About kg_triple.txt that generates node_group_one_hot_vector.pkl and so on file.

[zhiweihu1103]

Hello, I meet another question that can not reproduce the paper's result. The problem lies in the generation of node_group_one_hot_vector.pkl, node_group_one_hot_vector_multi.pkl, group_adj_matrix.pkl, and group_adj_matrix_multi.pkl files. I have carried out two designs to generate different above *.pkl files. Take FB15k-237 as example:

• experiment one: for kg_triple.txt, I merge the train, valid, and test triples together to produce the above *.pkl files, after training, the test results is similar to the paper's results.
• experiment two: for kg_triple.txt, I only merge the train, and valid triples together besides the test triple to produce the above *.pkl files, after training, the performance dropped sharply, only slightly better than Query2Box, and compared with experiment one above, it dropped by about 0.2.

I want to ask whether the kg_triple.txt used in your experiment aggregates train, valid, and test or only used train and valid triples?

Supplementary note: I checked your data/FB15k-237/kg_triple.txt. For the FB15k-237 data, there are 272115, 17535 and 20466 triples in the training set, validation set and test set respectively. When are processed by create_queries.py (from BetaE paper), there are a total of 620,158 triples in the integration, but the kg_triple.txt you provided has a total of 781,694 triples. I would like to ask how the extra 161536 triples are generated? Besides, I have checked the data/FB15k-237/from_to_map.pkl, the total triples in from_to_map.pkl is 620,158. Very confused for me, what is the actual number of triples used? Looking forward to your reply.

[cx51666666]

node_group_one_hot_vector.pkl, node_group_one_hot_vector_multi.pkl, group_adj_matrix.pkl, and group_adj_matrix_multi.pkl files.
Could you please send me a copy of the above documents?Thank you very much
[email protected]

[cx51666666]

The graphics card I used was 2080Ti, but I had a memory overflow problem.
Command:
CUDA_VISIBLE_DEVICES=1 Python -U codes/ run_model_newlook. py --do_train -- CUDa -- do_VALID --do_test --data_pathData /FB15k --model BoxNewLook -n 128-b 256 -d 400 -g 24 -a 1.0 -lr 0.0001 --max_steps 50000 -- CPU_num 1- 16-0.02 - geo center_reg test_batch_size box - task 1 c. c. 2 3 c. i. 2 3 i. IC. Ci. 2 u. Uc. 2 d. 3 d. Dc -- stepsforpath 50000-- Areopagivarearev --print_on_screen.

What type of graphics card do you use? What is the run command?
Thanks!