Is LFW evaluation accuracy for resnet-27 0.989 not 0.998?

[AmitRozner]

When running evaluate.py I get:

best: 6000 -11.621302604675293 0.989

Is the resnet-27 accuracy a typo or am I missing something?

[huangyangyu]

@AmitRozner The result should be 0.9980. You could check the detail progress. Make sure all the environment is the same with the readme describe.

[ekarabulut]

@huangyangyu Same here.

According to the paper, the results should be the following:

I have downloaded SeqFace and run it as is with the precalculated feature files. It quickly completes the process and results in a score of 0.989 for LFW while I expected to see 0.998.

Are there any settings or anything else we should do to reach that higher score?

[huangyangyu]

@ekarabulut @AmitRozner Sorry to reply later. I will check it at weekend.

[huangyangyu]

@ekarabulut @AmitRozner Thank you find the accuracy bug. The problem is fixed now that feature vector is not normalized in code. Code is updated. The accuracy is 99.80% on LFW by using Resnet-27.