def call(self, node_embeddings, adjacency_lists, training=True):
node_embeddings = self.dropout1(node_embeddings, training=training)
x = self.gc1(GNNInput(node_embeddings, adjacency_lists))
x = tf.nn.relu(x)
x = self.dropout2(x, training=training)
x = self.gc2(GNNInput(x, adjacency_lists))
return tf.math.softmax(x, -1)
def decoder(self, hidden_embeddings, adjacency_lists, training):
for layer in range(self.num_layers):
hidden_embeddings = self.decoder_layers[layer](GNNInput(hidden_embeddings, adjacency_lists),
self.weight[-(layer+1)],
True,
training)
return hidden_embeddings
def call(self, inputs):
adjacency_lists = inputs.adjacency_lists
node_embeddings = inputs.node_embeddings
if self.use_bias:
node_embeddings = tf.linalg.matmul(node_embeddings,
self.weight) + self.bias
else:
node_embeddings = tf.linalg.matmul(node_embeddings, self.weight)
aggr_out = self.propagate(GNNInput(node_embeddings, adjacency_lists))
return aggr_out
def call(self, inputs, training):
adjacency_lists = inputs.adjacency_lists
node_embeddings = inputs.node_embeddings
node_embeddings = tf.linalg.matmul(node_embeddings, self.weight)
aggr_out = self.propagate(GNNInput(node_embeddings, adjacency_lists), training)
if self.concat is True:
aggr_out = tf.reshape(aggr_out, [-1, self.heads * self.out_features])
else:
aggr_out = tf.reduce_mean(aggr_out, 1)
if self.use_bias:
aggr_out += self.bias
return aggr_out
def encoder(self, node_embeddings, adjacency_lists, training):
for layer in range(self.num_layers):
node_embeddings = self.encoder_layers[layer](GNNInput(node_embeddings, adjacency_lists), self.weight[layer],
False, training)
return node_embeddings