Open LeoYML opened 1 year ago
same question
And I apply the ExternalEmbedding to my code, to enhance my train efficient, I found I can't get statisfied result. Does it's matters with dgl Graph.
And I apply the ExternalEmbedding to my code, to enhance my train efficient, I found I can't get statisfied result. Does it's matters with dgl Graph.
I have a question need your help! We can communicate together. On the code, I see pos_g has edges 1024 and neg_g also has edges 1024, it corrupts triplet 1 time, but on the paper you say corrupt every triplet k times. Is it correct? Or does it just mean 1024=4*256, 256 is chunk size, and on this way you finish group triplets into 4 chunks and corrupt them together then we get 1024
As I understand, It corrupts together, the candidates(neg_samples) for per triplet is same.
I have implemented another version that different triplets use different candidates, but the training is not efficient and the GPU memory is so large (as expected).
By the way, I am working on implement the adam by myself, I think this project is not active enough, If somebody would like to communicate with me directly, can email yueling.me@qq.com or add my wechat, I am working on improving this pipeline.
I have also tried to implement adam myself, but doesn't adam take up a lot of memory?
Now kge framework is a lot, pykeen framework is quite active, better than dglke
pykeen
Thanks, and are there any other KGE models to recommend besides pykeen, do you learn openKE?
openke is a relatively early project, has not been very active, currently see pykeen is always updated framework
Hi, Thanks for your excellent jobs!
Why not support Adam and Adagrad?
Is this project updating? or your team is working on new projct.