Max Berrendorf

This might be related: https://github.com/pytorch/pytorch/issues/105839

In theory, tests run now on mps, too; however running the example from https://github.com/mberr/torch-max-mem/issues/14#issuecomment-1732274087 does not succeed; the error seems to be related to https://discuss.pytorch.org/t/mps-back-end-out-of-memory-on-github-action/189773

After merging all modules, it makes sense to do some clean-up of utils (`pykeen.utils` vs. `pykeen.nn.utils` vs. `pykeen.nn.functional` vs. `pykeen.nn.compute_kernel`), and maybe also to split `pykeen.nn.modules` into smaller files.

Hi @febsonthomas , (I reduced you code example a bit by removing some imports you did not use in the snippet). Your code snippet looks mostly correct, but you need...

> Could you also confirm if using relation_to_id alone would suffice, or should I include entity_to_id as well? After training, the new dataset will contain both the previously trained triples...

It probably makes sense to decompose this MR into smaller ones: 1. https://github.com/pykeen/pykeen/pull/1573 2. https://github.com/pykeen/pykeen/pull/1576 3. https://github.com/pykeen/pykeen/pull/1574 4. https://github.com/pykeen/pykeen/pull/1575

> Should we make a protocol for each of these that we can more thoroughly document than the type hints we have so far? Yes, sounds like a good idea!

Is the `attribute` a categorical feature, i.e., has only a fixed set of possible values? If yes, you could convert them to individual relations, e.g., ```json { "source": "IfcWorkControl", "target":...

Your dataset appears quite sparse – with only about three examples per relationship type, it's unlikely that you'll be able to learn good representations. Given that your relation labels are...

In this particular dataset, `InductiveFB15k237`, the validation and test graphs are guaranteed to be disjoint - so we don't need to filter for them; however, it would not hurt much...