Neural network models face challenges such as an excessive number of parameters,high computational load,and rapidly increasing memory overhead.To accommodate the diversity of ap-plication scenarios and devices,it is essential to optimize neural network models for specific appli-cations.This work proposes a deep neural network pruning method that is aware of application categories.This method analyzes the distinct roles of various filters in extracting class features during the forward propagation process in convolutional neural networks.It identifies the relation-ship between the importance of filters and application categories,and carries out customized prun-ing optimization for different target categories in specific applications.The work is designed and implemented based on the PyTorch deep learning framework.Experimental results demonstrate that the application-aware neural network optimization method proposed in this paper can effec-tively optimize convolutional neural networks.
维普
万方数据