unable to reproduce the result with lfw at FAR 1e-5

Hi,

I tried to generate a new large scale lfw benchmark, the protocol is similar to blurf (http://www.cbsr.ia.ac.cn/users/scliao/projects/blufr/)

From the model zoo, I test glint360k_r50 and got 99% accuracy at FAR 1e-5, but I can't reproduce the result from the training code. my result is only 95%

any idea?

Start by comparing the model-zoo glint360k_r50 evaluation with the training-code configuration and the BLUFR-like LFW protocol described in the issue. Confirm whether the same data, checkpoint, and FAR 1e-5 measurement produce the reported 99% result; otherwise document the discrepancy and its cause.

Written by the indexing model from the issue text.