Pytorch PartialFC worse than Mxnet PartialFC

Hi,

Thank you for this amazing project, the work you're doing here is great.

I evaluated the pretrained pytorch R100 partialFC model against the equivalent Mxnet model and saw that the Mxnet model does better on IJB-C. Furthermore, when I try to retrain the partialFC model using the arcface_torch module, I still see the same result (Mxnet < Pytorch).

Were you able to replicate the Mxnet result with Pytorch? If so, can you share the training strategy you followed to make that model.

Thank you again for the excellent work.

Start with the arcface_torch module and compare the pretrained R100 PartialFC PyTorch model with its equivalent Mxnet model on IJB-C. Reproduce the reported evaluation and retraining setup, then document whether the discrepancy persists and which training strategy or configuration accounts for the result.

Written by the indexing model from the issue text.