InceptionV4/Inception-ResNet算法的简介(论文介绍)

      InceptionV4和Inception-ResNet是谷歌研究人员，2016年，在Inception基础上进行的持续改进，又带来的两个新的版本。

Abstract

       Very deep convolutional networks have been central to  the largest advances in image recognition performance in  recent years. One example is the Inception architecture that  has been shown to achieve very good performance at relatively  low computational cost. Recently, the introduction  of residual connections in conjunction with a more traditional  architecture has yielded state-of-the-art performance  in the 2015 ILSVRC challenge; its performance was similar  to the latest generation Inception-v3 network. This raises  the question of whether there are any benefit in combining  the Inception architecture with residual connections. Here  we give clear empirical evidence that training with residual  connections accelerates the training of Inception networks  significantly. There is also some evidence of residual Inception  networks outperforming similarly expensive Inception  networks without residual connections by a thin margin. We  also present several new streamlined architectures for both  residual and non-residual Inception networks. These variations  improve the single-frame recognition performance on  the ILSVRC 2012 classification task significantly. We further  demonstrate how proper activation scaling stabilizes  the training of very wide residual Inception networks. With  an ensemble of three residual and one Inception-v4, we  achieve 3.08% top-5 error on the test set of the ImageNet  classification (CLS) challenge.

摘要

       非常深的卷积网络是近年来图像识别性能最大进步的核心。一个例子是Inception 架构，已经证明它在相对较低的计算成本下获得了非常好的性能。最近，在2015年的ILSVRC挑战中，引入residual 连接和更传统的架构带来了最先进的性能；其性能类似于最新一代的Inception-v3网络。这就提出了这样一个问题：在将Inception 架构与residual 连接结合起来时是否有任何好处。在这里，我们给出了清晰的经验证据，证明使用residual 连接的训练显著加速了初始网络的训练。还有一些证据表明，residual Inception 架构网络的表现优于同样昂贵的Inception 网络，而无需residual 连接。我们还为残差和非残差初始网络提供了几种新的简化架构。这些变化显著提高了ILSVRC 2012分类任务的单帧识别性能。我们进一步证明了适当的激活比例如何稳定非常广泛的residual Inception网络的训练。利用三个residual 和一个Inception-v4,的集合，我们在ImageNet分类（CLS）挑战的测试集上实现了3.08% top-5 错误。

Conclusions

      We have presented three new network architectures in detail:

• Inception-ResNet-v1: a hybrid Inception version that has a similar computational cost to Inception-v3 from [15].

• Inception-ResNet-v2: a costlier hybrid Inception version with significantly improved recognition performance.

• Inception-v4: a pure Inception variant without residual connections with roughly the same recognition performance as Inception-ResNet-v2.

      We studied how the introduction of residual connections leads to dramatically improved training speed for the Inception architecture. Also our latest models (with and without residual connections) outperform all our previous networks, just by virtue of the increased model size.

结论

      我们详细介绍了三种新的网络架构:

•Inception-ResNet-v1:一个混合的Inception版本，其计算成本与[15]版本的incep -v3相似。

•Inception-ResNet-v2:一个成本更高的混合Inception版本，显著提高了识别性能。

•Inception-v4:一个没有residual 连接的Inception，与Inception-ResNet-v2的识别性能大致相同。

      我们研究了如何引入residual 连接来显著提高Inception体系结构的训练速度。此外，我们最新的模型(包括和不包括residual 连接)的性能优于所有以前的网络，这仅仅是因为模型的大小有所增加。

1、实验结果

1、Single crop -single model experimental results

Reported on the non-blacklisted subset of the validation set of ILSVRC 2012

单crop -单模型试验结果：在ILSVRC 2012验证集的非黑名单子集上的报告

2、144 crops evaluations -single model experimental results

采用了144个crops比single效果更好。

Reported on the all 50000 images of the validation set of ILSVRC 2012

3、Ensemble results with 144 crops/dense evaluation.

集成学习效果更好！

For Inception-v4(+Residual), the ensemble consists of one pure Inception-v4 and three Inception-ResNet-v2 models and were evaluated both on the validation and on the test-set.

4、训练过程中的速度比较

其中红色的Inception-resnet-v2效果性能***

(1)、Top-5 error evolution of all four models (single model, single crop)

模型尺寸较大时，性能改进。

尽管残差版本收敛得更快，但最终的准确性似乎主要取决于模型的大小。

(2)、Top-1 error evolution of all four models (single model, single crop)

This paints a similar picture as the top-5 evaluation.

其中红色的Inception-resnet-v2效果性能***

论文

Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, Alex Alemi.

Inception-v4, Inception-ResNetand the Impact of Residual Connections on Learning, 2016

https://arxiv.org/abs/1602.07261

Inception-v4算法的架构详解

DL之InceptionV4/ResNet：InceptionV4/Inception-ResNet算法的架构详解之详细攻略

Inception-ResNet算法的架构详解

     Inception-ResNet网络： 改进的Inception模块和残差连接的结合。引入residual connection直连，把Inception和ResNet结合起来，让网络又宽又深。

DL之InceptionV4/ResNet：InceptionV4/Inception-ResNet算法的架构详解之详细攻略

InceptionV4/Inception-ResNet算法的案例应用

后期更新……