How to efficiently batch multiple instances of the same graph (only difference being node embeddings)

[jasperhyp]

Hi! I have one graph and a number of different sets of node embeddings for this graph. Using a larger adjacency matrix does not sound efficient in this case. I think the simplest way is to feed in 3D node embedding input ([batch_size, node_num, feat_dim]), so I experimented and it looks like GATv2Conv does not allow for 3D node embeddings. How should I deal with this?

[rusty1s]

Yes, not all GNN operators can operate on 3D tensors. It is currently recommended to follow the standard PyG batching in other cases. This will solely replicate the edge_index for each example in the mini-batch once, which is negligible in terms of memory consumption.

[jasperhyp]

Ok cool, thank you!