I’m currently working on a personal mini project and I wanted to utilise/combine both standard PyTorch as well as PyTorch Geometric, however, there are couple of things that seem to confuse me, primarily with regards to creating DataLoaders.
The project aims to predict a continuous quantity with respect to an agent that is surrounded by multiple other agents. Let’s call the agent of interest parent and other agents within the scene children.
The parent object has various attributes, some of which I want to process/encode with MLP and other with CNN. I also wanted to encode the relationship between parent and children through the use of Graph Neural Net. This would allow me to combine the output of MLP + CNN with the output of GNN (encoding of the nodes and graph) that could be processed further to get final prediction.
I have an issue with understanding how would one create a
torch.utils.data.DataLoader that yields a batch of samples with regards to features of a parent agent as well as
torch_geometric.data.DataLoader that yields the same batch with node features, edge indices etc, of both parent and children.
I went over couple of examples from the main PyTorch Geometric side but couldn’t find anything that would address my issues. Most of these example deal with data that is in pure graph form.
Thank you for any suggestions.