wordman2

Results 1 comments of wordman2

> we use deformable attention :) it's less compute intensive with modern model compilation. > > try it and see what happens! the rf-detr base at 560 should be similar...