当前位置:网站首页>深度学习——网络中的网络以及1x1卷积
深度学习——网络中的网络以及1x1卷积
2022-06-30 07:22:00 【头发没了还会再长】
1x1卷积
在架构内容设计方面,其中一个比较有帮助的想法是使用 1×1 卷积。
1x1卷积?
过滤器为 1×1,这里是数字 2,输入一张 6×6×1 的图片,然后对它做卷积,起过滤器大小为 1×1×1,结果相当于把这个图片乘以数字 2,所以前三个单元格分别是 2、4、6 等等。用 1×1 的过滤器进行卷积,似乎用处不大,只是对输入矩阵乘以某个数字。但这仅仅是对于6×6×1 的一个通道图片来说,1×1 卷积效果不佳。
如果是一张 6×6×32 的图片,那么使用 1×1 过滤器进行卷积效果更好。具体来说,1×1 卷积所实现的功能是遍历这 36 个单元格,计算左图中 32 个数字和过滤器中 32 个数字的元素积之和,然后应用 ReLU 非线性函数。这个 1×1×32 过滤器中的 32 个数字可以这样理解,一个神经元的输入是 32 个数字(输入图片中左下角位置 32 个通道中的数字),即相同高度和宽度上某一切片上的 32 个数字,这 32 个数字具有不同通道,乘以 32 个权重(将过滤器中的 32 个数理解为权重),然后应用 ReLU 非线性函数,在这里输出相应的结果。一般来说,如果过滤器不止一个,而是多个,就好像有多个输入单元,其输入内容为一个切片上所有数字,输出结果是 6×6 过滤器数量。
所以 1×1 卷积可以从根本上理解为对这 32 个不同的位置都应用一个全连接层,全连接层的作用是输入 32 个数字(过滤器数量标记为𝑛𝐶[𝑙+1],在这 36 个单元上重复此过程),输出结果是 6×6×#filters(过滤器数量),以便在输入层上实施一个非平凡(non-trivial)计算。这种方法通常称为1×1 卷积,有时也被称为Network in Network.
example
假设这是一个 28×28×192 的输入层,你可以使用池化层压缩它的高度和宽度,这个过程我们很清楚。但如果通道数量很大,该如何把它压缩为 28×28×32 维度的层呢?你可以用 32个大小为 1×1 的过滤器,严格来讲每个过滤器大小都是 1×1×192 维,因为过滤器中通道数量必须与输入层中通道的数量保持一致。但是你使用了 32 个过滤器,输出层为 28×28×32,
这就是压缩通道数(𝑛𝑐)的方法,对于池化层我只是压缩了这些层的高度和宽度。
边栏推荐
- November 9, 2020 [wgs/gwas] - whole genome analysis (association analysis) process (Part 2)
- STM32 control LED lamp
- Final review -php learning notes 6- string processing
- Installation software operation manual (continuous update)
- ADC basic concepts
- November 21, 2021 [reading notes] - bioinformatics and functional genomics (Chapter 5 advanced database search)
- Graphic explanation pads update PCB design basic operation
- Minecraft 1.16.5 module development (50) guide book
- January 23, 2022 [reading notes] - bioinformatics and functional genomics (Chapter 6: multiple sequence alignment)
- DS1302 digital tube clock
猜你喜欢

Use of ecostruxure (3) creating composite function blocks
![Experiment 1: comprehensive experiment [process on]](/img/19/6c6e18d7e1f042bfd3ee4832b78542.png)
Experiment 1: comprehensive experiment [process on]

Deloitte: investment management industry outlook in 2022

2022 Research Report on China's intelligent fiscal and tax Market: accurate positioning, integration and diversity

Analysys analysis: online audio content consumption market analysis 2022

Efga design open source framework openlane series (I) development environment construction

Tencent and Fudan University "2021-2022 yuan universe report" with 102 yuan universe collections

期末复习-PHP学习笔记9-PHP会话控制

Investment and financing analysis report of Supply Chain & logistics industry in 2021
![2021-10-29 [microbiology] a complete set of 16s/its analysis process based on qiime2 tool (Part I)](/img/9d/37c531b1b439770f69f715687685f5.jpg)
2021-10-29 [microbiology] a complete set of 16s/its analysis process based on qiime2 tool (Part I)
随机推荐
Video player (II): video decoding
342 maps covering exquisite knowledge, one of which is classic and pasted on the wall
Periodic planning work
November 22, 2021 [reading notes] - bioinformatics and functional genomics (Section 5 of Chapter 5 uses a comparison tool similar to blast to quickly search genomic DNA)
RT thread kernel application development message queue experiment
Final review -php learning notes 3-php process control statement
Mailbox application routine of running wild fire RT thread
Graphic explanation pads update PCB design basic operation
Efga design open source framework openlane series (I) development environment construction
How to batch modify packaging for DXP schematic diagram
November 21, 2021 [reading notes] - bioinformatics and functional genomics (Chapter 5 advanced database search)
Application of stack -- using stack to realize bracket matching (C language implementation)
2021-10-29 [microbiology] qiime2 sample pretreatment form automation script
National technology n32g45x series about timer timing cycle calculation
Final review -php learning notes 5-php array
Analysis of cross clock transmission in tinyriscv
Solve the linear equation of a specified point and a specified direction
实验一、综合实验【Process on】
Efga design open source framework fabulous series (I) establishment of development environment
Distance from point to line


