soumyasanyal
commented
4 years ago For Transe variant of the CompGCN model, we found that setting the prior rel_inv = -rel helps the model performance (also seemed logical to do so). Its not directly clear how to use such priors for other models, hence we didn't try for them. You can check this line to understand that we initialize with half the embedding size for Transe variant and later concat with the negative vector.
r = self.init_rel if self.p.score_func != 'transe' else torch.cat([self.init_rel, -self.init_rel], dim=0) why cat them when use ‘transe’？
cheers