search

DingXiaoH / RepVGG

RepVGG: Making VGG-style ConvNets Great Again
MIT License
3.34k stars 435 forks source link

Weight decay will be applied to the weight of BN in identity branch and the weight of FC when using --custwd? #73

Closed wangjinwei94 closed 3 years ago

wangjinwei94 commented 3 years ago

It seems we skipped all parameters other than rbr_identity.weight and linear.weight ( https://github.com/DingXiaoH/RepVGG/blob/main/train.py#L86 ) Is it by-design or not?

DingXiaoH commented 3 years ago

Yes. It is by design. This is because we would not use L2 twice (weight_decay in optimizer + L2 loss in the total loss).

wangjinwei94 commented 3 years ago

Thanks! Sorry I missed the get_custom_L2