afshinsadeghi / mde_graphical Goto Github PK

before trainin should prcess files using process_files_gmde/store_graph_properties.py it generates a folder inside the folder of each dataset that includes the graphicaö geatures

running is similar to RotatE and MDE_adv , plus a folder that incldues node features in the path: -node_feat_path

with best hyperparameters on wn18rr

CUDA_VISIBLE_DEVICES=2 python run.py --do_train --do_test -save ../experiments/kge_baselines_wn18rr --data_path ../data/WN18RR --model MDE -n 624 -b 210 -d 300 -g 2.5 -a 2.5 -adv -lr .0005 --max_steps 250000 --test_batch_size 2 --valid_steps 10000 --log_steps 10000 --do_valid -node_feat_path ../data/WN18RR/train_node_features --cuda -psi 14.0

in cas it does not fit in memory reduce -n and id not the -d and reduce the --max_steps if trains fast

running is without --triples_are_mapped when runned without dic files are generated in process_datasets/convert_files.py