Why IJB test result different when using different batch size?

The code i used: https://github.com/deepinsight/insightface/blob/master/evaluation/IJB/IJB_11.py

I got different test results with different batch size:

for IJB-C TAR@FAR=1E-5:

  | bs=32 (default) | bs=1024
-- | -- | --
Model1 | 52.55 | 57.87
Model2 | 53.09 | 58.31
Model3 | 51.27 | 56.07
Model4 | 53.87 | 59.48
Model5 | 53.93 | 58.84

The IJB-B also has similar behavior
It seems that bigger batch size will get better result, it's weird...

Start with evaluation/IJB/IJB_11.py and reproduce the reported IJB-C and IJB-B results using batch sizes 32 and 1024. Trace how batch size affects evaluation, then compare the outputs for the listed models; the issue is done when the discrepancy is explained and evaluation results are consistent or the behavior is clearly documented.

Written by the indexing model from the issue text.