首页|基于Winograd算法的3D卷积神经网络权重剪枝方法

基于Winograd算法的3D卷积神经网络权重剪枝方法

扫码查看
针对3D卷积神经网络在资源有限的环境下高计算成本的挑战,文章提出了一种融合Winograd算法和网络剪枝技术的3D卷积神经网络优化方法.首先,将标准3D卷积层替换为效率更高的3D Winograd层,实现对卷积操作的优化.接着,对3D Winograd层的权重进行重要性评估,保留重要的权重单元并剪枝获得稀疏模型.最后,对稀疏模型进行重训练,恢复剪枝后网络的性能.通过结合Winograd算法和网络剪枝技术,能够在提高识别准确度的同时,显著降低了模型的计算需求.实验结果证实,与其他优化技术相比,本方法能有效减少计算资源消耗,同时保持甚至提高识别性能.
3D Convolutional Neural Networks Weight Pruning Based on Winograd Algorithm
In response to the high computational demands of 3D Convolutional Neural Networks(CNNs)in environments with limited resources,this paper proposes an optimization approach that synergizes the Winograd algorithm with network pruning techniques.Firstly,the method replaces conventional 3D convolution layers with more efficient 3D Winograd layers to expedite convolution processes.Secondly,it involves assessing the importance of weights within the 3D Winograd layers,which facilitates the formation of a sparse model via pruning.Finally,the sparse model is retrained to further recover network performance.This combined use of the Winograd algorithm for computational efficiency and network pruning for model simplification results in a substantial reduction in computational needs while maintaining or improving the ac-curacy of recognition.Comparative experimental results affirm the proposed approach signifi-cantly minimizes computational resource use,maintaining or even boosting recognition capabili-ties compared to other optimization methods.

Optimization of 3D CNNsWinograd AlgorithmNetwork Pruning

邹贵、秦子然、吴捷、刘国梁、赵军、王迎雪、林晖、林巍峣

展开 >

上海交通大学电子信息与电气工程学院,上海 200240

中国航空无线电电子研究所,上海 200233

中国电子科技集团有限公司电子科学研究院,北京 100041

3D卷积神经网络优化 Winograd算法 网络剪枝

2024

长江信息通信
湖北通信服务公司

长江信息通信

影响因子:0.338
ISSN:2096-9759
年,卷(期):2024.37(8)