摘要
随着深度神经网络在人工智能领域的广泛应用,其模型参数也越来越庞大,神经网络剪枝就是用于在资源有限设备上部署深度神经网络。该文通过新的优化策略-加速近端梯度(APG)、轻量级网络设计、非结构化剪枝和神经网络结构搜索(NAS)等手段相结合,实现对目标分类和目标检测等常见卷积神经网络模型的压缩剪枝,实验表明压缩剪枝后模型准确率不变,参数量下降91.1%,计算量下降84.0%。最后将压缩剪枝后模型的推断过程在嵌入式架构中实现,为深度学习在边缘端设备平台上的实现奠定了基础。
With wide applications of deep neural network in the field of artificial intelligence,the model parameters are getting larger and larger,and neural network pruning is used to deploy deep neural network on resource-limited devices. In this paper,through the combination of new optimization strategies-accelerated proximal gradient(APG),lightweight network design,unstructured pruning,and neural network structure search(NAS),common convolutions such as object classification and object detection are realized. The experiment results show that the accuracy of the model remains unchanged after compression pruning,the amount of parameters decreases by 91.1%,and the amount of calculation decreases by 84.0%. Finally,the inference process of the compressed and pruned model is implemented in the embedded architecture,which lays a foundation for the implementation of deep learning on the edge device platform.
作者
蒲亮
石毅
PU Liang;SHI Yi(Huazhong Institute of Electro-Optics-Wuhan National Laboratory for Optoelectronics,Wuhan 430223,China)
出处
《自动化与仪表》
2023年第2期15-18,24,共5页
Automation & Instrumentation
关键词
模型压缩
卷积神经网络
神经网络剪枝
神经网络结构搜索
model compression
convolutional neural network
neural network pruning
neural architecture search
