摘要
尽管神经网络在各个领域都取得了长足的进步,但是在移动设备上的应用却受限于巨大模型所需要的计算能力以及存储空间。神经网络剪枝技术可以使训练后的网络参数减少,减少存储需求,提高计算效率。最近一些研究基于权重的绝对值(L1范数)进行剪枝,可以在不损失过多精度的情况下有效地压缩模型。本文在此基础上结合了权重的变化程度进行迭代剪枝,能够对网络模型进一步压缩。文章提出的重要性评估方法分别在结构化与非结构化剪枝策略中在全连接网络以及卷积网络进行了实验,结果表明均优于仅依靠权重绝对值的剪枝方法。
Although neural network has made great progress in various fields, its application in mobile devices is limited by the computing power and storage space required by the huge model. Neural network pruning technology can reduce network parameters after training, reduce storage requirements and improve calculation efficiency. Recently, some researches based on the weight of the absolute value(L1 norm) pruning can effectively compress the model without losing too much precision. On this basis, this paper combines the degree of weight change to carry out iterative pruning, which can further compress the network model. The importance evaluation methods proposed in this paper are tested in the full connection network and convolution network respectively in the structured and unstructured pruning strategies. The results show that they are better than the pruning methods only relying on the absolute weight.
出处
《信息通信》
2019年第12期165-167,共3页
Information & Communications
作者简介
徐嘉荟(1995-),男,江苏南通人,工学硕士,东南大学信息科学与工程学院学生,研究方向:信号处理,深度学习。
