Julia符号微分

huangapple go评论54阅读模式
英文:

Julia symbolic differentiation

问题

尽管我查阅了几页关于Julia中各种微分工具的文档,但我仍然没有找到以下简单功能。我想定义一个可以手动微分的函数,例如,f(x) = 2x.^3,然后调用一个符号计算函数来获得fprime = derivative(f)。这段代码应该产生与fprime(x) = 6x.^2相同的结果。是否有一个类似于上面虚构的derivative函数的库函数?

英文:

Though I've looked through several pages documenting various differentiation tools in Julia, I have yet to find the following simple functionality. I want to define a function which is differentiable by hand, e.g, f(x) = 2x.^3, then call a symbolic computation function to obtain fprime = derivative(f). This code should produce the same results as fprime(x) = 6x.^2. Is there a library function which acts like the made-up function derivative above?

答案1

得分: 2

Symbolics.jl包能够实现你想要的功能,尽管你需要明确声明你的变量。

In [2]: using Symbolics

In [3]: @variables x
Out[3]: 1-element Vector{Num}:
 x

In [4]: D = Differential(x)
Out[4]: (::Differential) (generic function with 2 methods)

In [5]: f(t) = 2*t^3
Out[5]: f (generic function with 1 method)

In [6]: expand_derivatives(D(f(x)))
Out[6]: 6(x^2)

你可以在这里了解更多信息:https://symbolics.juliasymbolics.org/dev/manual/derivatives/

英文:

The Symbolics.jl package is capable of what you're looking for, although you need to explicitly declare what your variables are

In [2]: using Symbolics

In [3]: @variables x
Out[3]: 1-element Vector{Num}:
 x

In [4]: D = Differential(x)
Out[4]: (::Differential) (generic function with 2 methods)

In [5]: f(t) = 2*t^3
Out[5]: f (generic function with 1 method)

In [6]: expand_derivatives(D(f(x)))
Out[6]: 6(x^2)

You can read more about it here: https://symbolics.juliasymbolics.org/dev/manual/derivatives/

答案2

得分: 1

If you work numerically with symbolic derivatives you can also use code-to-code symbolic differentiation by using Zygote

Assume you have f(x) = 2x^3 than having Zygote loaded you can do

julia> f'(5)
150.0

to understand what just happened peek into the compilation process to see that the function f has been actually symbolically differentiated.

julia> @code_llvm f'(5)
define double @"julia_#79_991"(i64 signext %0) #0 {
  top:
       %1 = mul i64 %0, 3
       %2 = mul i64 %1, %0
       %3 = sitofp i64 %2 to double
       %4 = fmul double %3, 2.000000e+00
  ret double %4
}
英文:

If you work numerically with symbolic derivatives you can also use code-to-code symbolic differentiation by using Zygote

Assume you have f(x) = 2x^3 than having Zygote loaded you can do

julia> f'(5)
150.0

to understand what just happened peek into the compilation process to see that the function f has been actually symbolically differentiated.

julia> @code_llvm f'(5)
define double @"julia_#79_991"(i64 signext %0) #0 {
  top:
       %1 = mul i64 %0, 3
       %2 = mul i64 %1, %0
       %3 = sitofp i64 %2 to double
       %4 = fmul double %3, 2.000000e+00
  ret double %4
}

答案3

得分: 1

以下是翻译好的部分:

在Julia中,有各种用于自动微分(AD)的包。对于像您描述的单变量函数,最简单且最快的可能是ForwardDiff包,它允许您执行以下操作:

using ForwardDiff
fprime(x) = ForwardDiff.derivative(f, x)

ForwardDiff和其他AD包确实计算精确的解析导数,相当于您可能手动编写的6x^2。(然而,算法与您可能想象的“形成大型符号表达式然后进行微分”的方法有些不同。)

英文:

There are various packages for automatic differentiation (AD) in Julia. For a single-variable function like the one you describe, the simplest and fastest is probably the ForwardDiff package, which lets you do:

using ForwardDiff
fprime(x) = ForwardDiff.derivative(f, x)

ForwardDiff and other AD packages indeed compute the exact analytical derivative, equivalent to the 6x^2 you might write by hand. (However, the algorithms are a bit different from the "form a big symbolic expression then differentiate" that you might imagine.)

huangapple
  • 本文由 发表于 2023年5月26日 01:06:02
  • 转载请务必保留本文链接:https://go.coder-hub.com/76334723.html
匿名

发表评论

匿名网友

:?: :razz: :sad: :evil: :!: :smile: :oops: :grin: :eek: :shock: :???: :cool: :lol: :mad: :twisted: :roll: :wink: :idea: :arrow: :neutral: :cry: :mrgreen:

确定