search

wy1iu / sphereface

Implementation for <SphereFace: Deep Hypersphere Embedding for Face Recognition> in CVPR'17.
MIT License
1.59k stars 541 forks source link

Gamma and power settings for lambda? #6

Closed NothingL closed 7 years ago

NothingL commented 7 years ago

Hi wy1iu,

Asoftmax is a great work.I have some questions about how to set gamma and power for lambda. Using the settings in your prototxt ( base: 1000 gamma: 0.12 power: 1 lambda_min: 5 ) lambda will drop very quickly to lambda_min, is it really possible to converge? what is the best way to setting these for different networks and different size of training data?

Thanks.

wy1iu commented 7 years ago

The current parameter setting can converge on CASIA with our network architectures.

I would say there is no best way, but it is generally not difficult to set these parameters. One practical trick here is to use softmax to train a model first and then use A-Softmax to finetune on it. The other trick is to save intermedia models more frequently so that you can restart your training with slightly larger lambda when your network diverges.