Some details confuses me.

[LCorleone]

hi, first excellent job! but I have some problems.

1. In train_multi_gpu.py, your '_parse_function' use

   image = tf.subtract(image,127.5)
   image = tf.div(image,128.)

   However, when testing, you use prewhiten function

   def prewhiten(x):
   mean = np.mean(x)
   std = np.std(x)
   std_adj = np.maximum(std, 1.0/np.sqrt(x.size))
   y = np.multiply(np.subtract(x, mean), 1/std_adj)
   return y  

   It is a little different from above! Will these two different operations of image preprocessing degrades the performance when testing?

2. thank you for providing the pretrained model! Will you release the newest model?

[yule-li]

I'm sorry that I forget updating the test.py. Now the image preprocess within train phase and test phase is consistent. The lastest trained model is here.

[LCorleone]

@yule-li Thank you so much！I test on my own dataset and it seems that the performance is enhanced.

[Gabit07]

@yule-li Thank you so much！I test on my own dataset and it seems that the performance is enhanced.

Hello, I am just curious how many datasets did you use on training?

[LCorleone]

@Gabit07 hello, webface dataset. But till now i find cosface is not so good for face verification. Arcface is more accurate.

[Gabit07]

wow, thank you for your quick reply!

[Gabit07]

@Gabit07 hello, webface dataset. But till now i find cosface is not so good for face verification. Arcface is more accurate.

which implementation in github do you recommend to try?

[LCorleone]

@Gabit07 Insightface based on mxnet.