Multi-Agent Pruning and Quantization in Mixed-Precision Models

In order to improve the size reduction of the deep learning models for deploying to edge devices, this study employs the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) method, combined with the Preserve Ratio for each layer, to formulate different compression strategies without manual parameter tuning. Additionally, it evaluates the model compression effectiveness using metrics such as Top-1 accuracy, model size, parameter count, FLOPs, and latency. The results show that compressing MobileNet-V2 at FP32 precision on the target hardware reduces inference latency by 45.8%. Moreover, the model size decreases by 79.23%, the parameter counts by 60%, and the model computational load by 90.35%.