-
Notifications
You must be signed in to change notification settings - Fork 7.2k
Closed
Description
🚀 The feature
Currently the meta-data of weights contain a few problematic fields that should be removed.
The first one is publication_year
field which originally meant to store the year an architecture is published. Unfortunately the current approach, doesn't account for cases where a technique is a combination of 2 or more papers (which is especially the case for post-paper optimizations). Adding the paper on the documentation (as we currently do) should be enough.
Another problematic field is interpolation
because it is stored twice (once on the transforms and once on the meta-data). There is no reason to have duplicate information so we should remove it.
cc @datumbox