CFP_FP performance reproduce issue

First of all, thanks to all contributors for these valuable resources.

I downloaded `r100` model trained with glint360k from [this link](https://onedrive.live.com/?authkey=%21AFZjr283nwZHqbA&id=4A83B6B633B029CC%215582&cid=4A83B6B633B029CC)

And, when evaluating it with [this code](https://github.com/deepinsight/insightface/blob/996d844e12ed315b44935131c1f25bc4aef4e15e/recognition/arcface_torch/utils/utils_callbacks.py#L12-L48) and dataset file shared in this repo .

It reported the 97.x accuracy for `cfp_fp`.

However, I am afraid that this value is not consistent with that of README where the accuracy is reported as 99.x for `cfp_fp`.
Below is the provided log.

Is anyone who has ever encountered this issue or knows the reason?

Thank you!