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).

[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.