Mmdetection: How to apply focal loss in Cascade RCNN?
Hi, due to the high imbalance of the class distribution of my own dataset, I want to use focal loss to solve it. I tried change all the 'loss_cls' in cascade rcnn to
loss_cls=dict(
type='FocalLoss',
use_sigmoid=True,
gamma=2.0,
alpha=0.25,
loss_weight=0.005),
but the loss gets bigger and bigger. I wonder how to change the config file of cascade RCNN to use focal loss, or do you have any advice to help with the imbalance problem? Thanks!
wynterr
>All comments
It is not surprised. FocalLoss may not work well on two-stage detectors. For the imbalanced problem, it is beyond the scope of the codebase usage. My personal suggestion is that you may read some related papers, to name a few, OHEM, GHM, Libra R-CNN, etc.
hellock
on 29 Feb 2020
Related issues
fengxiuyaun
路
3Comments
hust-kevin
路
3Comments
happog
路
3Comments
dereyly
路
3Comments
henbucuoshanghai
路
3Comments
Most helpful comment
It is not surprised. FocalLoss may not work well on two-stage detectors. For the imbalanced problem, it is beyond the scope of the codebase usage. My personal suggestion is that you may read some related papers, to name a few, OHEM, GHM, Libra R-CNN, etc.