I am trying to write Graph NN algorithms without using pytorch-geometric. I will have no sparse learnable parameters, however I will have a lot of embeddings and MLPs.
When documentation states ‘doesn’t support computing derivaties’ (torch.sparse.mm) or 'may not have autograd support" ( torch.nn.functional.linear), do you mean I cannot use sparse tensors at all, or only cannot learn sparse tensors?
What would you recommend to use in my case? My adjacency matrices are quite large, they cannot be dence.