does the code of classification branch has the loss of reg?

[zimenglan-sysu-512]

in the paper, before the grid-branch, the proposals from rpn will go through the classificaiton branch. so in the classification branch, it only has the cls-loss? or still has the both cls-loss and reg-loss?

[luxiin]

Only cls-loss. We simply discard reg-loss by overriding loss function in BBoxHead.

[zimenglan-sysu-512]

hi @luxiin
can u show me the code of it, since i can't find it.
thanks.

[luxiin]

Grid-R-CNN/mmdet/models/bbox_heads/convfc_bbox_head.py

Line 218 in 23ef988

losses['loss_cls'] = weighted_cross_entropy(

[luxiin]

@zimenglan-sysu-512 We override the loss function of BBoxHead, it's original version contains reg loss like this:

Grid-R-CNN/mmdet/models/bbox_heads/bbox_head.py

Line 103 in 23ef988

losses['loss_reg'] = weighted_smoothl1(

[zimenglan-sysu-512]

thanks @luxiin
btw, in mmdetection, it seems to use reg-loss in the config file and in the detectors/grid_rcnn.py.

[luxiin]

In mmdetection, one can disable the reg-loss explicit in config file like this:
https://github.com/open-mmlab/mmdetection/blob/713e98bc1bfb842760df6be73b9520ee775d3c06/configs/grid_rcnn/grid_rcnn_gn_head_r50_fpn_2x.py#L36
The feature is not implemented in this repo.

[zimenglan-sysu-512]

u are right.
thanks.