MixFaceNets Evaluation Issue

[HamadYA]

Hi,
Thanks a lot for your work.

I tried to run the evaluation script of (in pocketnets github repo) for both pocketnets and mixfacenets. All models were evaluated as expected on all datasets except MixFaceNets version (without shuffle operation): MixFaceNet-XS, MixFaceNet-S, and MixFaceNet-M., as they achieved very low accuracy scores.

Do you have any suggestions or guidelines on what can be the issue?