search

facebookresearch / fvcore

Collection of common code that's shared among different research projects in FAIR computer vision team.
Apache License 2.0
1.93k stars 226 forks source link

don't use fp16 in giou loss #113

Closed lbin closed 1 year ago

lbin commented 1 year ago

There is no speed benefit using fp16 in losses anyway. There might be numerical problems instead since it's doing sigmoid / cross entropy. So should be good to change regardless of the bug.