🚀🫀 Extend scoring, evaluation, and inference for relation side #728
Conversation
cthoyt
changed the title
Co-authored-by: Max Berrendorf <berrendorf@dbs.ifi.lmu.de>
Co-authored-by: Max Berrendorf <berrendorf@dbs.ifi.lmu.de>
tests/cases.py
Outdated
| elif target == LABEL_TAIL: | ||
| sel_col, start_col = 2, 0 | ||
| elif target == LABEL_RELATION: | ||
| raise NotImplementedError(target) | ||
| else: | ||
| raise ValueError(target) | ||
| stop_col = start_col + 2 |
There was a problem hiding this comment.
do we need to adjust this one, too?
There was a problem hiding this comment.
is this one resolved? from the code below it looks like the range start_col:stop_col is not working for all but the first case
There was a problem hiding this comment.
idk. I just added the relation one with what seemed to fit the pattern 🤷
|
@mberr everything here is done - I marked the SampledRankingEvaluator with a |
| # Checking for self.inverse is not necessary since we have score_h_inverse. | ||
| # Since we have inverse triples, where the order of entities is inverted, | ||
| # and the relation replaced by an inverse relation. |
| @@ -523,6 +529,19 @@ def score_t_inverse(self, hr_batch: torch.LongTensor, slice_size: Optional[int] | |||
| r_inv_h = self._prepare_inverse_batch(batch=hr_batch, index_relation=1) | |||
| return self.score_h(rt_batch=r_inv_h, slice_size=slice_size) | |||
|
|
|||
| def score_r_inverse(self, ht_batch: torch.LongTensor, slice_size: Optional[int] = None): | |||
| """Score all rels for a batch of (h,t)-pairs using the rel predictions for the inverses $(t,*_{inv},h)$.""" | |||
There was a problem hiding this comment.
What is described in the docstring does make sense from my point of view: We compute scores for all relations, but this time, using the inverse representation and flipped head & tail.
score[i] = score(t, r_i_inv, h)
implementing seems to depend on fixing what score_r computes, which currently is equal to scores for normal and inverse relations (cf. #752 )
tests/cases.py
Outdated
| elif target == LABEL_TAIL: | ||
| sel_col, start_col = 2, 0 | ||
| elif target == LABEL_RELATION: | ||
| raise NotImplementedError(target) | ||
| else: | ||
| raise ValueError(target) | ||
| stop_col = start_col + 2 |
There was a problem hiding this comment.
is this one resolved? from the code below it looks like the range start_col:stop_col is not working for all but the first case
Most of the PyKEEN pipeline was built around the scoring and evaluation of head/tail corruptions. However, there were a few places where we added in functionality for training where the relations are corrupted. This PR uses the pre-existing templates for handling of heads and tails in the
pykeen.models.Modelandpykeen.evaluation.Evaluatorclasses to enable these things. This PR is missing a few things:pykeen.models.Model.score_r_inversepykeen.models.Model.predict_rbe updated to usepykeen.models.Model.score_r_inversesimilarly to howpykeen.models.Model.predict_husespykeen.models.Model.score_h_inverse?process_relation_scores_function?Dependencies