Setting of deploy

[banxia1994]

Should we save the norm1 layer for deploy? or just get the output from fc5.

[happynear]

Here is an example of deployment.
https://github.com/happynear/AMSoftmax/blob/master/prototxt/face_deploy_mirror_normalize.prototxt

I apply the mirror face trick and normalization inside the network. These two operations are implemented using CUDA, so the speed is faster than doing them outside on CPU.

[banxia1994]

thx

[banxia1994]

Hi, i trained AMsoftmax with webface which including 10572 subjects, get the accuracy 99.27% when the deploy file without norm1 layer in LFW, and get the accuracy 60% with norm1 layer. but In your example of deploy add the norm layer and some tricks, am i missed something in valid process? or some wrong with my valid code?(not your code)

[happynear]

It's weird. The normalize layer just does what it should do: normalization, which you would also do during testing. Maybe you need to check your codes...

[banxia1994]

https://github.com/happynear/FaceVerification/blob/master/test_lfw.m Is this your test code?

[banxia1994]

sorry to bother you again, Have you released your test code in LFW? thx kindly!

[happynear]

We use the same network architecture as SphereFace. You may use SphereFace's experimental settings.