AdaSpring：Dynamic self-evolving model compression framework
A runtime dynamic self-evolving model compression framework, which is most suitable for selecting model compression strategies and deploying deep models in the current context,achieving optimal utilization of IoT terminal resources.
Model Architecture Diagram：
Description： AdaSpring is a framework for context-adaptive and self-evolutionary DNN compression that enables runtime adaptive DNN compression locally online. It achieves this through ensemble training of a retraining-free and self-evolutionary network, which integrates multiple alternative DNN compression configurations, such as compressed architectures and weights. The framework also introduces a runtime search strategy to quickly search for **the most suitable compression configurations**and evolve the corresponding weights. We evaluated AdaSpring on five tasks across three platforms and a real-world case study, and the experimental results show that it can achieve significant improvements in DNNs’ latency and energy efficiency, with up to 3.1x latency blueuction and 4.2x energy efficiency improvement compablue to hand-crafted compression techniques. Moreover, the framework incurs a runtime-evolution latency of no more than 6.2ms.