search

facebookresearch / ConvNeXt

Code release for ConvNeXt model
MIT License
5.79k stars 696 forks source link

how to use syncbn when training with more than 1 card? #68

Closed clw5180 closed 2 years ago

clw5180 commented 2 years ago

It seems that ConvNext use LN, and can't change to syncbn

liuzhuang13 commented 2 years ago

Hi,

I think you can replace LN with syncbn. The way to write the code should be the same as syncbn in any other models. In practice, for imagenet classification, syncbn or normal bn shouldn't make a difference since the batch size is large enough for each gpu.

clw5180 commented 2 years ago

Hi,

I think you can replace LN with syncbn. The way to write the code should be the same as syncbn in any other models. In practice, for imagenet classification, syncbn or normal bn shouldn't make a difference since the batch size is large enough for each gpu.

Thanks to reply, I'm training object detection model, and I will try it.