金华专业网站建设公司,哈密建设局网站,互联网上市公司一览表,如何做视频类网站plotneuralnet和netron结合绘制模型架构图
一、plotneuralnet 本身的操作
模型结构图的可视化#xff0c;能直观展示模型的结构以及各个模块之间的关系。最近借助plotneuralnet python库#xff08;windows版#xff09;绘制了一个网络结构图#xff0c;有一些经验和心得…plotneuralnet和netron结合绘制模型架构图
一、plotneuralnet 本身的操作
模型结构图的可视化能直观展示模型的结构以及各个模块之间的关系。最近借助plotneuralnet python库windows版绘制了一个网络结构图有一些经验和心得记录在这里。 这个库里贴心地给了许多的示例用于理解代码的含义并表达自己的意思。 每一种模块都有固定的几个参数多数不难理解重点是理解其中的两个s_filer和n_filer。 如果input是5125124s_filer对应两个512代表了图形的height和depth也是参数n_filer对应4代表了图形的width。 需要注意的是两者的尺度不一样在s_filer中如果512对应40256对应32128对应25等而在n_filer中512对应的width值是7左右256对应的是6等等 这样图形看起来就是例子中“树立的扁扁的方块”。
二、获取模型各个部分的输入和输出尺寸
绘图部分有了着落之后还要获得网络各个部分的输入和输出的尺寸可以用netron这个网站和他的Python库来实现站内也有介绍。
这样数据和绘图部分都有了着落就可以实现绘图啦。
附生成文中图的代码
import sys
sys.path.append(../)
from pycore.tikzeng import *
from pycore.blocks import *
s_filer对应height和depth之间的对应关系512-40256-32128-2564-16
n_filer对应width之间的对应关系256-6128-5.564-4.532-3.516-2.5,4-1.5arch [ to_head(..), to_cor(),to_begin(),#inputto_input( demo.jpg,to(-3,0,0),nameinput_b1 ),#block-001to_ConvConvRelu( nameccr_b1, s_filer512, n_filer(16,16), offset(0,0,0), to(0,0,0), width(2.5,2.5), height40, depth40,captionStem Block),to_Pool(namepool_b1, offset(0,0,0), to(ccr_b1-east), width1, height30, depth30, opacity0.5),*block_2ConvPool( nameb2, bottonpool_b1, toppool_b2, s_filer256, n_filer32, offset(1,-8,0), size(32,32,3.5), opacity0.5 ),*block_2ConvPool( nameb3, bottonpool_b2, toppool_b3, s_filer128, n_filer64, offset(1,-8,0), size(25,25,4.5), opacity0.5 ),*block_2ConvPool( nameb4, bottonpool_b3, toppool_b4, s_filer64, n_filer128, offset(1,-8,0), size(16,16,5.5), opacity0.5 ),#Bottleneck#block-005to_Conv( nameconv1_aspp, s_filer64, n_filer256, offset(2,0,0), to(pool_b4-east), width6, height16, depth16 ),to_connection( pool_b4, conv1_aspp),to_Conv( nameconv2_aspp, s_filer64, n_filer256, offset(0,-1,0), to(conv1_aspp-east), width6, height16, depth16),to_connection( pool_b4, conv2_aspp),to_Conv( nameconv3_aspp, s_filer64, n_filer256, offset(0,-1,0), to(conv2_aspp-east), width6, height16, depth16 ),to_connection( pool_b4, conv3_aspp),to_Conv( nameconv4_aspp, s_filer64, n_filer256, offset(0,-1,0), to(conv3_aspp-east), width6, height16, depth16 ),to_connection( pool_b4,conv4_aspp),to_Sum(namesum_aspp,offset(2,3,0),to(conv4_aspp-east),radius3.5,opacity0.6),to_connection(conv1_aspp,sum_aspp),to_connection(conv2_aspp,sum_aspp),to_connection(conv3_aspp,sum_aspp),to_connection(conv4_aspp,sum_aspp),to_Conv( nameconv5_aspp, s_filer64, n_filer256, offset(2,0,0), to(sum_aspp-east), width6, height16, depth16 ),to_connection( sum_aspp, conv5_aspp),#Decoder*block_Unconv( nameb6, bottonconv5_aspp, topend_b6, s_filer64, n_filer256, offset(2.1,0,0), size(16,16,6.0), opacity0.5 ),to_skip( ofccr_b4, toccr_res_b6, pos1.25),*block_Unconv( nameb7, bottonend_b6, topend_b7, s_filer128, n_filer128, offset(2.1,8,0), size(25,25,5.5), opacity0.5 ),to_skip( ofccr_b3, toccr_res_b7, pos1.25), *block_Unconv( nameb8, bottonend_b7, topend_b8, s_filer256, n_filer64, offset(2.1,8,0), size(32,32,4.5), opacity0.5 ),to_skip( ofccr_b2, toccr_res_b8, pos1.25), *block_Unconv( nameb9, bottonend_b8, topend_b9, s_filer512, n_filer32, offset(2.1,8,0), size(40,40,3.5), opacity0.5 ),to_skip( ofccr_b1, toccr_res_b9, pos1.25),#aspp# to_ConvConvRelu( nameb10, s_filer512, n_filer(16,16), offset(2,0,0), to(end_b9-east), width(2.5,2.5), height40, depth40, captionASSP ),to_Conv( nameconv1_aspp2, s_filer512, n_filer16, offset(2,0,0), to(end_b9-east), width2.5, height40, depth40 ),to_connection( end_b9, conv1_aspp2),to_Conv( nameconv2_aspp2, s_filer512, n_filer16, offset(0,-1,0), to(conv1_aspp2-east), width2.5, height40, depth40),to_connection( end_b9, conv2_aspp2),to_Conv( nameconv3_aspp2, s_filer512, n_filer16, offset(0,-1,0), to(conv2_aspp2-east), width2.5, height40, depth40 ),to_connection( end_b9, conv3_aspp2),to_Conv( nameconv4_aspp2, s_filer512, n_filer16, offset(0,-1,0), to(conv3_aspp2-east), width2.5, height40, depth40 ),to_connection( end_b9,conv4_aspp2),to_Sum(namesum_aspp2,offset(2,3,0),to(conv4_aspp2-east),radius3.5,opacity0.6),to_connection(conv1_aspp2,sum_aspp2),to_connection(conv2_aspp2,sum_aspp2),to_connection(conv3_aspp2,sum_aspp2),to_connection(conv4_aspp2,sum_aspp2),to_Conv( nameconv5_aspp2, s_filer512, n_filer16, offset(2,0,0), to(sum_aspp2-east), width2.5, height40, depth40 ),to_connection( sum_aspp2, conv5_aspp2),to_Conv( nameb10, s_filer512, n_filer16, offset(2,0,0), to(conv5_aspp2-east), width2.5, height40, depth40 ),to_connection( conv5_aspp2, b10),to_Conv( nameb11, s_filer512, n_filer4, offset(2,0,0), to(b10-east), width1.5, height40, depth40 ),to_connection( b10, b11), to_ConvSoftMax( namesoft1, s_filer512, offset(0.75,0,0), to(b11-east), width1, height40, depth40, captionSOFT ),to_connection( b11, soft1),to_input( mask.png,to(64,0,0),nameoutput ),to_end() ]def main():namefile str(sys.argv[0]).split(.)[0]to_generate(arch, namefile .tex )if __name__ __main__:main()
