Edward-Sun
commented
4 years ago Hi, Peixiang,
For RotatE, you can first transform relation embedding into phases like:
where self.embedding_range.item() is model-dependent. Then you can calculate the score by either head-batch mode or tail-batch mode, they are equivalent for a single triplet.
For example, in head-batch mode, you can write something like:
re_head, im_head = torch.chunk(head, 2, dim=2) re_tail, im_tail = torch.chunk(tail, 2, dim=2) re_relation = torch.cos(phase_relation) im_relation = torch.sin(phase_relation) re_score = re_relation re_tail + im_relation im_tail - re_head im_score = re_relation im_tail - im_relation re_tail - im_head score = torch.stack([re_score, im_score], dim = 0) score = score.norm(dim = 0) score = - score.sum(dim = 2)
I don't quite understand your training script about head-batch. Anyway, I have pretrained ent and rel embeddings without understanding it.
Now I want to figure out that given a triplet:
h [x1, ..., x200], r=[r1, ..., r100] and t=[y1, ..., y200], how to compute its score?A detailed description will be much appreciated.