Optimal Corpus Aware Training for Neural Machine Translation