Graph Pooling without batch

[dgm2]

I have a batch of graphs features of dimension bs, 18, 18 and its adjacency matrix also bs,18,18.
what is a way to pool this into bs,1 e.g. each tensor pooled in single value ?
Any better way just getting the mean e.g. with some nn layer?
I looked into pyg pooling layers but they require some batch tensor that I don't have.

many thanks

[wsad1]

You have two options.

1. Use Dense Pooling Layers
2. Convert your adjacency matrix to sparse version using dense_to_sparse, and reshape your feature vector 'x.view(bs*18, 18)'. Then use any pooling layer

[dgm2]

thanks, for option 1, how do I get the parameter s (Tensor): Assignment tensor.
The issue I am finding, is there is this other parameter about "clusters" required in pooling?
I only have x and adj

[rusty1s]

The s tensor is usually learned, e.g., through an MLP on the node embeddings.

[dgm2]

thanks for explaining.
what should be the dimension of s in this case? are there minimal examples of pooling ?

[wsad1]

you can learn s using a GNN like done here.