Pooling layers vs aggregation operators

[vakker]

I'm a bit confused about the difference between the pooling layers and the aggregation operators.

To me they seem to provide the same functionality, which is aggregating the features from a set of nodes (which can be the whole graph in case of global pooling/aggregation), e.g. aggr.MeanAggregation does the same as pool.global_mean_pool.

Is there any particular theoretical difference, or pooling and aggregation are essentially the same concepts?
If the latter, then would it make sense to unify them?

Thanks!

[rusty1s]

global_mean_pool is just the functional variant of MeanAggregation (and mostly just exist for backward compatibility reasons by now).

[agunter7]

I'm confused on what exactly determines whether something is organized as an Aggregation Operator vs Pooling Layer. Not everything in the Pooling Layers section is purely functional, e.g. EdgePooling.

Is it just a historical distinction or am I misunderstanding something?

[vakker]

Thanks for the reply.
So, in general pooling and aggregation layers seem conceptually the same to me. Is there a fundamental difference?

[rusty1s]

Yes, they are basically the same (except for the module vs function difference), similar to nn.Dropout and torch.nn.functional.dropout.