When I train insightface with train_parallel.py scripts, which is the model parallelism implementation of MXNet, I change the para of config.num_workers, and ge
config.num_workers