摘要
学习率的大小影响着模型的训练速度和收敛精度,为了解决常用学习率AdaGrad的历史梯度干扰、AdaDec中幂指函数相关性不强的问题,从学习率衰减方式出发,提出一种基于衰减时效性的学习率改进方法(AdaRecur)。此学习率改进有两方面:1)通过设置衰减速率ρ减小历史梯度作用并结合当前梯度共同调整学习率;2)根据当前网络梯度的变化,将初始学习率替换为上轮训练中的学习率,以递推的方式调整学习率大小。目标位姿估计中LineMod数据集测试结果表明,在相同训练次数的情况下,AdaRecur比AdaGrad和AdaDec的平移、角度误差小,其中角度误差降低了2.378%,平移误差降低了2.216%,位姿估计的效果更加完美。
The size of learning rate affects the training speed and convergence accuracy of the model. In order to solve the problem of historical gradient interference of AdaGrad and weak correlation of power exponential function in AdaDec,an improved learning rate method(AdaRecur) is proposed based on the attenuation timeliness. There are two ways to improve the learning rate. Firstly,the learning rate is adjusted by setting the decay rateρ to reduce the effect of historical gradient and combining the current gradient.Secondly,according to the change of current network gradient,the initial learning rate is replaced by the learning rate in the previous round of training,and the learning rate is adjusted recursively. The test results of Line Mod data set in target pose estimation show that under the same training times,the translation and angle errors of AdaRecur are smaller than those of AdaGrad and AdaDec. The angle error and translation error are reduced by 2. 378% and 2. 216%,respectively. The effect of pose estimation is more perfect.
作者
张德
李国璋
王怀光
张峻宁
Zhang De;Li Guozhang;Wang Huaiguang;Zhang Junning(Army Engineering University Shijiazhuang Campus, Shijiazhuang 050003, China)
出处
《电子测量与仪器学报》
CSCD
北大核心
2019年第6期51-58,共8页
Journal of Electronic Measurement and Instrumentation
基金
国家自然科学基金(51205405、51305454)资助项目
关键词
深度学习
神经网络
位姿估计
自适应学习率
计算机视觉
deep learning
neural network
pose estimation
adaptive learning rate
computer vision