To prune, or not to prune: exploring the efficacy of pruning for model compression

   Abstract