The rapid evolution of deep learning (DL) models necessitates effective compression and scaling methods for their development. Conventional approaches typically involve modifying the model’s components, which affects the outcomes across all cases. However, the complexity of inputs varies. More complex samples require larger models, while smaller models are enough for simpler samples. Additionally, the preference of model scales also depends on the scenario. A fully charged device can handle full computation, while a low-battery device needs to conserve power. This project aims to develop dynamic DL systems which can dynamically allocating computation based on the varying input complexity and scenario.