How to reproduce iQIYI_VID baseline?(without MLP, 79.8%)

Hello, I want to reproduce iQIYI_VID benchmark using ArcFace. In ArcFace paper, ArcFace gets 79.8% MAP without MLP and 88.65% with MLP.

I already checked out MLP code and trained it.

I can get frame-level face features and face label. But, i can't get video-level face features and face label without MLP which is introduced in your github.

Please tell me how to get iQIYI_VID MAP without MLP.

thank you :)