Upsample or Upweight? Balanced Training on Heavily Imbalanced Datasets