akirato / perm-gaussiankg Goto Github PK

Hi, your work is very interesting and inspiring. However, I had some difficulties when I tried to repeat your experiments.

1. When I try your first code in run.sh (I just changed the dimension to 100 and keep all other arguments), it raises the error of tensor size mismatching. In fact, it occurs when i in some tasks. (I gauss there are some problems in the CenterIntersection class.)
2. In your paper, the part of the intersection seems to be non-parameterized, I wonder what the role of the two networks (GaussianOffsetIntersection and CenterIntersection ) on the intersection is in your model.
3. I gauss self.rank is the 'r' mentioned in Implementation Details in your paper. But why is the dimension of the center also multiplied by the self.rank? (For example, when I set the dimension to 100, the relation_embedding and offset_embedding are both 1600d. 16 is the self.rank.)
4. It seems that your memory L (size: d * r) instead of S=L*L^T as the inverse of the covariance matrix. However, I find you implement S_1+S_2 by L_1+L_2, but L_1L_1^T + L_2L_2^T is not the same as (L_1+L_2)(L_1+L_2)^T. (line 242-249 in models_gaussian.py)

I am sorry for asking you so many questions. Your work is so wonderful that I really look forward to hearing from you so that I can fully understand your work.

Hi,
Thanks for the very interesting work! I was wondering if you are going to share the model gaussian file as in the README?

Hi, could you please share your datasets used in the paper? Thanks a lot!

Hi, the repo does not have any links to the datasets used in the paper, esp. the new ones on DRKG and DBpedia. Do you have any plans sharing those?