about running time about relphormer HOT 4 CLOSED

It depends on the data size. For dataset such as WN18RR or FB15K237, the whole procedure usually take one or two days in one 3090 gpu (24 GB).

from relphormer.

It depends on the data size. For dataset such as WN18RR or FB15K237, the whole procedure usually take one or two days in one 3090 gpu (24 GB).

Thanks. But I used your default parameters to test on 'umls' dataset, and found that there is a certain gap between the running results and the paper data. Is there any modification in other parts?

from relphormer.

Hi, it is caused by the random sampling strategy or pre-training procedure, especially for the small dataset such as UMLS. You can try to pre-train the task with more epochs or use our pre-training checkpoints (which will be attached in this repository).

from relphormer.

Sorry for the inconvenience.

from relphormer.