Quantization and Deployment Study of Classification Models for Embedded Platforms
Zihan Huang, Jin Jin, Chaolong Zhang, Zhijie Xu, Yuanping Xu, Chao Kong, Wen Qin, Dan Tang · 2023
Deep learning models find extensive applications across various domains. However, their large number of parameters, high storage requirements, and computational overhead pose challenges for deploying these models on resource-constrained embedded devices. This study focuses on addressing this issue by exploring techniques to optimize and deploy lightweight models on embedded devices. The approach involves optimization and adjustment of the model, followed by model conversion, quantization, and quantization calibration, aimed at reducing model size and improving inference speed. Notably, improvements are made to the quantization calibration algorithm to mitigate accuracy loss caused by model quantization. The experimental results demonstrate that light quantization significantly reduces model size, facilitating storage on embedded devices. Although there is a slight reduction in accuracy, the inference speed is substantially improved, enabling real-time human face recognition in video scenarios.