Accessing aligned pairs of trained model

[calyptis]

I am following the entity alignment tutorial tutorial/entity_alignment/main.py and would like to store all closest alignments for each entity in model.kgs.kg1.entities_set.

Question: How can I access them?

Currently, I am using the following repurposed code (function added to tutorial/entity_alignment/mtranse.py):

def get_matches(self, remove_dangling=False):

     candidate_list_kg2 = self.kgs.kg2.entities_set
     if remove_dangling:
         candidate_list_kg2 = (
             candidate_list_kg2
             - set([x for x, _ in self.kgs.train_unlinked_entities2])
             - set([x for x, _ in self.kgs.valid_unlinked_entities2])
         )

     candidate_list_kg2 = list(candidate_list_kg2)

     embeds1 = tf.nn.embedding_lookup(self.ent_embeds, list(self.kgs.kg1.entities_set)).eval(session=self.session)
     embeds2 = tf.nn.embedding_lookup(self.ent_embeds, candidate_list_kg2).eval(session=self.session)
     mapping = self.mapping_mat.eval(session=self.session) if self.mapping_mat is not None else None

     alignment_rest_12, _, _, sim_list = test(
         embeds1, embeds2, mapping, self.args.top_k, self.args.test_threads_num,
         metric=self.args.eval_metric, normalize=self.args.eval_norm, csls_k=0, accurate=True
     )
     
     return alignment_rest_12

But the problem is that the tuple in alignment_rest_12 cannot be mapped to the entity URIs. The integer values seem not to be referring to entity IDs since

assert all([i[0] in model.kgs.kg1.entities_id_dict.values() for i in alignment_rest_12])

fails.

[calyptis]

Writing my question down helped me find a solution (see below).

The issue is that alignment_rest_12 does not store the entity IDs but rather the indexes of the entity IDs.

Perhaps there's an easier way. Happy to hear it if it exists.

def get_matches(self, remove_dangling: bool = False) -> list:
    candidate_list_kg2 = self.kgs.kg2.entities_set
    if remove_dangling:
        candidate_list_kg2 = (
            candidate_list_kg2
            - set([x for x, _ in self.kgs.train_unlinked_entities2])
            - set([x for x, _ in self.kgs.valid_unlinked_entities2])
        )

    candidate_list_kg1 = sorted(list(self.kgs.kg1.entities_set))
    candidate_list_kg2 = sorted(list(candidate_list_kg2))

    embeds1 = tf.nn.embedding_lookup(self.ent_embeds, candidate_list_kg1).eval(session=self.session)
    embeds2 = tf.nn.embedding_lookup(self.ent_embeds, candidate_list_kg2).eval(session=self.session)
    mapping = self.mapping_mat.eval(session=self.session) if self.mapping_mat is not None else None

    alignment_rest_12, _, _, sim_list = test(
        embeds1, embeds2, mapping, self.args.top_k, self.args.test_threads_num,
        metric=self.args.eval_metric, normalize=self.args.eval_norm, csls_k=0, accurate=True
    )

    candidates_kg1 = np.array(candidate_list_kg1)
    candidates_kg2 = np.array(candidate_list_kg2)
    matches = np.array(list(alignment_rest_12))

    # Map from position (index) to ID
    matches_reindexed = np.dstack((
        candidates_kg1[matches[:, 0]],
        candidates_kg2[matches[:, 1]]
    )).squeeze()

    # Map from ID to name
    mapping_kg1_int_name = {v: k for k, v in self.kgs.kg1.entities_id_dict.items()}
    mapping_kg2_int_name = {v: k for k, v in self.kgs.kg2.entities_id_dict.items()}

    matches_names = np.dstack((
        [mapping_kg1_int_name.get(i) for i in matches_reindexed[:, 0]],
        [mapping_kg2_int_name.get(i) for i in matches_reindexed[:, 1]]
    )).squeeze().tolist()

    return matches_names