Depth study found a drawing tool from the public summary of number, it is practical: https://mp.weixin.qq.com/s/MMzvZA55Xb2sOA7rJiXiEw
1 NN-SVG
This tool can be very convenient to draw various types of diagram, below is the little brother of the development, from the Massachusetts Institute of Technology Frankel biological engineering lab that develops visualization and machine learning tools for the analysis of biological data .
-
github地址：https://github.com/zfrenchee -
画图工具体验地址：http://alexlenail.me/NN-SVG/
Can be plotted in the form of nodes comprises FCNN style shown in this drawing is particularly suitable for conventional fully connected neural network.
LeNet style tile to the network structure shown, two-dimensional manner, suitable for viewing the size and number of channels of each layer featuremap.
Show a three-dimensional block form AlexNet style, can be more realistic picture of the scale of change in high-dimensional data convolution process, currently only supports full convolution layer and connection layer.
This tool can be exported very definition of SVG drawing, worth experiencing.
2 PlotNeuralNet
This tool is a computer science student Saarland University developed a computer look like a college thing.
First we look at the results, it github link below, nearly 4000 star:
https://github.com/HarisIqbal88/PlotNeuralNetPeople look at this visual map fcn-8, the high value of the odd color.
The threshold used relatively speaking, higher, and with the language LaTex editor, so you can play on the big room, you see the following softmax layer, which is going to write the code of the advantages.
One part of the code is that, write it.
-
\pic[shift={(0,0,0)}] at (0,0,0) {Box={name=crp1,caption=SoftmaxLoss: $E_\mathcal{S}$ ,% -
fill={rgb:blue,1.5;red,3.5;green,3.5;white,5},opacity=0.5,height=20,width=7,depth=20}};
There is a similar tool: https: //github.com/jettan/tikz_cnn
3 ConvNetDraw
ConvNetDraw is configured to use a neural network CNN command of drawing tools, the developer is a programmer in Hong Kong, Cédric cbovar.
采用如下的语法直接配置网络，可以简单调整x，y，z等3个维度，github链接如下：
https://cbovar.github.io/ConvNetDraw/
使用方法如上图所示，只需输入模型结构中各层的参数配置。
挺好用的不过它目标分辨率太低了，放大之后不清晰，达不到印刷的需求。
4 Draw_Convnet
这一个工具名叫draw_convnet，由Borealis公司的员工Gavin Weiguang Ding提供。
简单直接，是纯用python代码画图的，
https://github.com/gwding/draw_convnet看看画的图如下，核心工具是matplotlib，图不酷炫，但是好在规规矩矩，可以严格控制，论文用挺合适的。
类似的工具还有：https://github.com/yu4u/convnet-drawer
5 Netscope
下面要说的是这个，我最常用的，caffe的网络结构可视化工具，大名鼎鼎的netscope，由斯坦福AILab的Saumitro Dasgupta开发，找不到照片就不放了，地址如下：
https://github.com/ethereon/netscope
左边放配置文件，右边出图，非常方便进行网络参数的调整和可视化。这种方式好就好在各个网络层之间的连接非常的方便。
其他
再分享一个有意思的，不是画什么正经图，但是把权重都画出来了。
http://scs.ryerson.ca/~aharley/vis/conv/
看了这么多，有人已经在偷偷笑了，上PPT呀，想要什么有什么，想怎么画就怎么画
draw.io
还要别人推荐的这个