class GraphAttention

GAT layer skeleton (Veličković et al.).

Efficient on sparse graphs via tf.sparse. For dense graphs, uses masked attention over the adjacency.