zeta1999 / torchoptim

A bit like 'stats::optim', but with torch

torchoptim

The goal of torchoptim is to experiment with building a stats::optim like function powered by the torch optimizers.

Experimental, just for fun and does not support constraint optimization. If this is a good idea remains to be seen :)

Installation

remotes::install_github("dirkschumacher/torchoptim")

Example

As an example we optimize the Rosenbrock function and compare it to the solution from stats::optim.

library(torch)
library(torchoptim)

# from the R docs of stats::optim
fr <- function(x) {   ## Rosenbrock Banana function
    x1 <- x[1]
    x2 <- x[2]
    100 * (x2 - x1 * x1)^2 + (1 - x1)^2
}
grr <- function(x) { ## Gradient of 'fr'
    x1 <- x[1]
    x2 <- x[2]
    c(-400 * x1 * (x2 - x1 * x1) - 2 * (1 - x1),
       200 *      (x2 - x1 * x1))
}
# first with stats::optim
stats::optim(c(-1.2,1), fr, grr, method = "L-BFGS-B")
#> $par
#> [1] 0.9999997 0.9999995
#> 
#> $value
#> [1] 2.267577e-13
#> 
#> $counts
#> function gradient 
#>       47       47 
#> 
#> $convergence
#> [1] 0
#> 
#> $message
#> [1] "CONVERGENCE: REL_REDUCTION_OF_F <= FACTR*EPSMCH"

And then with torch:

optim_torch(
 torch_tensor(c(-1.2, 1), requires_grad = TRUE),
 fr,
 method = "lbfgs",
 control = list(maxiter = 10)
)
#> $par
#> torch_tensor
#>  1.0000
#>  1.0000
#> [ CPUFloatType{2} ]
#> 
#> $value
#> torch_tensor
#> 1e-13 *
#>  3.6948
#> [ CPUFloatType{1} ]
#> 
#> $converged
#> [1] TRUE

Why is this cool you ask? We can optimize a function using lbfgs, but without having to manually figure out it’s gradient and we can also optimize the function on cuda.