Engineers typically run NAS on smaller proxy datasets and transfer their learned CNN architectures to the target task. This generalization method reduces the model's accuracy, however. Moreover, the same outputted architecture also is applied to all hardware platforms, which leads to efficiency issues.