Attention Coefficient Matrix conversion

[tsjshg]

In GATConv layer, if we set return_attn_coef argument True we can get the attention coefficient matrix with each head. But it does not have the shape n_nodes x n_nodes, I convert it to the squared matrix form with the following code. Could you please check whether these are correct or not?

a is adjacency matrix as SparseTensor type
att is the returned attention coefficient matrix
n_nodes is the number of nodes as int

When add_self_loops argument is True

from spektral.layers.ops import  add_self_loops_indices

self_looped_idx = add_self_loops_indices(a.indices)
tf.scatter_nd(self_looped_idx, att[..., 0, 0], (n_nodes, n_nodes))

When add_self_loops argument is False

tf.scatter_nd(a.indices, att[..., 0, 0], (n_nodes, n_nodes))

I've confirmed them with CATConv code, I think this might be correct...

[danielegrattarola]

Yes, this looks correct.
The attention indices of the GAT layer are mapped 1:1 to the edges, and the self-loops are added at the end of the edge list.
Were you able to achieve what you wanted to do?

[tsjshg]

Thank you for your kind reply.
Yes, I've successfully made my first graph attention autoencoder for the drug repositioning research. Spektral is easy to use, so much helpful for me! This is a great library! Thanks.