🩹 👾 Fix ERMLP functional form

[mberr]

This PR fixes an issue with the ER-MLP functional form.

I encountered the issue trying to run inference on a single element batch

result = pipeline(
    interaction="ermlp",
    interaction_kwargs=dict(embedding_dim=32, hidden_dim=32),
    dimensions=dict(d=32),
    dataset="nations",
    loss="bcewithlogits",
    training_kwargs=dict(num_epochs=1),
)
result.model.predict_t(result.training.mapped_triples[:1, 1:])

[cthoyt]

Can you encode the failing functionality in a test?

[mberr]

Can you encode the failing functionality in a test?

The error in the implementation was that SizeInformation extracted the wrong size information.

In nn.functional we always use the "more canonical" form (b, nh, nr, nt, *), whereas the score_* methods of modules only use the canonical form (b, n*, *).

For new-style interaction functions, we only test the score_* methods of the interaction module. The ERModel does not use them but always directly goes to the generic score one (without suffix).

I can try to come up with a test, but right now I just stumbled across this error and pushed my local fix here without having time for a deeper investigation of the root cause.

[mberr]

A smaller snippet for reproducing the error without any training:

from pykeen.models.resolve import make_model_cls
from pykeen.triples.generation import generate_triples_factory

triples_factory = generate_triples_factory()
model_cls = make_model_cls(
    interaction="ermlp", 
    dimensions=dict(d=3), 
    interaction_kwargs=dict(
        embedding_dim=3, 
        hidden_dim=5,
    ),
)
model = model_cls(triples_factory=triples_factory)
model.predict_t(triples_factory.mapped_triples[:1, :-1])

[mberr]

An even smaller one

import torch
from pykeen.nn.modules import ERMLPInteraction

num = 5
dim = 3
batch_size = 1
interaction = ERMLPInteraction(embedding_dim=dim, hidden_dim=1)
interaction.score_t(h=torch.rand(batch_size, dim), r=torch.rand(batch_size, dim), all_entities=torch.rand(num, dim))

[mberr]

This is strange, since we have

pykeen/tests/test_nn/test_modules.py

Line 106 in 9b7adc6

class ERMLPTests(cases.InteractionTestCase):

and

pykeen/tests/cases.py

Lines 445 to 453 in 9b7adc6

	def test_score_t(self):
	"""Test score_t."""
	h, r, t = self._get_hrt(
	(self.batch_size,),
	(self.batch_size,),
	(self.num_entities,),
	)
	scores = self.instance.score_t(h=h, r=r, all_entities=t)
	self._check_scores(scores=scores, exp_shape=(self.batch_size, self.num_entities))

which essentially should do the same.

[mberr]

batch_size=1

seems to be what causes the problem.

[mberr]

@cthoyt 53ed846 this catches the old problem. Maybe we want this for the other score_* functions, too?

[cthoyt]

@mberr yes let’s implement this check for all models/score_*

[mberr]

@cthoyt there seems to be a new version of mypy, since there are now some mypy errors in non-modified code parts, cf. https://github.com/pykeen/pykeen/pull/444/checks?check_run_id=2793677666#step:9:12

[cthoyt]

Fine for me. Squash+merge when ready

[cthoyt]

🦜 💃 ⏩