harryprince/torch

torch from R!

torch

torch from R!

A proof of concept for calling libtorch functions from R. API will change! Use at your own risk. Most libtorch’s functionality is not implemented here too.

Installation

Installation is very simple:

CPU

Sys.setenv(TORCH_HOME="/libtorch")
devtools::install_github("dfalbel/torch")

Code above will check whether libtorch is installed to TORCH_HOME dir. If not it will automatically download libtorch binaries from pytorch.org and unpack them to TORCH_HOME. After that it will install torch R package. If you don’t set the TORCH_HOME env var it will use /libtorch as default.

Alternatively you can provide URL for binaries download by adding setting the TORCH_BINARIES environment variable.

Note: The package will return std::bad_alloc errors (and they crash the R session) if compiled with recent versions of g++ (eg. the default version of Ubuntu Xenial - 5.4.0). It’s recommended to compile the package with g++-4.9. In order to do it:

sudo apt-get install g++-4.9

And add the following to your .R/Makevars (usethis::edit_r_makevars()):

CXX=g++-4.9
CXX11=g++-4.9

You may need to reinstall the Rcpp package.

GPU

On Linux you can also install torch with CUDA 9.0 support (still very initial stage)

Install CUDA 9.0

• follow these instructions and add necessary repositories
• install cuda-9.0 - sudo apt-get install cuda-9-0
• install cuDNN > 7 - follow the instructions here.

Install libtorch and torch R package

Sys.setenv(TORCH_BACKEND = "CUDA")
devtools::install_github("dfalbel/torch")

Example

Currently this package is only a prrof of concept and you can only create a Torch Tensor from an R object. And then convert back from a torch Tensor to an R object.

library(torch)
x <- array(runif(8), dim = c(2, 2, 2))
y <- tensor(x)
y
#> tensor 
#> (1,.,.) = 
#>   0.0899  0.5985
#>   0.9210  0.0111
#> 
#> (2,.,.) = 
#>   0.6249  0.2564
#>   0.3077  0.3371
#> [ Variable[CPUFloatType]{2,2,2} ]
identical(x, as.array(y))
#> [1] FALSE

Simple Autograd Example

In the following snippet we let torch, using the autograd feature, calculate the derivatives:

x <- tensor(1, requires_grad = TRUE)
w <- tensor(2, requires_grad = TRUE)
b <- tensor(3, requires_grad = TRUE)

y <- w * x + b
y$backward()

x$grad
#> tensor 
#>  2
#> [ Variable[CPUFloatType]{1} ]
w$grad
#> tensor 
#>  1
#> [ Variable[CPUFloatType]{1} ]
b$grad
#> tensor 
#>  1
#> [ Variable[CPUFloatType]{1} ]

Linear Regression

In the following example we are going to fit a linear regression from scratch using torch’s Autograd.

Note all methods that end with _ (eg. sub_), will modify the tensors in place.

x <- matrix(runif(100), ncol = 2)
y <- matrix(0.1 + 0.5 * x[,1] - 0.7 * x[,2], ncol = 1)

x_t <- tensor(x)
y_t <- tensor(y)

w <- tensor(matrix(rnorm(2), nrow = 2), requires_grad = TRUE)
b <- tensor(0, requires_grad = TRUE)

lr <- 0.5

for (i in 1:100) {
  y_hat <- tch_mm(x_t, w) + b
  loss <- tch_mean((y_t - y_hat)^2)
  
  loss$backward()
  
  with_no_grad({
    w$sub_(w$grad*lr)
    b$sub_(b$grad*lr)   
  })
  
  w$grad$zero_()
  b$grad$zero_()
}

print(w)
#> tensor 
#>  0.4999
#> -0.6990
#> [ Variable[CPUFloatType]{2,1} ]
print(b) 
#> tensor 
#> 0.01 *
#>  9.9514
#> [ Variable[CPUFloatType]{1} ]