Deployment, Versioning and Scoring Tools for 'inlabru' Models

Serves Bayesian spatial and latent Gaussian models fit with 'inlabru' as versioned Representational State Transfer (REST) application programming interfaces (APIs). Includes a 'vetiver' bridge (versioning, REST API serving with posterior uncertainty, 'Docker' packaging) for 'bru' objects, a 'parsnip' model constructor scoped to fixed-effects models, and predictive scoring helpers (continuous ranked probability, log, squared-error and Dawid-Sebastiani scores) computed from inlabru::generate() posterior samples via 'scoringRules'.

inlaops

Serve inlabru models as versioned REST APIs that answer with posterior uncertainty, not just a point estimate.

Why

inlabru fits Bayesian spatial and latent Gaussian models well, but a fitted model lives in an R session. Getting its predictions to the people and systems that need them usually means re-running scripts by hand or exporting static tables.

inlaops connects inlabru to vetiver, the standard R tooling for model deployment. A bru fit can then be pinned with version history, served over HTTP and packaged in a container like any lm() or tidymodels fit. Each prediction carries the posterior mean, a 95% credible interval and the posterior standard deviation.

For example, inlabru's gorilla nests article fits a log-Gaussian Cox process to nest locations in the Kagwene Gorilla Sanctuary. Pinned and served with inlaops, that intensity surface (~ exp(field + Intercept)) becomes an endpoint: a field team's app, or a service in any language, sends coordinates to /predict and gets back the expected nest density there, with its credible interval. Nobody on the consuming side runs R. When the model is refitted on a new survey, the new version is pinned and served, and the old versions stay available for comparison and rollback.

Installation

You can install the released version of inlaops from CRAN with:

install.packages("inlaops")

You can install the development version of inlaops from GitHub with:

# install.packages("pak")
pak::pak("novica/inlaops")

Example

library(inlabru)
library(vetiver)
library(pins)
library(plumber)
library(inlaops)

# config = TRUE keeps what the API needs for posterior sampling.
fit <- bru(
  y ~ Intercept(1) + x1 + x2,
  family  = "gaussian",
  data    = dat,
  options = bru_options(control.compute = list(config = TRUE))
)

v <- vetiver_model(fit, "my_model",
  metadata = list(pred_formula = "~ Intercept + x1 + x2")
)

board <- board_folder("model-store")
vetiver_pin_write(board, v)

pr() |> vetiver_api(v) |> pr_run(port = 8088)

From any HTTP client:

curl -X POST http://127.0.0.1:8088/predict \
  -H "Content-Type: application/json" \
  -d '[{"x1": 0.5, "x2": 1.0}]'
#> [{".pred":1.99,".pred_low":1.88,".pred_high":2.11,".pred_sd":0.058}]

Point-referenced spatial fits (sf / sp points, SPDE fields, LGCP intensity surfaces) are served the same way: the request carries coordinates. Polygon / areal fits are not served yet.

inlaops also includes bru_reg(), a parsnip engine for fixed-effects inlabru models, and bru_score() for proper scoring rules (CRPS, log-score) on held-out data.

Installation

# INLA is not on CRAN
install.packages("INLA",
  repos = c(getOption("repos"), INLA = "https://inla.r-inla-download.org/R/stable")
)

# Once on CRAN
install.packages("inlaops")

# Development version
# install.packages("pak")
pak::pak("novica/inlaops")