search

cavalleria / cavaface

face recognition training project(pytorch)
MIT License
463 stars 87 forks source link

MobileNetV3 #66

Open FeiMiBa opened 3 years ago

FeiMiBa commented 3 years ago

Hi,

For MobileNetV3, seems that you have done several modifications on the original implementation: disable h-swish, remove non-linear funtion in DW, add extra non-linear funtion after SElayer, change last layer to GDC, etc.

Does doing this particularly help a lot in face recognition task? What's the performance of the vanilla MobileNetV3? And what's you training strategy?

Thanks a lot.

John1231983 commented 3 years ago

I have tried mobinetv3 with small configure but it does not work. Any reason?

bobo0810 commented 3 years ago

Hi,

For MobileNetV3, seems that you have done several modifications on the original implementation: disable h-swish, remove non-linear funtion in DW, add extra non-linear funtion after SElayer, change last layer to GDC, etc.

Does doing this particularly help a lot in face recognition task? What's the performance of the vanilla MobileNetV3? And what's you training strategy?

Thanks a lot.

GhostNet also made similar changes. What is the reason?