survivalGPU

GPU-accelerated survival analysis — Cox Proportional Hazards (CoxPH) and Weighted Cumulative Exposure (WCE) models, built on PyTorch (with KeOps additionally powering WCE). Classical survival models don’t scale well to large datasets or to heavy bootstrap resampling on CPU; survivalGPU runs the core computations on the GPU to make both practical, while still working without one (CPU fallback).

Two interfaces, one implementation

survivalGPU is usable from both Python and R, and both give you the same models with the same results — there’s a single canonical implementation (in Python), not two separate codebases to keep in sync.

import numpy as np
from survivalgpu import CoxPHSurvivalAnalysis

stop = np.array([1, 1, 2], dtype=np.int64)
event = np.array([0, 1, 1], dtype=np.int64)
covariates = np.array([[1.0], [0.0], [4.0]])

model = CoxPHSurvivalAnalysis(ties="efron")
model.fit(covariates, stop, event=event)

print(model.coef_)
library(survivalGPU)
library(survival)

data <- lung
data$status <- lung$status - 1  # recode to 0/1

fit <- coxphGPU(Surv(time, status) ~ age + sex, data = data)
summary(fit)

coxphGPU()/wceGPU() are built as drop-in companions to survival::coxph() and WCE::WCE() respectively, so R users familiar with those functions should feel at home.

These two R packages are the reference implementations of the models, and survivalGPU is tested against them: survival by Terry Therneau, and WCE by Marie-Pierre Sylvestre and colleagues, which implements the method of Sylvestre and Abrahamowicz (2009), doi:10.1002/sim.3701.

The reticulate bridge

The R package doesn’t reimplement the models — coxphGPU() and wceGPU() are thin R wrappers that call directly into the same Python classes shown above (CoxPHSurvivalAnalysis, WCESurvivalAnalysis), via reticulate. Concretely: the R package’s source tree bundles the actual survivalgpu Python package (as a copy of the same code, not a separate reimplementation), and on load, points reticulate at it with a delayed import — so loading the R package itself never requires Python, only actually calling coxphGPU()/wceGPU() does.

In practice, this means:

  • A fix or new feature in the Python implementation is immediately available from R too, with no separate porting step.

  • The R side needs a working Python environment (PyTorch, + KeOps for WCE) set up once — see Installation — but nothing else Python-specific to think about afterwards.