🌯 🌪️ Encapsulate inverse relations #752
Conversation
|
What's the roadmap for this PR? I think that it's fine to keep |
| if dataset_instance.create_inverse_triples and not create_inverse_triples: | ||
| assert dataset_instance.training.num_relations % 2 == 0 | ||
| dataset_instance.training.num_relations //= 2 |
There was a problem hiding this comment.
We might need to keep that, if we want to be able to load from old cache files.
| @@ -112,18 +115,18 @@ def __init__( | |||
| else: | |||
| self.loss = loss_resolver.make(loss, pos_kwargs=loss_kwargs) | |||
|
|
|||
| self.use_inverse_triples = triples_factory.create_inverse_triples | |||
| self.use_inverse_triples = create_inverse_triples | |||
There was a problem hiding this comment.
Maybe we should agree on some consistent naming
There was a problem hiding this comment.
let's do that in a follow-up PR, since the diff is already so big
tests/cases.py
Outdated
| @@ -1083,6 +1084,7 @@ def test_score_r(self) -> None: | |||
| self._test_score( | |||
| score=self.instance.score_r, | |||
| columns=[0, 2], | |||
| # switch to effective_num_relations? | |||
There was a problem hiding this comment.
TODO - either switch to self.instance.effective_num_relations or update the tests to new have parameters
There was a problem hiding this comment.
in 3104af9 I witched it to self.instance.effective_num_relations and this made the tests pass again, except inductive GNN nodepiece
|
|
||
| # FIXME do these need to get passed the relation inverter? or how should | ||
| # this math get updated | ||
| inverse_edge_index = edge_index.flip(0) |
There was a problem hiding this comment.
@mberr I also tracked down some issues in GNN - do you have an idea how to update the CompGCN layer to work better with the new way of doing inverses?
Fixes #749
Tasks:
inverse_triplesflags fromTriplesFactoryandDatasets.Dependencies
New Structure
score_hrt(h, r, t)score_h(r, t, slice_size)score_r(h, t, slice_size)score_t(h, r, slice_size)score_hrt_extended(h, r, t, invert_relation)score_h_extended(r, t, slice_size, invert_relation)score_r_extended(h, t, slice_size, invert_relation)score_t_extended(h, r, slice_size, invert_relation)score_h):predict_hrt(h, r, t, invert_relation)predict_h(r, t, slice_size, invert_relation)predict_r(h, t, slice_size, invert_relation)predict_t(h, r, slice_size, invert_relation)score_{hrt,h,t}_inverse-> move to prediction workflows?