Measuring the Effects of Non-Identical Data Distribution for Federated Visual Classification