一种面向电磁识别模型的分散计算方法

Decentralized calculation of neural network model for electromagnetic object detection

摘要:

基于张量分裂技术，设计了一种面向电磁目标识别的神经网络模型分散计算方法。该方法根据不同的隐藏层选择特定的张量分裂方法，将权重无损地分散到多个分布式节点上，以分散、聚合的方式完成协同推理计算。在树莓派设备上进行的仿真实验表明，该方法可以对集中式电磁识别模型进行无损拆分并分布式部署，可以保持与原始模型完全相同的准确率。并且当原始模型由于参数量过大而无法加载到内存中进行处理时，该方法仍可以正常完成计算。

Abstract:

Based on tensor splitting technique, a decentralized computing method of neural network model for electromagnetic object detection is introduced. In this method, different tensor splitting techniques are selected according to different hidden layers, and the weights are distributed to multiple distributed nodes losslessly. The simulation results on Raspberry PI show that this method can decompose and deploy the centralized detection model losslessly, and ensure the same accuracy as the original model. And when the original model is too heavy to be loaded into memory for calculation, this method can still complete the calculation properly.