bxshi/proje

Embedding Projection for Knowledge Graph Completion

proje's Issues

Negative Sampling

Hi, which function is for generating negative samples?

Question about sampled_softmax

Hi, it's me again :)
I wonder what's the purpose of rescale in the sampled_softmax function?

about ProjE_wlistwise

hi, it's me again~
I'm confused about ProjE_wlistwise training process in the Appendix.

why no in the above formula？
what is the difference between ProjE_listwise and ProjE_wlistwise？

Reproducing relation prediction results

Hi Baoxu,

I was wondering whether you could provide a script to reproduce relation prediction results provided in the paper.

Cheers,

Tensorflow version

Hi,

Which version of tensorflow and python is used in the project?

Thanks.

ValueError

I tried executing the softmax version of projE using python 3 and tensorflow 1.1.0
I get this value error. How do I fix this?

I'm using the downloaded data as suggested in the readme (FB15k and WN18)

Entity embedding initialization

How can pre-trained entity embeddings be used as initialized embeddings rather than sampling from uniform distribution (as stated in the paper)?

Freezing

I was wondering what the specs of the machine were when you ran this test. I was unable to complete the test (following the instructions exactly) because at about a quarter of the way through, my entire machine froze. This could be due to the fact that I am running a virtual machine, but I gave it over 11 GB of memory and 4 processors.

Thank you.

about the parameter size

hi, @bxshi, in your paper, you said

Therefore, compared to simple models like TransE, ProjE only increases the number of parameters by 5k + 1, where 1, 4k, and k are introduced as the projection bias, combination weights, and combination bias respectively

but the combination weights De and Dr both are k*k diagonal matrices, each of them should have k parameters so that there would be 2k, why 4k in your paper?

About predict relation

Hi! I read "ProjE: Embedding Projection for Knowledge Graph Completion" and i want to result about relationship prediction task. but i think It does not seem to exist in this code about testing relationship prediction task. How can i training and evaluating that task?

About performance

Hi bxshi,
In gen_ht_r, it seems that you use both the validation set and test set to generate ht_r and tr_h. When I change it to use only the training set to generate ht_r and tr_h(I run ProjE_softmax.py as you recommended in the README file), filtered mean rank and hits@10 only reaches 74.3 and 0.675 after 12 iters, but according to the appendix in your AAAI paper, the model should yield filtered hits@10 over 0.8 around 12 iters.

Why yield test_triple in validation_data method?

I think the validation_data method should yields self.__valid_triple, but in ProjE_softmax.py it yields self.__test_triple, why?