Why not batch normalization?

wy1iu / sphereface

Implementation for <SphereFace: Deep Hypersphere Embedding for Face Recognition> in CVPR'17.

MIT License

1.58k stars 543 forks source link

Open huangbiubiu opened 6 years ago

huangbiubiu commented 6 years ago

I found that SphereFace network removed batch normalization layer? Why?

jimeffry commented 6 years ago

Maybe the batch_norm layer could give a fixed value , when the result of the layer is Nan.