googLeNet--Inception四部曲四Inception-ResNet and the Impact of Residual Connections on Learning
2017-01-11 14:07
609 查看
原论文:Going deeper with convolutions
作者:Christian Szegedy,Wei Liu,Yangqing Jia,Pierre Sermanet,Scott Reed,Dragomir Anguelov,Dumitru Erhan,Vincent Vanhoucke,Andrew Rabinovich
时间:February 2012
本文的大部分观点来自于这篇论文,并且加入了一些自己的理解。该博客纯属读书笔记。
基于对TensorFlow系统的使用,使得Inception architecture不需要过多的考虑分布式训练带来的诸多问题,不需要考虑模型的分割。这样我们的模型的结构会更简洁,计算量也更少。
文章提到当卷积核的数量超过1000个时,残差学习会变得不稳定,当随着迭代次数增加,参数基本为0,也就是说残差学习最终得到的只是一个恒等映射。而通过缩放残差因子x可以使得训练更加稳定。文章还否定了ResNet的two-phase 训练方法,先使用较低的学习率训练进行预热,然后使用高学习率。但是这种做法在卷积核数量很多的情况下也行不通。所以文章最终只采用缩放残差因子的方法进行训练。
实验结果如图1所示。总体效果差强人意,毕竟残差学习是微软的成果,而且文章只是把效果提升归结于模型规模的提升。
图1
作者:Christian Szegedy,Wei Liu,Yangqing Jia,Pierre Sermanet,Scott Reed,Dragomir Anguelov,Dumitru Erhan,Vincent Vanhoucke,Andrew Rabinovich
时间:February 2012
本文的大部分观点来自于这篇论文,并且加入了一些自己的理解。该博客纯属读书笔记。
Pure Inception blocks
基于对TensorFlow系统的使用,使得Inception architecture不需要过多的考虑分布式训练带来的诸多问题,不需要考虑模型的分割。这样我们的模型的结构会更简洁,计算量也更少。
Scaling of the Residuals
文章提到当卷积核的数量超过1000个时,残差学习会变得不稳定,当随着迭代次数增加,参数基本为0,也就是说残差学习最终得到的只是一个恒等映射。而通过缩放残差因子x可以使得训练更加稳定。文章还否定了ResNet的two-phase 训练方法,先使用较低的学习率训练进行预热,然后使用高学习率。但是这种做法在卷积核数量很多的情况下也行不通。所以文章最终只采用缩放残差因子的方法进行训练。实验结果如图1所示。总体效果差强人意,毕竟残差学习是微软的成果,而且文章只是把效果提升归结于模型规模的提升。
图1
相关文章推荐
- Inception系列3_Inception-v4:Inception-ResNet and the Impact of Residual Connections on Learning
- 【笔记】Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning
- Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning
- 《Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning》笔记
- 论文笔记 | Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning
- GoogleNetV4 Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning
- Inception-ResNet and the Impact of Residual Connections on Learning 论文阅读
- How to design DL model(2):Inception(v4)-ResNet and the Impact of Residual Connections on Learning
- Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning 论文笔记
- 使用JavaScript展开/折叠TreeView中所有节点(Expand and Collapse All Nodes of asp.net Treeview on the client with javascript)
- Get to the Top on Google: Tips and Techniques to Get Your Site to the Top of the Search Engine Ranki
- 【原】用使用JavaScript展开/折叠TreeView中所有节点(Expand and Collapse All Nodes of asp.net Treeview on the client with javascript)
- We're on the cusp of deep learning for the masses. You can thank Google later
- Machine-Learning Maestro Michael Jordan on the Delusions of Big Data and Other Huge Engineering Effo
- Zero-shot learning on the ImageNet of over 20000 classes
- HOW TO: Tune and Scale Performance of Applications That Are Built on the .NET Framework -zt from MS
- The Evolution Of LINQ And Its Impact On The Design Of C#
- HOW TO: Tune and Scale Performance of Applications That Are Built on the .NET Framework -zt from MS
- The Search : How Google and Its Rivals Rewrote the Rules of Business and Transformed Our Culture
- Latest preview of the ASP.Net MVC available on CodePlex