Package autodiffr provides an R wrapper for Julia packages ForwardDiff.jl and ReverseDiff.jl through JuliaCall to do automatic differentiation for native R functions.
Installation
Julia is needed to use autodiffr. You can download a generic Julia binary from https://julialang.org/downloads/ and add it to the path. Pakcage autodiffr is not on CRAN yet. You can get the development version of autodiffr by
Important: Note that currently Julia v0.6.x, v0.7.0 and v1.0 are all supported by autodiffr, but to use autodiffr with Julia v0.7/1.0, you need to get the development version of JuliaCall by:
Basic Usage
library(autodiffr)
## Do initial setup
ad_setup()
#> Julia version 1.0.1 at location /Applications/Julia-1.0.app/Contents/Resources/julia/bin will be used.
#> Loading setup script for JuliaCall...
#> Finish loading setup script for JuliaCall.
#> Loading ReverseDiff...
#> Loading ForwardDiff...
#> Finish autodiffr setup.
## If you want to use a julia at a specific location, you could do the following:
## ad_setup(JULIA_HOME = "the folder that contains julia binary"),
## or you can set JULIA_HOME in command line environment or use `options(...)`
## Define a target function with vector input and scalar output
f <- function(x) sum(x^2L)
## Calculate gradient of f at [2,3] by
ad_grad(f, c(2, 3)) ## deriv(f, c(2, 3))
#> [1] 4 6
## Get a gradient function g
g <- makeGradFunc(f)
## Evaluate the gradient function g at [2,3]
g(c(2, 3))
#> [1] 4 6
## Calculate hessian of f at [2,3] by
ad_hessian(f, c(2, 3))
#> [,1] [,2]
#> [1,] 2 0
#> [2,] 0 2
## Get a hessian function h
h <- makeHessianFunc(f)
## Evaluate the hessian function h at [2,3]
h(c(2, 3))
#> [,1] [,2]
#> [1,] 2 0
#> [2,] 0 2
## Define a target function with vector input and vector output
f <- function(x) x^2
## Calculate jacobian of f at [2,3] by
ad_jacobian(f, c(2, 3))
#> [,1] [,2]
#> [1,] 4 0
#> [2,] 0 6
## Get a jacobian function j
j <- makeJacobianFunc(f)
## Evaluate the gradient function j at [2,3]
j(c(2, 3))
#> [,1] [,2]
#> [1,] 4 0
#> [2,] 0 6Advanced Usage
Functions with Multiple Arguments
## Define a target function with mulitple arguments
f <- function(a = 1, b = 2, c = 3) a * b ^ 2 * c ^ 3
## Calculate gradient/derivative of f at a = 2, when b = c = 1 by
ad_grad(f, 2, b = 1, c = 1) ## deriv(f, 2, b = 1, c = 1)
#> [1] 1
## Get a gradient/derivative function g w.r.t a when b = c = 1 by
g <- makeGradFunc(f, b = 1, c = 1)
## Evaluate the gradient/derivative function g at a = 2
g(2)
#> [1] 1
## Calculate gradient/derivative of f at a = 2, b = 3, when c = 1 by
ad_grad(f, list(a = 2, b = 3), c = 1)
#> $a
#> [1] 9
#>
#> $b
#> [1] 12
## Get a gradient/derivative function g w.r.t a and b when c = 1 by
g <- makeGradFunc(f, c = 1)
## Evaluate the gradient/derivative function g at a = 2, b = 3
g(list(a = 2, b = 3))
#> $a
#> [1] 9
#>
#> $b
#> [1] 12Trouble Shooting and Way to Get Help
Julia is not found
Make sure the Julia installation is correct. autodiffr is able to find Julia on PATH, and there are three ways for autodiffr to find Julia not on PATH.
- Use
ad_setup(JULIA_HOME = "the folder that contains julia binary") - Use
options(JULIA_HOME = "the folder that contains julia binary") - Set
JULIA_HOMEin command line environment.
How to Get Help
The GitHub Pages for this repository host the documentation for the development version of
autodiffr: https://non-contradiction.github.io/autodiffr/.And you are more than welcome to contact me about
autodiffrat lch34677@gmail.com or cxl508@psu.edu.
Suggestion and Issue Reporting
autodiffr is under active development now. Any suggestion or issue reporting is welcome! You may report it using the link: https://github.com/Non-Contradiction/autodiffr/issues/new. Or email me at lch34677@gmail.com or cxl508@psu.edu.
