Installation

survivalGPU has two parts to install: the R package itself, from CRAN, and the Python libraries it uses behind the scenes (via the reticulate R package).

1. Before you start

You need R >= 4.1 and Python >= 3.10. The R package itself has no compiled code, but the WCE model needs a C++ compiler:

System

Install first

Linux

Nothing for the Cox model.
For the WCE model, a C++ compiler, usually already there. If not, on Debian/Ubuntu: sudo apt install build-essential

macOS

Nothing for the Cox model.
For the WCE model, Apple’s compiler: run xcode-select --install in a terminal, and also brew install libomp (needs Homebrew).

Windows

Nothing.

The compiler is used by the WCE model, which compiles its routines the first time it runs.

If you don’t have Python >= 3.10, reticulate can install it for you: reticulate::install_python("3.11").

2. Install the R package

install.packages("survivalGPU")

To install the development version from GitHub instead:

install.packages("remotes")
remotes::install_github("jeanfeydy/survivalGPU", subdir = "R")

The R package has no compiled code, so this doesn’t need a compiler.

3. Install the Python libraries

Create a Python environment called survivalGPU, and install the libraries in it. You only do this once; PyTorch is a large download, so it takes a few minutes.

library(reticulate)

virtualenv_create("survivalGPU", version = ">=3.10")
virtualenv_install("survivalGPU", c("numpy", "torch", "pandas", "scipy",
                                    "matplotlib", "beartype", "jaxtyping"))

# Only for the WCE model, wceGPU() -- not available on Windows:
virtualenv_install("survivalGPU", "pykeops")

4. Use it

At the start of each R session, select the Python environment before loading survivalGPU:

library(reticulate)
use_virtualenv("survivalGPU")

library(survivalGPU)
use_cuda()   # TRUE if a GPU is detected; FALSE means survivalGPU runs on CPU

Then use coxphGPU() and wceGPU(): see the user guide for examples.

Platform-specific considerations

  • Windows: the Cox model (coxphGPU()) works, but the WCE model (wceGPU()) doesn’t: it needs pykeops, which doesn’t support Windows. Skip the pykeops line above. To use WCE on Windows, install R inside WSL2 and follow the Linux instructions.

  • macOS: there’s no NVIDIA GPU on Macs, so survivalGPU runs on CPU.

  • GPU on Windows: the PyTorch installed by default on Windows only uses the CPU. To use an NVIDIA GPU, reinstall PyTorch from its GPU package index, with the CUDA version that matches your setup (see pytorch.org), for example:

    reticulate::virtualenv_install("survivalGPU", "torch",
      pip_options = "--index-url https://download.pytorch.org/whl/cu124")
    
  • GPU and the WCE model: to run wceGPU() on an NVIDIA GPU, KeOps also needs the CUDA toolkit (nvcc), not only the GPU driver.

Troubleshooting

  • virtualenv_create() fails with “Suitable Python installation for creating a venv not found”? The Python it picked can’t create virtual environments (e.g. on Ubuntu, without the python3.X-venv package). Give it another Python explicitly:

    virtualenv_create("survivalGPU", python = "/usr/bin/python3.10")
    # or: virtualenv_create("survivalGPU", python = install_python("3.11"))
    
  • Your own Python environment: you can also install these libraries in an environment of your own, and select it with reticulate::use_virtualenv(), reticulate::use_condaenv() or the RETICULATE_PYTHON environment variable before loading survivalGPU. See the reticulate documentation.