Have you trained sphereface 64 when m=0.35 and s=64, just in the condition of the paper? I

MuggleWang / CosFace_pytorch

Pytorch implementation of CosFace

MIT License

242 stars 52 forks source link

Have you trained sphereface 64 when m=0.35 and s=64, just in the condition of the paper? I #5

Open Reneea1 opened 6 years ago

Reneea1 commented 6 years ago

I trained sphereface64 network in the same super-parameter setting, but can not reach 99.% accuracy in LFW, but sphereface20 network can get 99.20% accuracy in LFW, which make me confused...why the deeper network works worse than the shallow network?

weihua04 commented 5 years ago

which dataset do u use? may be training set is not large enough for the deeper network