怎么画高大上的神经网络结构？试试这个！- 专知

会员服务 ·

怎么画高大上的神经网络结构？试试这个！

专知出品

【导读】现在这年头，做深度学习的，如果不在自己的文章里画一个神经网络结构图，都不好意思出门跟人说话。今天小编给大家介绍一个工具，让你简单又快速的搞定神经网络结构图：PlotNeuralNet。

Github 地址：

https://github.com/HarisIqbal88/PlotNeuralNet

编辑报道：

专知-huaiwen

「工具简介」

PlotNeuralNet工具，具如其名，plot neural net用的，首先我们看看效果：

FCN-8

VGG-16

Holistically-Nested Edge Detection

「安装使用」

PlotNeuralNet的使用方法非常简单，首先将这个项目克隆下来：

git clone https://github.com/HarisIqbal88/PlotNeuralNet

然后，你可以自己写一个python脚本，想要用什么结构，就从PlotNerualNet里找对应的模块，然后，把它们拼起来就行, 好比UNet:

你只需要按部就班的堆：

import sys
sys.path.append('../')
from core.tikzeng 
import *from core.blocks  import *

arch = [ 
    to_head('..'), 
    to_cor(),
    to_begin(),    
    #input
    to_input( '../examples/fcn8s/cats.jpg' ),    

    #block-001
    to_ConvConvRelu( name='ccr_b1', s_filer=500, n_filer=(64,64), offset="(0,0,0)", to="(0,0,0)", width=(2,2), height=40, depth=40  ),
    to_Pool(name="pool_b1", offset="(0,0,0)", to="(ccr_b1-east)", width=1, height=32, depth=32, opacity=0.5),    
    
    *block_2ConvPool( name='b2', botton='pool_b1', top='pool_b2', s_filer=256, n_filer=128, offset="(1,0,0)", size=(32,32,3.5), opacity=0.5 ),    
    *block_2ConvPool( name='b3', botton='pool_b2', top='pool_b3', s_filer=128, n_filer=256, offset="(1,0,0)", size=(25,25,4.5), opacity=0.5 ),    
    *block_2ConvPool( name='b4', botton='pool_b3', top='pool_b4', s_filer=64,  n_filer=512, offset="(1,0,0)", size=(16,16,5.5), opacity=0.5 ),    
    
    #Bottleneck
    #block-005
    to_ConvConvRelu( name='ccr_b5', s_filer=32, n_filer=(1024,1024), offset="(2,0,0)", to="(pool_b4-east)", width=(8,8), height=8, depth=8, caption="Bottleneck"  ),
    to_connection( "pool_b4", "ccr_b5"),    
    
    #Decoder
    *block_Unconv( name="b6", botton="ccr_b5", top='end_b6', s_filer=64,  n_filer=512, offset="(2.1,0,0)", size=(16,16,5.0), opacity=0.5 ),
    to_skip( of='ccr_b4', to='ccr_res_b6', pos=1.25),    
    *block_Unconv( name="b7", botton="end_b6", top='end_b7', s_filer=128, n_filer=256, offset="(2.1,0,0)", size=(25,25,4.5), opacity=0.5 ),
    to_skip( of='ccr_b3', to='ccr_res_b7', pos=1.25),    
    *block_Unconv( name="b8", botton="end_b7", top='end_b8', s_filer=256, n_filer=128, offset="(2.1,0,0)", size=(32,32,3.5), opacity=0.5 ),
    to_skip( of='ccr_b2', to='ccr_res_b8', pos=1.25),    
    
    *block_Unconv( name="b9", botton="end_b8", top='end_b9', s_filer=512, n_filer=64,  offset="(2.1,0,0)", size=(40,40,2.5), opacity=0.5 ),
    to_skip( of='ccr_b1', to='ccr_res_b9', pos=1.25),
    
    to_ConvSoftMax( name="soft1", s_filer=512, offset="(0.75,0,0)", to="(end_b9-east)", width=1, height=40, depth=40, caption="SOFT" ),
    to_connection( "end_b9", "soft1"),
     
    to_end() 
    ]

def main():
    namefile = str(sys.argv[0]).split('.')[0]
    to_generate(arch, namefile + '.tex' )
if __name__ == '__main__':
    main()