打印syms / matlabFunction慢

时间：2015-10-16 16:06:32

标签： matlab symbolic-math mupad

我在尝试使符号替换变得更快时遇到了很多麻烦 - 也就是说，在符号表达式中替换变量并获得双精度。

我正在创建一个复杂的函数f，并计算其jacobian df。这是合理的速度，我可以将它保存到文件中。但是当我尝试使用matlabFunction甚至是disp或fprintf时，系统会挂起并无法继续进行（即使matlabFunction设置为未优化）。这是一个主要问题，因为我需要能够合理地快速替换。

f矢量是24个元素，Jacobian是24 x 78（虽然这里只显示了70个变量，所以这可以压缩到70列;但我怀疑这是问题）。

我也知道f和df的某些元素很简单，单独访问时工作正常，但f和df的某些更复杂的元素无法显示。我想它们很长，但是因为它们计算得很好，所以我不能将它们转换成matlab函数或显示它们。

更奇怪的是，我可以替换所有符号变量，但是完全替换f矢量的最终显示（例如，通过disp），或者转换为double（通过double（））似乎需要永远

如果你想使用.mat文件，你可以得到它here（filedropper链接，它是288kb）。我该怎么做才能在合理的时间内写出这个文件？

1 个答案:

答案 0 :(得分：7)

注意：从评论的角度来看，我专注于这个问题：

  

这是矩阵乘法和加法的结果。我已经尝试使用simplify命令简化它，但是速度非常慢。我已经成功地将f写入未经优化的文件（没有运气df） - 花了2个小时 - 但随后评估它需要0.8秒，这太慢了。我需要能够在大约0.02秒内完成评估。

我开始查看f中的元素，并且f(12)很简单。但是，f(13)释放了地狱：

>> inp.f(13)

ans =

(2289*l4)/100 - (11371197146449238679*l3)/8112963841460668169578900514406400 - (2289*l2)/100 + (11371197146449238679*l5)/8112963841460668169578900514406400 - (2289*l8)/100 - (11371197146449238679*l9)/8112963841460668169578900514406400 + (2289*l10)/100 + (11371197146449238679*l11)/8112963841460668169578900514406400 - (2289*l14)/100 - (11371197146449238679*l15)/8112963841460668169578900514406400 + (2289*l16)/100 + (11371197146449238679*l17)/8112963841460668169578900514406400 - (2289*l20)/100 - (11371197146449238679*l21)/8112963841460668169578900514406400 + (2289*l22)/100 + (11371197146449238679*l23)/8112963841460668169578900514406400 - (2289*l26)/100 - (11371197146449238679*l27)/8112963841460668169578900514406400 + (2289*l28)/100 + (11371197146449238679*l29)/8112963841460668169578900514406400 - (2289*l32)/100 - (11371197146449238679*l33)/8112963841460668169578900514406400 + (2289*l34)/100 + (11371197146449238679*l35)/8112963841460668169578900514406400 - h1*(((cos(x4/2)*cos(x6/2) + sin(x4/2)*sin(x5/2)*sin(x6/2))*(cos(x4/2)*sin(x6/2) - cos(x6/2)*sin(x4/2)*sin(x5/2)) + (sin(x4/2)*sin(x6/2) + cos(x4/2)*cos(x6/2)*sin(x5/2))*(cos(x6/2)*sin(x4/2) - cos(x4/2)*sin(x5/2)*sin(x6/2)) - cos(x5/2)^2*cos(x6/2)*sin(x6/2))*(((x17*(cos(x4/2)*cos(x5/2)*(cos(x6/2)*(cos(x6/2)*sin(x4/2) - cos(x4/2)*sin(x5/2)*sin(x6/2)) + sin(x6/2)*(sin(x4/2)*sin(x6/2) + cos(x4/2)*cos(x6/2)*sin(x5/2))) - cos(x5/2)*sin(x4/2)*(cos(x6/2)*(cos(x4/2)*cos(x6/2) + sin(x4/2)*sin(x5/2)*sin(x6/2)) + sin(x6/2)*(cos(x4/2)*sin(x6/2) - cos(x6/2)*sin(x4/2)*sin(x5/2)))))/2 - (x18*(cos(x4/2)^2*cos(x5/2)^2 + cos(x5/2)^2*sin(x4/2)^2 + sin(x5/2)^2))/2 + (x16*(sin(x5/2)*(cos(x5/2)^2*cos(x6/2)^2 + cos(x5/2)^2*sin(x6/2)^2 + sin(x5/2)^2) + cos(x5/2)*sin(x4/2)*(cos(x5/2)*sin(x4/2)*sin(x5/2) + cos(x5/2)*cos(x6/2)*(cos(x4/2)*sin(x6/2) - cos(x6/2)*sin(x4/2)*sin(x5/2)) - cos(x5/2)*sin(x6/2)*(cos(x4/2)*cos(x6/2) + sin(x4/2)*sin(x5/2)*sin(x6/2))) + cos(x4/2)*cos(x5/2)*(cos(x4/2)*cos(x5/2)*sin(x5/2) - cos(x5/2)*cos(x6/2)*(sin(x4/2)*sin(x6/2) + cos(x4/2)*cos(x6/2)*sin(x5/2)) + cos(x5/2)*sin(x6/2)*(cos(x6/2)*sin(x4/2) - cos(x4/2)*sin(x5/2)*sin(x6/2)))))/2 - (x19*cos(x5/2)*sin(x4/2))/2)*((LEG_MASS*((cos(x7/2)*(sin(x4/2)*sin(x6/2) + cos(x4/2)*cos(x6/2)*sin(x5/2)) + cos(x5/2)*cos(x6/2)*sin(x7/2))*((cos(x4/2)*cos(x6/2) + sin(x4/2)*sin(x5/2)*sin(x6/2))*(x2/2 - BASE_ORIGIN_Z*(cos(x6/2)*sin(x4/2) - cos(x4/2)*sin(x5/2)*sin(x6/2)) - (cos(x4/2)*cos(x6/2) + sin(x4/2)*sin(x5/2)*sin(x6/2))*(BASE_LINK_EXTENTS_Y/2 - BASE_ORIGIN_Y + LEG_LINK_EXTENTS_Y/2) + cos(x5/2)*sin(x6/2)*(BASE_ORIGIN_X - BASE_LINK_EXTENTS_X/4)) - (cos(x4/2)*sin(x6/2) - cos(x6/2)*sin(x4/2)*sin(x5/2))*(x1/2 + BASE_ORIGIN_Z*(sin(x4/2)*sin(x6/2) + cos(x4/2)*cos(x6/2)*sin(x5/2)) + (cos(x4/2)*sin(x6/2) - cos(x6/2)*sin(x4/2)*sin(x5/2))*(BASE_LINK_EXTENTS_Y/2 - BASE_ORIGIN_Y + LEG_LINK_EXTENTS_Y/2) + cos(x5/2)*cos(x6/2)*(BASE_ORIGIN_X - BASE_LINK_EXTENTS_X/4)) + cos(x5/2)*sin(x4/2)*(x3/2 - sin(x5/2)*(BASE_ORIGIN_X - BASE_LINK_EXTENTS_X/4) + BASE_ORIGIN_Z*cos(x4/2)*cos(x5/2) - cos(x5/2)*sin(x4/2)*(BASE_LINK_EXTENTS_Y/2 - BASE_ORIGIN_Y + LEG_LINK_EXTENTS_Y/2))) - (cos(x4/2)*sin(x6/2) - cos(x6/2)*sin(x4/2)*sin(x5/2))*((sin(x5/2)*sin(x7/2) - cos(x4/2)*cos(x5/2)*cos(x7/2))*(x3/2 - sin(x5/2)*(BASE_ORIGIN_X - BASE_LINK_EXTENTS_X/4) + BASE_ORIGIN_Z*cos(x4/2)*cos(x5/2) - cos(x5/2)*sin(x4/2)*(BASE_LINK_EXTENTS_Y/2 - BASE_ORIGIN_Y + LEG_LINK_EXTENTS_Y/2)) - (cos(x7/2)*(sin(x4/2)*sin(x6/2) + cos(x4/2)*cos(x6/2)*sin(x5/2)) + cos(x5/2)*cos(x6/2)*sin(x7/2))*(x1/2 + BASE_ORIGIN_Z*(sin(x4/2)*sin(x6/2) + cos(x4/2)*cos(x6/2)*sin(x5/2)) + (cos(x4/2)*sin(x6/2) - cos(x6/2)*sin(x4/2)*sin(x5/2))*(BASE_LINK_EXTENTS_Y/2 - BASE_ORIGIN_Y + LEG_LINK_EXTENTS_Y/2) + cos(x5/2)*cos(x6/2)*(BASE_ORIGIN_X - BASE_LINK_EXTENTS_X/4)) + (cos(x7/2)*(cos(x6/2)*sin(x4/2) - cos(x4/2)*sin(x5/2)*sin(x6/2)) - cos(x5/2)*sin(x6/2)*sin(x7/2))*(x2/2 - BASE_ORIGIN_Z*(cos(x6/2)*sin(x4/2) - cos(x4/2)*sin(x5/2)*sin(x6/2)) - (cos(x4/2)*cos(x6/2) + sin(x4/2)*sin(x5/2)*sin(x6/2))*(BASE_LINK_EXTENTS_Y/2 - BASE_ORIGIN_Y + LEG_LINK_EXTENTS_Y/2) + cos(x5/2)*sin(x6/2)*(BASE_ORIGIN_X - BASE_LINK_EXTENTS_X/4))))*(sin(x7/2)*(sin(x4/2)*sin(x6/2) + cos(x4/2)*cos(x6/2)*sin(x5/2)) - cos(x5/2)*cos(x6/2)*cos(x7/2)) - LEG_MASS*((sin(x7/2)*(sin(x4/2)*sin(x6/2) + cos(x4/2)*cos(x6/2)*sin(x5/2)) - cos(x5/2)*cos(x6/2)*cos(x7/2))*((cos(x4/2)*cos(x6/2) + sin(x4/2)*sin(x5/2)*sin(x6/2))*(x2/2 - BASE_ORIGIN_Z*(cos(x6/2)*sin(x4/2) - cos(x4/2)*sin(x5/2)*sin(x6/2)) - (cos(x4/2)*cos(x6/2) + sin(x4/2)*sin(x5/2)*sin(x6/2))*(BASE_LINK_EXTENTS_Y/2 - BASE_ORIGIN_Y + LEG_LINK_EXTENTS_Y/2) + cos(x5/2)*sin(x6/2)*(BASE_ORIGIN_X - BASE_LINK_EXTENTS_X/4)) - (cos(x4/2)*sin(x6/2) - cos(x6/2)*sin(x4/2)*sin(x5/2))*(x1/2 + BASE_ORIGIN_Z*(sin(x4/2)*sin(x6/2) + cos(x4/2)*cos(x6/2)*sin(x5/2)) + (cos(x4/2)*sin(x6/2) - cos(x6/2)*sin(x4/2)*sin(x5/2))*(BASE_LINK_EXTENTS_Y/2 - BASE_ORIGIN_Y + LEG_LINK_EXTENTS_Y/2) + cos(x5/2)*cos(x6/2)*(BASE_ORIGIN_X - BASE_LINK_EXTENTS_X/4)) + cos(x5/2)*sin(x4/2)*(x3/2 - sin(x5/2)*(BASE_ORIGIN_X - BASE_LINK_EXTENTS_X/4) + BASE_ORIGIN_Z*cos(x4/2)*cos(x5/2) - cos(x5/2)*sin(x4/2)*(BASE_LINK_EXTENTS_Y/2 - BASE_ORIGIN_Y + LEG_LINK_EXTENTS_Y/2))) + (cos(x4/2)*sin(x6/2) - cos(x6/2)*sin(x4/2)*sin(x5/2))*((cos(x7/2)*sin(x5/2) + cos(x4/2)*cos(x5/2)*sin(x7/2))*(x3/2 - sin(x5/2)*(BASE_ORIGIN_X - BASE_LINK_EXTENTS_X/4) + BASE_ORIGIN_Z*cos(x4/2)*cos(x5/2) - cos(x5/2)*sin(x4/2)*(BASE_LINK_EXTENTS_Y/2 - BASE_ORIGIN_Y + LEG_LINK_EXTENTS_Y/2)) + (sin(x7/2)*(sin(x4/2)*sin(x6/2) + cos(x4/2)*cos(x6/2)*sin(x5/2)) - cos(x5/2)*cos(x6/2)*cos(x7/2))*(x1/2 + BASE_ORIGIN_Z*(sin(x4/2)*sin(x6/2) + cos(x4/2)*cos(x6/2)*sin(x5/2)) + (cos(x4/2)*sin(x6/2) - cos(x6/2)*sin(x4/2)*sin(x5/2))*(BASE_LINK_EXTENTS_Y/2 - BASE_ORIGIN_Y + LEG_LINK_EXTENTS_Y/2) + cos(x5/2)*cos(x6/2)*(BASE_ORIGIN_X - BASE_LINK_EXTENTS_X/4)) - (sin(x7/2)*(cos(x6/2)*sin(x4/2) - cos(x4/2)*sin(x5/2)*sin(x6/2)) + cos(x5/2)*cos(x7/2)*sin(x6/2))*(x2/2 - BASE_ORIGIN_Z*(cos(x6/2)*sin(x4/2) - cos(x4/2)*sin(x5/2)*sin(x6/2)) - (cos(x4/2)*cos(x6/2) + sin(x4/2)*sin(x5/2)*sin(x6/2))*(BASE_LINK_EXTENTS_Y/2 - BASE_ORIGIN_Y + LEG_LINK_EXTENTS_Y/2) + cos(x5/2)*sin(x6/2)*(BASE_ORIGIN_X - BASE_LINK_EXTENTS_X/4))))*(cos(x7/2)*(sin(x4/2)*sin(x6/2) + cos(x4/2)*cos(x6/2)*sin(x5/2)) + cos(x5/2)*cos(x6/2)*sin(x7/2)) + LEG_MASS*(cos(x4/2)*sin(x6/2) - cos(x6/2)*sin(x4/2)*sin(x5/2))*((cos(x7/2)*(sin(x4/2)*sin(x6/2) + cos(x4/2)*cos(x6/2)*sin(x5/2)) + cos(x5/2)*cos(x6/2)*sin(x7/2))*((cos(x7/2)*sin(x5/2) + cos(x4/2)*cos(x5/2)*sin(x7/2))*(x3/2 - sin(x5/2)*(BASE_ORIGIN_X - BASE_LINK_EXTENTS_X/4) + BASE_ORIGIN_Z*cos(x4/2)*cos(x5/2) - cos(x5/2)*sin(x4/2)*(BASE_LINK_EXTENTS_Y/2 - BASE_ORIGIN_Y + LEG_LINK_EXTENTS_Y/2)) + (sin(x7/2)*(sin(x4/2)*sin(x6/2) + cos(x4/2)*cos(x6/2)*sin(x5/2)) - cos(x5/2)*cos(x6/2)*cos(x7/2))*(x1/2 + BASE_ORIGIN_Z*(sin(x4/2)*sin(x6/2) + cos(x4/2)*cos(x6/2)*sin(x5/2)) + (cos(x4/2)*sin(x6/2) - cos(x6/2)*sin(x4/2)*sin(x5/2))*(BASE_LINK_EXTENTS_Y/2 - BASE_ORIGIN_Y + LEG_LINK_EXTENTS_Y/2) + cos(x5/2)*cos(x6/2)*(BASE_ORIGIN_X - BASE_LINK_EXTENTS_X/4)) - (sin(x7/2)*(cos(x6/2)*sin(x4/2) - cos(x4/2)*sin(x5/2)*sin(x6/2)) + cos(x5/2)*cos(x7/2)*sin(x6/2))*(x2/2 - BASE_ORIGIN_Z*(cos(x6/2)*sin(x4/2) - cos(x4/2)*sin(x5/2)*sin(x6/2)) - (cos(x4/2)*cos(x6/2) + sin(x4/2)*sin(x5/2)*sin(x6/2))*(BASE_LINK_EXTENTS_Y/2 - BASE_ORIGIN_Y + LEG_LINK_EXTENTS_Y/2) + [...] ... Output truncated.  Text exceeds maximum line length of 25,000 characters for Command Window display.

（实际输出被截断以适应SO的30k字符限制，但是你得到了交易）

我打赌matlabFunction的解析器不适合这种程度的输入。那里还有一些奇怪的东西：像8e33那样的整数字符串文字。

所以我仔细看了看你的功能。幸运的是，你可以将你的函数转换为字符串，然后处理那些只占用CPU时间而不是内存的字符串。

Preproc：

for k=1:24
   fstring2{k}=char(inp.f(k));
end

功能长度：

>> cellfun(@length,fstring2)

ans =

  Columns 1 through 12

          11          11          11          11          11          11          11          11          11          11          11          11

  Columns 13 through 24

     2301006     2300241     2299996     8425640     8416273     8424306     1375443     1305245     1302440     1237876     1381084     1310884

休斯顿，我们遇到了问题。

这些巨大的符号函数攻击破坏了matlabFunction的解析器，或者更可能的是，在操作期间你的内存耗尽。我确实在尝试simplify f(13)时，在几秒钟内失去了8 GB的更好的一半。

正如概念验证一样，我试图嘲笑你的功能所涉及的计算工作量。我检查了f(13)（第一个野兽）。有关所涉及的操作的一些信息：

>> length(strfind(char(inp.f(13)),'*'))

ans =

      134710

>> length(strfind(char(inp.f(13)),'+'))

ans =

       36932

>> length(strfind(char(inp.f(13)),'-'))

ans =

       26855

>> length(strfind(char(inp.f(13)),'/'))

ans =

      183380

>> length(strfind(char(inp.f(13)),'ln'))

ans =

     0

>> length(strfind(char(inp.f(13)),'exp'))

ans =

     0

>> length(strfind(char(inp.f(13)),'cos'))

ans =

       78700

>> length(strfind(char(inp.f(13)),'sin'))

ans =

       84142

我试图计算涉及相似数量操作的模拟计算：

x=zeros(36000,1);
tic;
for k=1:36000
   x(k)=(((sin(sin(((cos(cos(3.1+2.1)*3.1)*6.1)*5.1)*9.1)/4.1)/3.1)/6.1)/5.1)/8.1;
end
toc;

Elapsed time is 0.010895 seconds.

这包括36000次增加，144000次增殖，180000分割和72000对sin和cos的调用。

现在，如果我们假设这是一个正确的大概数字，如果我们假设你的函数有类似的操作分布，那么你正在查看函数的40080434个字符，这是17个等价的{ {1}}单位。这表明即使您可以转换为正确的matlab函数，您的运行时只需调用f(13)（我们根本没有查看f）将采取至少0.1-0.2秒。

由于您的问题的性质，我不确定是否有办法绕过它。我可能会尝试使用df中的sympy来做同样的事情，在那里你也可以转换为python（等效于匿名函数的python），用于数值计算。如果 会成功，那么至少你可以尽快使用你的功能。

更新

在发布不太乐观的答案之后，我相信我已成功将您的功能转换为匿名功能。它很脏，但似乎有效。

首先，如上所述将函数转换为字符串，然后使用lambda提取变量名称。然后使用这些函数名创建函数定义;不幸的是，我只能使用symvar来破解它。 应该是一种更优雅的方式，但无论如何我们对可实现的运行时感兴趣。

eval

转换是讨厌的，但匿名函数的实际构造很快，而且内存不太难。虚拟运行时：

varcell=symvar(fstring2{13}); %variables of inp.f(13)
vars2=strcat(varcell,','); %add a comma to each var
vars3=[vars2{:}]; %put them into a single string
vars3=vars3(1:end-1); %remove trailing comma

f13=eval(['@(' v3 ') ' fstring2{13}]); %this is your numeric function

它可以更加用户友好，例如允许在函数中进行数组操作，或者将所有58个输入作为单个数组输入传递。但是你的运行时会是一样的。这只是一个功能，你大约有17个功能。你可能永远不会得到你希望的加速。

（无论如何，我确实开始了

  

线程“AWT-EventQueue-0”中的异常java.lang.OutOfMemoryError：Java堆空间

在整个考验之后出现

错误，所以成功也可能取决于你对“成功”的定义;）

相关问题
最新问题