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. |
macOS |
Nothing for the Cox model. |
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 needspykeops, which doesn’t support Windows. Skip thepykeopsline 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 thepython3.X-venvpackage). 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 theRETICULATE_PYTHONenvironment variable before loading survivalGPU. See the reticulate documentation.