第七十四篇：机器学习优化方法及超参数设置综述

墨蓝 2023-06-04 02:52 8阅读 0赞

# 第七十四篇：机器学习优化方法及超参数设置综述 #

置顶 2019-08-25 23:03:44 [廖佳才][Link 1] 阅读数 207更多

分类专栏： [深度学习][Link 2]

本文链接： [https://blog.csdn.net/liaojiacai/article/details/100069249][https_blog.csdn.net_liaojiacai_article_details_100069249]

# 第七十四篇：机器学习优化方法及超参数设置综述 #

ps：很久没碰博客，长草了。。。应该坚持的。

# 摘 要 #

机器学习及其分支深度学习主要任务是模拟或者实现人类学习行为，这些学习方法近年来在目标分类、语音识别等各项任务中取得巨大突破。机器学的各种优化器极大了改善了学习模型的训练速度和泛化误差。优化方法和超参数作为观察训练模型的窗口，能够探索学习模型的结构和训练机制，是机器学习研究的重点之一。对机器学习的优化器与超参数理论研究进行了综述，回顾了超参数的一般搜索方法，对和优化器直接关联的批量大小、学习率超参数的设置方法进行了总结，对优化器和超参数需要进一步研究的问题进行了讨论。  
关键词：机器学习；深度学习；梯度下降；优化器；超参数；学习率；批量大小；反向传播：

## 引言 ##

梯度下降法(GD)\[1,2\] 是解决无约束最优问题的一种方法，被广泛的使用在当前的机器学习\[3,4\]优化算法中。随着机器学习的发展，训练数据达到百万以上，如ImageNet\[5\]，学习模型变得更深和更宽，训练时间长，收敛更慢，为了适应大数据和复杂模型的训练，解决这些训练问题，基于随机梯度下降扩展了各种优化器，并引入了更多的超参数，优化器的选择和超参数的设置影响网络的最终表现。深度学习的大多数模型可解释性较差，超参数及优化器可以作为一个观察、探索深度学习模型黑盒的一个工具，但是超参数设置仍然是当前机器学习训练的一个难题，手工调试超参数效率低下，近些年来提出了很多超参数的设置方法。本文从机器学习的问题、梯度下降和泛化误差的原理出发，分析梯度法与超参数的本质联系，并对基于理论产生的各种优化方法和超参数设置方法进行了总结。

## 机器学习优化算法：优化器及超参数 ##

![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 1]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 2]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 3]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 4]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 5]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 6]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 7]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 8]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 9]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 10]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 11]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 12]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 13]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 14]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 15]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 16]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 17]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 18]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 19]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 20]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 21]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 22]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 23]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 24]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 25]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 26]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 27]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 28]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 29]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 30]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 31]![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 32]![在这里插入图片描述][20190825230048494.JPG]  
后记：很久没写博客了，CSDN用上了markdown，感觉还是不方便，截图挺好的。我传了这篇文档的pdf，大家可以下载看一下，欢迎大家指正错误，一起进步。下载地址：[机器学习优化方法及超参数设置综述][Link 3]

### 参考文献 ###

\[1\] ROBBINS, H MONRO S. A stochastic approximation method\[J\]. Statistics, 1951: 102–109.  
\[2\] KALMAN R E. A New Approach to Linear Filtering and Prediction Problems\[J\]. Journal of Basic Engineering, 1960, 82(1): 35–45.  
\[3\] LECUN Y, BENGIO Y, HINTON G. Deep learning\[J\]. International Journal of Semantic Computing, 2016, 10(3): 417–439.  
\[4\] JÜRGENSCHMIDHUBER. Deep Learning in Neural Networks : An Overview\[J\]. Neural Networks, 2014, 61: 85–117.  
\[5\] DENG J, DONG W, RICHARD S et al. ImageNet : A Large-Scale Hierarchical Image Database\[J\]. 2009 IEEE Conference on Computer Vision and Pattern Recognition, IEEE, 2009: 248–255.  
\[6\] QING A. An Introduction to Optimization\[J\]. Antennas & Propagation Magazine IEEE, 2013, 38(2): 60.  
\[7\] YUAN Y. A Modified BFGS Algorithm for Unconstrained Optimization\[J\]. IMA Journal of Numerical Analysis, 1993, 11(3): 325–332.  
\[8\] MOKHTARI A, RIBEIRO A. RES : Regularized Stochastic BFGS Algorithm\[J\]. IEEE Transactions on Signal Processing, 2014, 62(23): 6089–6104.  
\[9\] ZHU C, H R, BYRD et al. Algorithm 778: L-BFGS-B: Fortran subroutines for large-scale bound-constrained optimization\[J\]. ACM Transactions on Mathematical Software, 1997, 23(4): 550–560.  
\[10\] WILLIAMS D E R E H J. Learning internal representations by error propagation\[M\]. MIT Press Cambridge, MA, USA ©1988, 1988.  
\[11\] HUTCHISON D, MITCHELL J C. Neural Networks: Tricks of the Trade\[M\]. 2012, 7700.  
\[12\] J. DUCHI, E. HAZAN and Y S. Adaptive Subgradient Methods for Online Learningand Stochastic Optimization\[J\]. 2010: :2121–2159.  
\[13\] GULCEHRE C, MOCZULSKI M, BENGIO Y. ADASECANT: Robust Adaptive Secant Method for Stochastic Gradient\[J\]. arXiv, 2014: arXiv:1412.7419.  
\[14\] ITERATION S S. A GENERALIZATION OF THE HERMITIAN ANDSKEW-HERMITIAN SPLITTING ITERATION\[J\]. BIT Numerical Mathematics, 2003, 43(5): 881–900.  
\[15\] SONG∗ S, CHAUDHURI† K, SARWATE‡ A D. Learning from Data with Heterogeneous Noise using SGD\[J\]. Journal of Machine Learning Research, 2015, 38: 1894–902.  
\[16\] CHATURAPRUEK S, DUCHI J C, R´E C. Asynchronous stochastic convex optimization : the noise is in the noise and SGD don’t care\[J\]. Neural Information Processing Systems 2015, 2015.  
\[17\] BOTTOU L, CURTIS F E, NOCEDAL J. Optimization Methods for Large-Scale Machine Learning\[J\]. SIAM Review, 2016, 60(2): 223–311.  
\[18\] BOTTOU L´EON O B. The Tradeoffs of Large Scale Learning\[J\]. Advances in Neural Information Processing Systems 20, Proceedings of the Twenty-First Annual Conference on Neural Information Processing Systems, Vancouver, British Columbia, Canada, December 3-6, 2007. Curran Associates Inc. 2007., 2007: 161–168.  
\[19\] PINHO M do R de, SHVARTSMAN I. Lipschitz continuity of optimal control and Lagrange multipliers in a problem with mixed and pure state constraints\[J\]. Discrete & Continuous Dynamical Systems - Series A (DCDS-A), 2012, 29(2): 505–522.  
\[20\] WOOD G R, ZHANG B E. Estimation of the Lipschitz Constant of a Function\[J\]. Journal of Global Optimization, 1996, 8(1): 91–103.  
\[21\] QIAN N. On the momentum term in gradient descent learning algorithms\[J\]. Neural Networks, 1999, 12(1): 145–151.  
\[22\] POLYAK, JUDITSKY. Acceleration of Stochastic Approximation by Averaging\[J\]. SIAM Journal on Control and Optimization, 2014, 30(4): 838–855.  
\[23\] YOSHUA B. Practical Recommendations for Gradient-Based Training of Deep Architectures\[J\]. arXiv, 2012: arXiv:1612.02803.  
\[24\] NESTEROV Y. Convex Optimization\[M\]. Springer, Boston, MA, 2013.  
\[25\] SU W, BOYD S, CANDES E J. A Differential Equation for Modeling Nesterov’s Accelerated Gradient Method: Theory and Insights\[J\]. Journal of Machine Learning Research, 2016, 17: 1–43.  
\[26\] BUBECK S, LEE Y T, SINGH M. A geometric alternative to Nesterov’s accelerated gradient descent\[J\]. Mathematics, 2015: 1–9.  
\[27\] DUCHI J, HAZAN E, YORAM SINGER et al. Adaptive Subgradient Methods for Online Learning and Stochastic Optimization ∗\[J\]. 2011, 12: 2121–2159.  
\[28\] JOHN C. ADADELTA: AN ADAPTIVE LEARNING RATE METHOD J. Mach. Learn. Res. 12 (2011), 2121 – 2159.\[J\]. 2011: 2009–2011.  
\[29\] BECKER S, BOBIN J, BOBIN J. NESTA : A Fast and Accurate First-Order Method for Sparse Recovery ∗\[J\]. SIAM Journal on Imaging Sciences, 2011, 4(1): 1–39.  
\[30\] DAUPHIN Y N, VRIES H De, BENGIO Y. Equilibrated adaptive learning rates for non-convex optimization\[J\]. NIPS’15 Proceedings of the 28th International Conference on Neural Information Processing Systems, 2015, 1: 1504–1512.  
\[31\] KINGMA D P, BA J L. ADAM: A METHOD FOR STOCHASTIC OPTIMIZATION\[J\]. ICLR, 2015: 1–15.  
\[32\] REDDI S J, KALE S, KUMAR S. On the Convergence of Adam and Beyond\[J\]. arXiv, 2019: arXiv:1904.09237.  
\[33\] DOZAT T. Incorporating Nesterov Momentum intoAdam\[J\]. - ICLR, 2016.  
\[34\] FENG J, ZHANG H, ZHANG, TIEYAN et al. Adaptive Natural Gradient Algorithm for Blind Convolutive Source Separation\[J\]. International Symposium on Neural Networks: Advances in Neural Networks. Springer-Verlag, 2007: 715–720.  
\[35\] AMARI S. Natural Gradient Works Efficiently in Learning\[M\]. Bradford Company Scituate, MA, USA ©1999, 1999.  
\[36\] PASCANU R, BENGIO Y. Natural Gradient Revisited for deep networks\[J\]. Computer Science, 2014, 37(10–11): 1655–1658.  
\[37\] DRIVALIARIS N A S J P D. Improved Jacobian Eigen-Analysis Scheme for Accelerating Learning in Feedforward Neural Networks\[J\]. Cognitive Computation, 2015, 7(1): 86–102.  
\[38\] KARAKIDA R, OKADA M, AMARI S. Adaptive Natural Gradient Learning Algorithms for Unnormalized Statistical Models\[J\]. Artificial Neural Networks and Machine Learning – ICANN 2016, 2016, 1: 427–434.  
\[39\] PARK H, AMARI S, FUKUMIZU K. Adaptive natural gradient learning algorithms for various stochastic models\[J\]. Neural Networks, 2000, 13(7): 755–764.  
\[40\] ZHAO J, YU X. Neurocomputing Adaptive natural gradient learning algorithms for Mackey – Glass chaotic time prediction\[J\]. Neurocomputing, Elsevier, 2015, 157: 41–45.  
\[41\] ZHANG Z, SUN H, PENG L. Natural gradient algorithm for stochastic distribution systems with output feedback ✩\[J\]. Differential Geometry and its Applications, Elsevier B.V., 2013, 31(5): 682–690.  
\[42\] GUO W, WEI H, TIANHONG LIU et al. An Adaptive Natural Gradient Method with Adaptive Step Size in Multilayer Perceptrons\[J\]. 2017 Chinese Automation Congress (CAC), 2017: 1593–1597.  
\[43\] MUTSCHLER M, ZELL A. PAL : A fast DNN optimization method based on curvature information\[J\]. arXiv, : arXiv:1903.11991.  
\[44\] HOSEINI F, SHAHBAHRAMI A, BAYAT P. AdaptAhead Optimization Algorithm for Learning Deep CNN Applied to MRI Segmentation\[J\]. Journal of Digital Imaging, Journal of Digital Imaging, 2019, 32(1): 105–115.  
\[45\] KESKAR N S, SOCHER R. Improving Generalization Performance by Switching from Adam to SGD\[J\]. arXiv, 2017: arXiv:1712.07628.  
\[46\] R. S Y, ROSE D C, THOMAS P. KARNOWSKI et al. Optimizing deep learning hyper-parameters through an evolutionary algorithm\[J\]. MLHPC ’15 Proceedings of the Workshop on Machine Learning in High-Performance Computing Environments, 2015(August 2016): 1–5.  
\[47\] TINCHCOMBE M. Multilayer Feedforward Networks are Universal Approximators\[J\]. Neural Networks, 1989, 2(5): 359–366.  
\[48\] CHANDRA P, SINGH Y. An activation function adapting training algorithm for sigmoidal feedforward networks\[J\]. Neurocomputing, 2004, 61: 429–437.  
\[49\] CHENEY M, BORDEN B, STATE C等. Regularization of Inverse Problems\[M\]. 2015.  
\[50\] BREUEL T M. The Effects of Hyperparameters on SGD Training of Neural Networks\[J\]. Computer Science, 2015: 1–19.  
\[51\] BENGIO Y. Continuous optimization of hyper-parameters\[J\]. Proceedings of the IEEE-INNS-ENNS International Joint Conference on Neural Networks. IJCNN 2000, 2000, 1: 1305.  
\[52\] SMITH S P. Differentiation of the Cholesky Algorithm\[J\]. Joumal of Computational and Graphical Statistics, 1995, 4(2): 134–147.  
\[53\] Y L, LÉON B, ORR G B et al. Efficient BackProp\[J\]. Neural Networks: Tricks of the Trade, this book is an outgrowth of a 1996 NIPS workshop . Springer-Verlag, 1998: 9–50.  
\[54\] BAO Y, LIU Z. A Fast Grid Search Method in Support Vector Regression Forecasting Time Series\[M\]. Intelligent Data Engineering and Automated Learning – IDEAL 2006, Springer Berlin Heidelberg, 2006.  
\[55\] BERGSTRA J, BENGIO Y. Random Search for Hyper-Parameter Optimization James\[J\]. Journal of Machine Learning Research, 2012, 13(1): 281–305.  
\[56\] BERGSTRA J, YAMINS D, COX D D. Making a science of model search: Hyperparameter optimization in hundreds of dimensions for vision architectures\[J\]. Proceedings of the 30th International Conference on Machine Learning, 2013: 115–123.  
\[57\] HUTTER F, HOOS H, LEYTON-BROWN K. Sequential Model-Based Optimization for General Algorithm Configuration Lecture Notes in Computer Science\[J\]. International Conference on Learning and Intelligent Optimization, 2011: 507–523.  
\[58\] BERGSTRA J, BARDENET R, YOSHUA BENGIO et al. Algorithms for Hyper-Parameter Optimization Algorithms for Hyper-Parameter Optimization\[J\]. 25th Annual Conference on Neural Information Processing Systems (NIPS 2011), 2011(December): 2546–2554.  
\[59\] LI L, JAMIESON K, GIULIA DESALVO et al. Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimization\[J\]. Journal of Machine Learning Research, 2016, 18: 1–52.  
\[60\] LORENZO P R, NALEPA J, MICHAL KAWULOK et al. Particle swarm optimization for hyper-parameter selection in deep neural networks-p481-ribalta\_lorenzo\[J\]. : 481–488.  
\[61\] KOUTNÍK J, SCHMIDHUBER J, GOMEZ F. Evolving deep unsupervised convolutional networks for vision-based reinforcement learning\[J\]. GECCO ’14 Proceedings of the 2014 Annual Conference on Genetic and Evolutionary Computation, 2014: 541–548.  
\[62\] KOUTNÍK J, SCHMIDHUBER J, GOMEZ F. Online evolution of deep convolutional network for vision-based reinforcement learning\[J\]. International Conference on Simulation of Adaptive Behavior, 2014: 260–269.  
\[63\] VERBANCSICS P, HARGUESS J. Image classification using generative neuro evolution for deep learning\[J\]. 2015 IEEE Winter Conference on Applications of Computer Vision, WACV 2015, IEEE, 2015: 488–493.  
\[64\] LEVY E, DAVID O E, NETANYAHU N S. Genetic algorithms and deep learning for automatic painter classification\[J\]. GECCO ’14 Proceedings of the 2014 Annual Conference on Genetic and Evolutionary Computation, 2014(Dl): 1143–1150.  
\[65\] LANDER S, SHANG Y. EvoAE-A New Evolutionary Method for Training Autoencoders for Deep Learning Networks\[J\]. 2015 IEEE 39th Annual International Computers, Software & Applications Conference, IEEE, 2015, 2: 790–795.  
\[66\] VERMA B, GHOSH R. Combination Strategies for Finding Optimal Neural Network Architecture and Weights\[J\]. Neural Information Processing: Research and Development, 2004: 294–319.  
\[67\] VILALTA R, DRISSI Y. A Perspective View and Survey of Meta-Learning\[J\]. Artificial Intelligence Review, 2002, 18(2): 77–95.  
\[68\] MOHAMED MAHER S S. SmartML : A Meta Learning-Based Framework for Automated Selection and Hyperparameter Tuning for Machine Learning Algorithms\[J\]. the 22nd International Conference on Extending Database Technology (EDBT), 2019: 554–557.  
\[69\] WISTUBA M, SCHILLING N, SCHMIDT-THIEME L. Hyperparameter optimization machines\[J\]. Proceedings - 3rd IEEE International Conference on Data Science and Advanced Analytics, DSAA 2016, IEEE, 2016: 41–50.  
\[70\] BARDENET R, BRENDEL M, BAL´AZS K´EGL et al. Collaborative hyperparameter tuning\[J\]. Applied Physics Express, 2011, 4(5): 199–207.  
\[71\] ILIEVSKI I, AKHTAR T, JIASHI FENG et al. Efficient Hyperparameter Optimization of Deep Learning Algorithms Using Deterministic RBF Surrogates\[J\]. AAAI, 2016: 822–829.  
\[72\] YOGATAMA D, MANN G. Efficient Transfer Learning Method for Automatic Hyperparameter Tuning\[J\]. Proceedings of the Seventeenth International Conference on Artificial Intelligence and Statistics, PMLR, 2014, 33: 1077–1085.  
\[73\] SNOEK J, LAROCHELLE H, ADAMS R P. Practical Bayesian Optimization of Machine Learning Algorithms\[J\]. NIPS’12 Proceedings of the 25th International Conference on Neural Information Processing Systems, 2012, 2: 1–12.  
\[74\] MUTNY M, KRAUSE A. Efficient High Dimensional Bayesian Optimization with Additivity and Quadrature Fourier Features\[J\]. 2018(NeurIPS): 9019–9030.  
\[75\] CHRIS THORNTON, FRANK HUTTER , HOLGER H. HOOS K L-B. Auto-WEKA: Combined Selection and Hyperparameter Optimization of Classification Algorithms\[J\]. Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining, 2013: 847–855.  
\[76\] AARON KLEIN S F, BARTELS, SIMON et al. Fast Bayesian Optimization of Machine Learning Hyperparameters on Large Datasets\[J\]. the 20th International Conference on Artificial In- telligence and Statistics (AISTATS) 2017, 2016, 54: 1–10.  
\[77\] MURUGAN P. Hyperparameters Optimization in Deep Convolutional Neural Network / Bayesian Approach with Gaussian Process Prior\[J\]. arXiv, 2017: arXiv:1712.07233.  
\[78\] KIM J, KIM S, CHOI S. Learning to Warm-Start Bayesian Hyperparameter Optimization\[J\]. NIPS, 2017: 1–14.  
\[79\] HINTON G E. A Practical Guide to Training Restricted Boltzmann Machines\[J\]. Neural Networks: Tricks of the Trade, 2012: 599–619.  
\[80\] CHENGWEI YAO D C, JIAJUN BU et al. Pre-training the deep generative models with adaptive hyperparameter optimization\[J\]. Neurocomputing, Elsevier B.V., 2017, 247: 144–155.  
\[81\] BOYAN A, MOORE A W. Learning evaluation functions for global optimization and boolean satisfiability\[J\]. AAAI, 1998: 3–10.  
\[82\] GAMBARDELLA L M, DORIGO M. Ant-Q: A Reinforcement Learning approach to the traveling salesman problem\[J\]. Machine Learning Proceedings 1995, 2014: 252–260.  
\[83\] ENGINEERS E. Global Search in Combinatorial Optimization using Reinforcement Learning Algorithms Global Search in Combinatorial Optimization using Reinforcement Learning Algorithms\[J\]. Proceedings of the 1999 Congress on Evolutionary Computation-CEC99, 1999: 189–196.  
\[84\] MOLL R, PERKINS T J, BARTO A G. Machine Learning for Subproblem Selection\[J\]. Proc. 17th International Conf. on Machine Learning, 2000: 615–622.  
\[85\] HANSEN S. Using Deep Q-Learning to Control Optimization Hyperparameters\[J\]. arXiv, 2016: arXiv:1602.04062.  
\[86\] POGGIO T. Regularization algorithms for learning that are equivalent to multilayer networks\[J\]. Science, 1990, 247(4945): 978–982.  
\[87\] BROWN B M. Martingale Central Limit Theorems\[J\]. The Annals of Mathematical Statistics, 1971, 42(1): 59–66.  
\[88\] BALLES L, ROMERO J, HENNIG P. Coupling Adaptive Batch Sizes with Learning Rates\[J\]. Association for Uncertainty in Artificial Intelligence, 2017: 410–419.  
\[89\] KROGH\* A, HERTZ J A. A Simple Weight Decay Can Improve Generalization\[J\]. NIPS’91 Proceedings of the 4th International Conference on Neural Information Processing Systems, 1991: 950–957.  
\[90\] BYRD R H, CHIN G M, NOCEDAL, JORGE et al. Sample size selection in optimization methods for machine learning\[J\]. Mathematical Programming, 2012, 134(1): 127–155.  
\[91\] MASTERS D, LUSCHI C. Revisiting Small Batch Training for Deep Neural Networks\[J\]. 2018: 1–18.  
\[92\] PEDREGOSA F. Hyperparameter optimization with approximate gradient\[J\]. the 33rd International Conference on Machine Learning, 2016, 48: 737–746.  
\[93\] LI M, ZHANG T, YUQIANG CHEN et al. Efficient mini-batch training for stochastic optimization\[J\]. Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining - KDD ’14, 2014: 661–670.  
\[94\] AKIBA T, SUZUKI S, FUKUDA K. Extremely Large Minibatch SGD: Training ResNet-50 on ImageNet in 15 Minutes\[J\]. arXiv, 2017: arXiv:1711.04325.  
\[95\] YOU Y, LI J, SASHANK REDDI et al. Large Batch Optimization for Deep Learning: Training BERT in 76 minutes\[J\]. arXiv, 2019: arXiv:1904.00962v3.  
\[96\] YIN P, LUO P, NAKAMURA T. Small Batch or Large Batch? Gaussian Walk with Rebound Can Teach\[J\]. Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2017: 1275–1284.  
\[97\] JADERBERG M, DALIBARD V, SIMON OSINDERO et al. Population Based Training of Neural Networks\[J\]. arXiv, 2017: arXiv:1711.09846.  
\[98\] HAZAN E, KLIVANS A, YUAN Y. Hyperparameter Optimization: A Spectral Approach\[J\]. ICLR, 2018: 1–18.  
\[99\] KESKAR N S, MUDIGERE D, NOCEDAL, JORGE et al. On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima\[J\]. ICLR, 2017.  
\[100\] HOCHREITER S, SCHMIDHUBER J. FLAT MINIMA\[J\]. Neural Computation, 1997, 9(1): 1–42.  
\[101\] DEVARAKONDA A, NAUMOV M, GARLAND M. AdaBatch: Adaptive Batch Sizes for Training Deep Neural Networks\[J\]. ARXIV, 2017: 1–14.  
\[102\] YOU Y, GITMAN I, GINSBURG B. Scaling SGD Batch Size to 32K for ImageNet Training\[J\]. arXiv, 2017: arXiv:1708.03888v2.  
\[103\] DE S, YADAV A, JACOBS D et al. Big Batch SGD: Automated Inference using Adaptive Batch Sizes\[J\]. International Conference on Artificial Intelligence and Statistics, 2016, 2017: 1–20.  
\[104\] DE S, YADAV A, DAVID J et al. Automated Inference with Adaptive Batches\[J\]. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, 2017, 54: 1504–1513.  
\[105\] HOFFER E, HUBARA I, SOUDRY D. Train longer, generalize better: closing the generalization gap in large batch training of neural networks\[J\]. Advances in Neural Information Processing Systems, 2017: 1729–1739.  
\[106\] SAMUEL L. SMITH∗ P-J K et al. Don’t Decay the Learning Rate, Increase the Batch Size\[J\]. ICLR, 2018.  
\[107\] SMITH S L, LE Q V. A Bayesian Perspective on Generalization and Stochastic Gradient Descent\[J\]. ICLR, 2018.  
\[108\] MANDIC D P, CHAMBERS J A. Towards the Optimal Learning Rate for Backpropagation\[J\]. Neural Processing Letters, 2000, 11(1): 1–5.  
\[109\] XIAO-HU YU, GUO-AN CHEN, SHI-XIN CHENG. Dynamic learning rate optimization of the backpropagation algorithm\[J\]. IEEE Transactions on Neural Networks, 1995, 6(3): 669–677.  
\[110\] SEONG S, LEE Y, YOUNGWOOK K et al. Towards Flatter Loss Surface via Nonmonotonic Learning Rate Scheduling\[J\]. In UAI2018 Conference on Uncertainty in Artificial Intelligence，Association for Uncertainty in Artificial Intelligence (AUAI), 2018.  
\[111\] POLYAK B T. Some methods of speeding up the convergence of iteration methods\[J\]. USSR Computational Mathematics and Mathematical Physics, 1964, 4(5): 1–17.  
\[112\] OBERMAN A M, PRAZERES M. Stochastic Gradient Descent with Polyak’s Learning Rate. (arXiv:1903.08688v1 \[math.OC\])\[J\]. arXiv Optimization and Control, 2019, 0167.  
\[113\] JEFFREY DEAN, GREG S. CORRADO R M. Large Scale Distributed Deep Networks\[J\]. NIPS’12 Proceedings of the 25th International Conference on Neural Information Processing Systems, 2012, 1: 1223–1231.  
\[114\] MARTENS J. Deep learning via Hessian-free optimization\[J\]. ICML’10 Proceedings of the 27th International Conference on International Conference on Machine Learning, 2010: 735–742.  
\[115\] SCHAUL T, ZHANG S, LECUN Y. No More Pesky Learning Rates\[J\]. the 30th International Conference on International Conference on Machine Learning, 2013, 28: 343–351.  
\[116\] RAVAUT M, GORTI S. Gradient descent revisited via an adaptive online learning rate\[J\]. arXiv, 2018: arXiv:1801.09136.  
\[117\] BENGIO Y. Practical Recommendations for Gradient-Based Training of Deep Architectures BT - Neural Networks: Tricks of the Trade: Second Edition\[M\]. Springer, 2012.  
\[118\] GEORGE A P, POWELL W B. Adaptive stepsizes for recursive estimation with applications in approximate dynamic programming\[J\]. Machine Learning, 2006, 65: 167–198.  
\[119\] LUO H, HANAGUD S. Dynamic learning rate neural network training and delamination detection\[J\]. AIAA, 2013, 35(9): 1522–1527.  
\[120\] SMITH L N. Cyclical learning rates for training neural networks\[J\]. Proceedings - 2017 IEEE Winter Conference on Applications of Computer Vision, WACV 2017, 2017, 1: 464–472.  
\[121\] MISHRA S, YAMASAKI T, IMAIZUMI H. Improving image classifiers for small datasets by learning rate adaptations\[J\]. arXiv, 2019: arXiv:1903.10726.  
\[122\] SCHAUL T, LECUN Y. Adaptive learning rates and parallelization for stochastic, sparse, non-smooth gradients\[J\]. arXiv, 2013: arXiv:1301.3764.  
\[123\] TAKASE T, OYAMA S, KURIHARA M. Effective neural network training with adaptive learning rate based on training loss\[J\]. Neural Networks, Elsevier Ltd, 2018, 101: 68–78.  
\[124\] YEDIDA R, SAHA S. A novel adaptive learning rate scheduler for deep neural networks\[J\]. arXiv, 2019: arXiv:1902.07399.  
\[125\] SMITH L N. A disciplined approach to neural network hyper-parameters: Part 1 – learning rate, batch size, momentum, and weight decay\[J\]. arXiv, 2018: arXiv:1803.09820.  
\[126\] LOSHCHILOV I, HUTTER F. SGDR: Stochastic Gradient Descent with Warm Restarts\[J\]. ICLR, 2017.  
\[127\] WU X, WARD R, BOTTOU L. WNGrad: Learn the Learning Rate in Gradient Descent\[J\]. arXiv, 2018: arXiv:1803.02865.  
\[128\] TAN† C, MA† S, AYU-HONG D et al. Barzilai-Borwein Step Size for Stochastic Gradient Descent\[J\]. the 30th International Conference on Neural Information Processing Systems, 2016: 685–693.  
\[129\] CHANDRA B, SHARMA R K. Deep learning with adaptive learning rate using laplacian score\[J\]. Expert Systems with Applications, Elsevier Ltd, 2016, 63: 1–7.  
\[130\] REN G, CAO Y, SHIPING W et al. A modified Elman neural network with a new learning rate scheme\[J\]. Neurocomputing, Elsevier B.V., 2018, 286: 11–18.  
\[131\] HSIEH H L, SHANECHI M M. Optimizing the learning rate for adaptive estimation of neural encoding models\[M\]. PLoS Computational Biology, 2018, 14(5).

转载于:https://www.cnblogs.com/think90/p/11610940.html

[Link 1]: https://me.csdn.net/liaojiacai
[Link 2]: https://blog.csdn.net/liaojiacai/article/category/9260513
[CC 4.0 BY-SA]: http://creativecommons.org/licenses/by-sa/4.0/
[https_blog.csdn.net_liaojiacai_article_details_100069249]: https://blog.csdn.net/liaojiacai/article/details/100069249
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70]: https://img-blog.csdnimg.cn/20190825221438618.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 1]: https://img-blog.csdnimg.cn/20190825221504337.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 2]: https://img-blog.csdnimg.cn/20190825224657206.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 3]: https://img-blog.csdnimg.cn/20190825224712312.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 4]: https://img-blog.csdnimg.cn/20190825224723856.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 5]: https://img-blog.csdnimg.cn/2019082522473998.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 6]: https://img-blog.csdnimg.cn/20190825224757495.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 7]: https://img-blog.csdnimg.cn/20190825224810984.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 8]: https://img-blog.csdnimg.cn/20190825224825537.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 9]: https://img-blog.csdnimg.cn/20190825224840719.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 10]: https://img-blog.csdnimg.cn/20190825224852371.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 11]: https://img-blog.csdnimg.cn/20190825224905104.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 12]: https://img-blog.csdnimg.cn/20190825224919878.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 13]: https://img-blog.csdnimg.cn/201908252249356.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 14]: https://img-blog.csdnimg.cn/20190825224947613.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 15]: https://img-blog.csdnimg.cn/20190825225353329.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 16]: https://img-blog.csdnimg.cn/20190825225412987.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 17]: https://img-blog.csdnimg.cn/2019082522544220.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 18]: https://img-blog.csdnimg.cn/20190825225506656.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 19]: https://img-blog.csdnimg.cn/20190825225532410.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 20]: https://img-blog.csdnimg.cn/20190825225601708.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 21]: https://img-blog.csdnimg.cn/2019082522562510.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 22]: https://img-blog.csdnimg.cn/20190825225648150.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 23]: https://img-blog.csdnimg.cn/20190825225710207.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 24]: https://img-blog.csdnimg.cn/20190825225733594.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 25]: https://img-blog.csdnimg.cn/2019082522580016.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 26]: https://img-blog.csdnimg.cn/20190825225822979.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 27]: https://img-blog.csdnimg.cn/20190825225839469.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 28]: https://img-blog.csdnimg.cn/20190825225915492.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 29]: https://img-blog.csdnimg.cn/20190825225934861.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 30]: https://img-blog.csdnimg.cn/20190825225954667.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 31]: https://img-blog.csdnimg.cn/20190825230012840.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk_size_16_color_FFFFFF_t_70 32]: https://img-blog.csdnimg.cn/20190825230030709.JPG?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2xpYW9qaWFjYWk=,size_16,color_FFFFFF,t_70
[20190825230048494.JPG]: /images/20230601/2b861908391b45e58d4bc85a40dad9e3.png
[Link 3]: https://download.csdn.net/download/liaojiacai/11612865