avg_pool_neighbor_x now returns only the pooled neighbors' feature vectors

[ldv1]

This is a small PR.

avg_pool_neighbor_x takes in data and modifies data.x; sometimes we only want to average pool the neighbors to get information as a patch, without overwriting data.x; so, now we can write either
data.x = avg_pool_neighbor_x(data.x, data.edge_index, data.num_nodes)
or
patch_x = avg_pool_neighbor_x(data.x, data.edge_index, data.num_nodes)

[rusty1s]

Thanks for the PR. I'm not a big fan of avg_pool_neighbor_x to be honest, and will probably remove it at some point in time.

The proposed implementation is already quite similar to a normal scatter call:

path_x = scatter(data.x, data.edge_index[0], dim_size=data.num_nodes, reduce='mean')

It's hard for me to think of a use-case where this might be useful.

[ldv1]

You can think of it as a "local average pooling" as opposed to "global average pooling". I concatenate this information to the feature vector. It basically characterizes how much a node stands out among its neighbors (the information of a second-order derivative as a matter of fact). That's useful.

[rusty1s]

Yes, I definitely agree that this is useful. It's just that this functionality is already equivalent to a normal scatter call.

[rusty1s]

On a side node. What about we integrate that functionality in a non-trainable MessagePassing module?