andreaskoepf/faster-rcnn.torch

about gradient scaling

Opened this issue · 0 comments

hi thanks for sharing your nice code
in objective.lua, you scaled gradient by cls_count, which means total number of roi across all mini batches

it seems correct for cnet because total number of batch size for cnet is equal to total number of roi

but in case of pnet, input seems 3 dimensional image and dividing gradparam by total number of roi seems incorrect for pnet. it seems gradParams of pnet should be divided by total number of batch(#batch)

what do you think about this?