Get 99.03% performance on LFW

[playboywlh]

I reproduce your experience training with CASIA-WebFace and evaluate with LFW using SpherefaceNet-20 twice, but I get the accuracy 98.98% and 99.03% which don't match your achivement 99.30%. The preprocess, train and test all follow you and the config all set default except that I use the official caffe and I use one GPU GTX1070. I know it's reasonable to exist some difference in different environment, but the error rate differ almost 30%. I'm confuse whether it's GPU number that make the difference. And what should I adjust if I want to get your accuracy using one GPU. Thanks. Here are the two train logs： (1) 1 98.67% 2 98.83% 3 98.33% 4 99.33% 5 99.00% 6 99.33% 7 98.83% 8 98.67% 9 99.83% 10 99.00%

AVE 98.98%

(2) 1 98.50% 2 98.83% 3 98.50% 4 99.33% 5 98.83% 6 99.50% 7 98.83% 8 98.83% 9 99.83% 10 99.33%

AVE 99.03%

[wy1iu]

I guess batch size could be the major reason. You can still adjust the iteration size to use large batch szie on one GPU.