miralab-ustc / kge-dura Goto Github PK

Hi, thanks for sharing the code, I have some questions about Equation 2 in this paper.

1. In the original ComplEx paper of ICML 2016, the scoring function is Re(<h, r, \overline{t}>). While in Equation 2 of your paper, the scoring function becomes Re(\overline{h}rt)
2. In Equation 2, why Re(\overline{h}rt) = Re(<h\overline{r}, t>)? I think Re(<h\overline{r}, t>) is not equal to the ComplEx' function Re(<h, r, \overline{t}>).

Looking forward to your response.

Since you mentioned before, h \overline{R} =(h_0r_0 - h_1r_1) + (h_0r_1 + h_1r_0) i, it can be seen as a complex vector where the real part is (h_0r_0 - h_1r_1) and the imaginary part is (h_0r_1 + h_1r_0), and the Euation (6) is the regularizer.

But why in your code the ||h \overline{R}||_2^2 = h^2 * r^2 = (lhs[0] ^ 2 + lhs[1] ^ 2) * (rel[0] ^ 2 + rel[1] ^ 2) = (h_0 ^2 + h_1 ^ 2) * (r_0 ^ 2 + r_1 ^ 2)

Since h \overline{R} is a new complex vector, where the real part is (h_0r_0 - h_1r_1) and the imaginary part is (h_0r_1 + h_1r_0), are you sure that its L2 norm's square ||h \overline{R}||_2^2 = (h_0 ^2 + h_1 ^ 2) * (r_0 ^ 2 + r_1 ^ 2)= ||h||_2^2 * ||r||_2^2 ?

Did I misunderstand anything about the L2 norm of a complex vector?

Thank you for your great work.

I'm a little confused about 2.1 Background, Tensor Factorization Based (TFB) KGC Models part, "Usually, a TFB KGC model factorizes Xj as $X_{j}` \approx Re(\overline{H}R_jT^T)$". May I ask why the conjugate of the head node is used here, instead of using the conjugate of the tail nodes like ComplEx?

Looking forward to your reply :)

您好，我在使用当前代码训练时模型无法收敛，模型性能始终为0。
请问这可能是什么情况呢？

环境：
torch 1.10
python 3.7
RTX 3090
cuda 11.4