Closed Filco306 closed 2 years ago
Hi, I think the problem is not that you use KvsAll instead of negative sampling but rather that the evaluation yields OOM. We score against all entities by default during evaluation. Especially for rotate this leads to a high memory consumption. You can reduce the memory overhead during evaluation by only scoring against chunks of all entities. You can do so with the option:
entity_ranking.chunk_size: 5000
The smaller the chunk size the smaller the memory consumption. You can find the relevant documentation here: https://github.com/uma-pi1/kge/blob/ed53b69aff350de33b236736c86e1ac4e33e3421/kge/config-default.yaml#L529
I agree. Note that the KvsAll line that you marked is the one form the default config, which is later modified by the search. You can see the actual config being used in the folder of the trial; all changes are also printed in the console after "Created trial ... with parameters".
I feel stupid now. Thank you, you are entirely correct. Still some things to learn about this super nice package :)
Hello!
I have this
.yaml
-config file for doing anAxSearchJob
:However, when I try to run it, I get OOM issues because it still performs a KvsAll training job. How do I disable the option to run KvsAll-jobs completely in the search?
Thanks!
Here is my log output: