mmdetection use common BatchNorm without freeze?
In SSD series batchNorm is unfrozen
I detectron and other faster RCNN batchNorm is frozen.
I want to train with BN and big batch size 4*16. Do I need change something?
PS: this trick is used in SNIPER and in competion WIDER winner.
question
dereyly
All 3 comments
You can set the arguments bn_eval and bn_frozen of backbone field in the config file to False, then parameters of BN layers will be updated during training.
hellock
on 21 Nov 2018
👍2
@hellock Thank you!
Batch Norm train with "synchronize BN"?
dereyly
on 22 Nov 2018
Synchronized BN is currently not supported and we have the plan to add it later. If you set bn_eval and bn_frozen as False, the normal BN is used.
hellock
on 23 Nov 2018
Was this page helpful?
0 / 5 - 0 ratings
Related issues
yangcong955
路
3Comments
namheegordonkim
路
3Comments
fatLime
路
3Comments
songyuc
路
3Comments
qifei123
路
3Comments
Most helpful comment
You can set the arguments
bn_evalandbn_frozenofbackbonefield in the config file toFalse, then parameters of BN layers will be updated during training.