我在使用您提供的网络结构训练之后，数据集采用CSAIA-webface，但是训练后在lfw上的精度只有97.87，我想问一下您是什么问题呢？

Thanks for your work. I want to use the mobilefacenet and it would be nice for you if you could provide the training log file. Thanks!

I test the model you give out, the acc is lower than yours
1 group acc is 0.968333
2 group acc is 0.976667
3 group acc is 0.980000
4 group acc is 0.983333
5 group acc is 0.995000
6 group acc is 0.988333
7 group acc is 0.990000
8 group acc is 0.988333
9 group acc is 0.990000
10 group acc is 0.978333
the acc is 0.983833

The lfw aligned face images I used come form insightface project, origin image size is 112X112, I crop it to 112*96 by cut width.
I also tried resize 112X112 to 112X96, the acc is 0.984000

is there any difference between my lfw and yours?

作者您好：
请问您训练的caffe 下的模型是否是，先只用softmax,训练到12万步，再此基础上，在模型中加入arcface损失层，进行finetune训练,, ？
最后训练的模型怎么进行LFW精度的测试昵？

normalize_layer

thanks for you great contribution! i just checked your training log but i saw the final accuracy on your training set is only 64%, can it achieve 99.5% accuracy on LFW dataset?

what's the tpr when far is at 0.0001%?

i using vggface2[alignement is same to sphereface] to train my model according to your model_prototxt and solver_prototxt, i get the result is 99.06%, the model is 8.5M. but i use your model i get the result is 99.48%. the model is 4M. Do you use any other skill? why the model size is different?

您好！为什么我训练到第三轮（300）的时候loss值就下降到1以下，请问是什么原因导致的？

你好。请问你用的是那个mtcnn版本以及对齐方式。因为mtcnn人脸检测和对齐的差异，会导致人脸识别的结果不一致。所以我想统一下。

这个训练后要怎么测试呢

你好！感谢分享
模型好像和论文有以下不一致的地方：
1、论文第一个bottleneck，n = 5，s=2，您给出的模型n=4,s=1，而且在此之前还有一个n=1,s=2的bottleneck，这样看上去似乎与后面更加一致，是有意修改的吗？
2、这个模型的loss公式并不是cos(θ+m)，而是cos(θ)+m，所以对应的LabelSpecificAdd层应该替换成https://github.com/xialuxi/arcface-caffe里的CosinAddm