TimDettmers
the number of relation
I have 11 relations in my dataset, but I get nun_relations is 24. I know the relation_inverse, but why it is not 22？ I also have another question, the index in str2var['e1'] and str2var['e2_multi1'] same or not, when I do like this? keys2keys['e1'] = 'e1' # entities keys2keys['rel'] = 'rel' # relations keys2keys['rel_eval'] = 'rel' # relations keys2keys['e2'] = 'e1' # entities keys2keys['e2_multi1'] = 'e1' # entity keys2keys['e2_multi2'] = 'e1' # entity input_keys = ['e1', 'rel', 'rel_eval', 'e2', 'e2_multi1', 'e2_multi2']
Yeah, that can be a bit confusing. It is a bit weird software artifact. It is 11 (normal) + 11 (reverse) + 1 OOV relation + 1 "" relation. The last two relations have no effect on training because they are never used. It is an artifact of the vocab class of the spodernet library.
I found that the number of entities in the calculation is also more than 1OOV entity +1"" entity, and these two parts appear in the first two entity embedding, are they also invalid in the training process? if I want to use conve as decoder, can I directly use self.emb_ e.weight.data [2:] = emb_e.data to replace the entity embedding?
