xiang song(charlie.song)
xiang song(charlie.song)
Can you install dgl-cuda10.1==0.4.3post2?
Can you add an assertion as `assert args.has_edge_importance == False` in kvclient and kvserver, as this is a workaround. You did not implement has_edge_importance for distributed training.
Did you see any OOM?
Did you check if the server is actually running?
GNN models are not supported yet, but we are planing to support it in next release along with the API refactor : https://github.com/awslabs/dgl-ke/issues/82
> ok. thank you very much. can I know your release date for the next version? I see the dgl has many GNN models, do you know it easily supports...
You can try use the dglke_predict and predict the scores for each r between h and t. The the top1 will be most relevant relation.
How about change the graph into only one relation type?
This can be a good point. We will provide python APIs in 0.2.0 release, at that time user can define their own Dataset loader.
The entity features mean the initial features? They you want to use a projection matrix W to convert them into embeddings like embs = feat @ W ?