Margin Base Softmax in Face recognition

I 've found that Margin base Softmax is config different for each dataset in Arcface_torch source.
But in Arcface_torch config only have margin base softmax( config.margin_list) for Glint360k, ms1mv, wf. I want to ask how to estimate or get the best params for each dataset. Such as Casia basically

Start with the Arcface_torch configuration and its config.margin_list entry, then compare how the settings are represented for Glint360k, ms1mv, wf, and Casia. Done would be a documented, dataset-specific method for estimating or selecting the Margin Base Softmax parameters.

Written by the indexing model from the issue text.