笔记——Learning Transferable Architectures for Scalable Image Recognition

笔记——Learning Transferable Architectures for Scalable Image Recognition

主要思想:

​通过NAS(Neural Architecture Search Framework)在CIFAR-10 dataset上找到最优的网络结构,再扩展到ImageNet dataset上。
​模型提升——​ This paper 1.2%better computer cost: a reductionof 28%

定义优化问题:

  • 定义两种类型的convolution layer:
    ​ a. Normal Cell: 输入输出的feature map大小相同
    ​ b. Reduction Cell: 输出feature map height/2, width/2
  • 上述两种layer中可以包含如下操作:

笔记——Learning Transferable Architectures for Scalable Image Recognition

  • 整体架构如Figure2所示
    每当feature map大小减小一半,filter个数加倍2x
    笔记——Learning Transferable Architectures for Scalable Image Recognition

  • 优化目标参数:Normal layer的重复个数N,初始层的filter参数,两种layer的内部结构

优化过程:(随便贴个图,反正知道了也没用,需要500+GPU训若干天)

  • NAS(Neural Architecture Search Framework)
    笔记——Learning Transferable Architectures for Scalable Image Recognition
    笔记——Learning Transferable Architectures for Scalable Image Recognition

最优结构:

越来越接近生物神界结构
笔记——Learning Transferable Architectures for Scalable Image Recognition

结果评测:

([email protected],表示Normal Cell重复次数N=6,网络中倒数第二层的filter个数为768)

​ 无论是任务迁移(从分类应用到检测)还是模型压缩(移动设备)方面,NAS都表现很优秀
笔记——Learning Transferable Architectures for Scalable Image Recognition
笔记——Learning Transferable Architectures for Scalable Image Recognition