论文地址 Inception V1 :Going Deeper with Convolutions Inception-v2 :Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Inception-v3 :Rethinking the Inception Architecture for Computer Vision Inception-v4 :Inception-Res
Google Inception Net,ILSVRC 2014比赛第一名.控制计算量.参数量,分类性能非常好.V1,top-5错误率6.67%,22层,15亿次浮点运算,500万参数(AlexNet 6000万).V1降低参数量目的,参数越多模型越庞大,需数据量越大,高质量数据昂贵:参数越多,耗费计算资源越大.模型层数更深,表达能力更强,去除最后全连接层,用全局平均池化层(图片尺寸变1x1),参数大减,模型训练更快,减轻过拟合(<Network in Network>论文),Inceptio
上一篇文章我们引出了GoogLeNet InceptionV1的网络结构,这篇文章中我们会详细讲到Inception V2/V3/V4的发展历程以及它们的网络结构和亮点. GoogLeNet Inception V2 GoogLeNet Inception V2在<Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift>出现,最大亮点是提出了Batch Normal