Closed rwightman closed 11 months ago
We have Flickr and COCO retrieval numbers on that table, are you thinking of some other datasets?
Right now we show average R@1 (image-to-text and text-to-image), but maybe we should expand this to show a breakdown (also R@5, R@10)
@gabrielilharco ah, didn't even realize that was retrieval w/o the R@
although, might make sense to break out into a diff table as zero-shot ranking and retrieval ranking be of separate interest to many down stream uses and those rankings are clearly different...
further, if we split retrieval into its own csv, makes sense to add more columns (R@1, R@5, R@10, image-to-text, and text-to-image) also useful because comparing to papers there isn't always consistency in terms of which numbers they're using
Will do @rwightman. We also have WinoGAViL as part of the 38 datasets (https://arxiv.org/abs/2207.12576). Maybe we can leave those three datasets (Flickr, MSCOCO and WinoGAViL) in the main table for completion and consistency with the datacomp benchmark, and keep a separate table for the three retrieval datasets?
@gabrielilharco yeah, could leave the main table as is, the full summary (also the link is out there so shouldn't change the name)... maybe even split off a zs-only variant since it's just post-filtering? I think more people would be happy to see a more detailed retrieval table as it's a big use case for these models.
Agreed, just did a PR, let me know how those look to you. Was surprised at how the rankings change!
@gabrielilharco awesome, yeah that's pretty useful to have them organized like this!
I feel the results csv has been getting a lot of eyes with it here closer to the models, esp with the new model additions.
Would it be worth adding retrieval results in a separate csv?