable=True)
        print('[GAT LAYER]: GAT W & b built.')
 
    def call(self, inputs, training=True):
        # 完成输入到输出的映射关系
        # inputs = tf.nn.l2_normalize(inputs, 1)
        raw_shape = inputs.shape
        inputs = tf.reshape(inputs, shape=(1, raw_shape[0], raw_shape[1]))  # (1, nodes_num, input_dim)
        mapped_inputs = keras.layers.Conv1D(self.output_dim, 1, use_bias=False)(inputs)  # (1, nodes_num, output_dim)
        # mapped_inputs = tf.nn.l2_normalize(mapped_inputs)
 
        sa_1 = keras.layers.Conv1D(1, 1)(mapped_inputs)  # (1, nodes_num, 1)
        sa_2 = keras.layers.Conv1D(1, 1)(mapped_inputs)  # (1, nodes_num, 1)
 
        con_sa_1 = tf.reshape(sa_1, shape=(raw_shape[0], 1))  # (nodes_num, 1)
        con_sa_2 = tf.reshape(sa_2, shape=(raw_shape[0], 1))  # (nodes_num, 1)
 
        con_sa_1 = tf.cast(self.support[0], dtype=tf.float32) * con_sa_1  # (nodes_num, nodes_num) W_hi
        con_sa_2 = tf.cast(self.support[0], dtype=tf.float32) * tf.transpose(con_sa_2, [1, 0])  # (nodes_num, nodes_num) W_hj
 
        weights = tf.sparse.add(con_sa_1, con_sa_2)  # concatenation
        weights_act = tf.SparseTensor(indices=weights.indices,
                                      values=tf.nn.leaky_relu(weights.values),
                                      dense_shape=weights.dense_shape)  # 注意力互相关系数
        attention = tf.sparse.softmax(weights_act)  # 输出注意力机制
        inputs = tf.reshape(inputs, shape=raw_shape)
        if self.coef_drop > 0.0:
            attention = tf.SparseTensor(indices=attention.indices,
                                        values=tf.nn.dropout(attention.values, self.coef_dropout),
                                        dense_shape=attention.dense_shape)
        if training and self.dropout_rate > 0.0:
            inputs = tf.nn.dropout(inputs, self.dropout_rate)
        if not training:
            print("[GAT LAYER]: GAT not training now.")
 
        attention = tf.sparse.reshape(attention, shape=[self.nodes_num, self.nodes_num])
        value = tf.matmul(inputs, self.kernel)
        value = tf.sparse.sparse_dense_matmul(attention, value)
 
        if self.use_bias:
            ret = tf.add(value, self.bias)
        else:
            ret = tf.reshape(value, (raw_shape[0], self.output_dim))
        return self.activation(ret)

参考

https://blog.csdn.net/weixin_36474809/article/details/89401552

https://github.com/PetarV-/GAT