AlexMRuch
commented
4 years ago I'd really appreciate this. For example, on https://aws-dglke.readthedocs.io/en/latest/train_user_data.html It's not super clear what should be in --data_path and --data_files.
For example, --data_path says "to specify the path to the knowledge graph dataset"; however, I presume this means "to specify the path to the folder containing the knowledge graph dataset".
Also, --data_files says "to specify the triplets of a knowledge graph as well as node/relation ID mapping"; however, it's not immediately clear the order of these files. For example, I would presume this would follow the order of the files listed under udd_[h|r|t]:
DGLBACKEND=pytorch dglke_train \
--data_path results_SXSW_2018 \
--data_files entities.tsv relations.tsv train.tsv valid.tsv test.tsv \
--format udd_hrt \
--model_name ComplEx \
--max_step 12000 --batch_size 1000 --neg_sample_size 200 --batch_size_eval 16 \
--hidden_dim 400 --gamma 19.9 --lr 0.25 --regularization_coef=1e-9 -adv \
--gpu 0 1 --async_update --force_sync_interval 1000 --log_interval 1000 \
--test^^^ But the order isn't clear. It seems like entities.txt and relations.tsv should go at the end since if someone uses to raw_udd_[h|r|t] option this would keep the first three elements consistently for training, validation, and testing files.
Perhaps there should be --data_tuple_files and --data_mapping_files options?
UPDATE: When I ran the code above, it gave me this output with FB_15k in the checkpoints, which doesn't seem right...
(dglke) amruch@wit:~/graphika/kg$ DGLBACKEND=pytorch dglke_train --data_path results_SXSW_2018 --data_files entities.tsv relations.tsv train.tsv valid.tsv test.tsv--format udd_hrt --model_name ComplEx --max_step 12000 --batch_size 1000 --neg_sample_size 200 --batch_size_eval 16 --hidden_dim 400 --gamma 19.9 --lr 0.25 --regularization_coef=1e-9 -adv --gpu 0 1 --async_update --force_sync_interval 1000 --log_interval 1000 --test
Using backend: pytorch
Logs are being recorded at: ckpts/ComplEx_FB15k_0/train.log
Reading train triples....
zheng-da
classicsong
We need to explain the arguments of commands.