How do you set the MEAN?

[LikaQie]

Thanks for sharing your project ! I just follow your steps of alignment and extract features using your model and caffe_ftr.py，but I can just get acc 94% on LFW. So I want to ask whether you set the mean of data. And if you do ,how do you make your test data mean file?

[AlfredXiangWu]

I don't use the mean file for training and testing.

[LikaQie]

Thanks for your reply!
I use the Yi sun's detection and alignment method, and about 200 photos fail to detect the faces, do you think it would impact my results? I think the impact should be very small.
And I can not find out what it's wrong. Do you have some advice to me?

[AlfredXiangWu]

According to my experiments and others' email, the situation you mentioned may lead to 1%~2% dropped.

To improve the results on LFW, some advice is as follows：

1. Check the alignment method between training set and testing set. The image size for training set is 144144, while it is 128128 for testing. So it is possible to make a mistake.
2. Don't decrease the learning rate too quickly. I fixed the learning rate(1e-3) for about 1,000,000 iterations.