paperswithcode / sotabench-api

Easily benchmark Machine Learning models on selected tasks and datasets

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Allow to run without cache

Randl opened this issue · comments

As discussed in Slack, having option to ignore cache can be useful sometimes.

In particular, if I run model someone already run, the result is returned from cache. However, the runtime isn't fetched (or cached?). Then, it is still shown in leaderboard. In case the original run is in leaderboard it is duplicated, otherwise it appears only once but without runtime. Both seem to me undesirable.

Probably, cache should be attached to a repository, and there should be a way of either clearing it or ignoring it for a run, for example with "Re-run without cache" button.

Same problem if build fails half-way, see for example https://sotabench.com/user/EvgeniiZh/repos/Randl/ECANet#latest-results
The models which were measured in first build do not have runtime associated with them because it was measured in build which later failed.
A simple band-aid would be caching runtime too (that probably would require to invalidate existing caches tho)