LoginSign Up
← Back to all sparks
I

inlabru

INFRA · APIS
Velocity2.5

Bayesian Latent Gaussian Modelling using INLA and Extensions

github.com/inlabru-org/inlabru ↗

A Bayesian spatial modelling package rebuilding its foundations one breaking release at a time

bayesian-modellingspatial-statisticsr-packageapi-consolidationdependency-pruningbreaking-changes
◆Current state
inlabru wraps INLA for spatial, point-process and latent-Gaussian models in R. It is mid-modernisation: since 2.12.0 cut the sp stack, each release has renamed part of the public surface, standardised how external packages attach custom mappers, or replaced internal machinery. 2.15.0 is the latest step, pairing a new predictor evaluation and linearisation implementation with broom's tidy(), glance() and augment() methods and four non-zero-truncated observation families.
◆Where it's heading
The arc is consolidation of the extension surface rather than expansion of the model catalogue. Every release adds mappers or families with one hand and removes a dependency, a re-export or a deprecated path with the other — plyr in 2.15.0, fmesher's Depends entry in 2.14.1, sp and ggmap in 2.12.0. The compatibility flag bru_compat_pre_2_14_enable and the temporary fm_int/fm_pixels re-exports show a maintainer sequencing breaks across releases instead of landing them together.
◆Prediction
The 2.14 compatibility flag is still defaulting to TRUE and the fmesher re-exports are described in the entries as temporary, so the next obvious move is a release that flips bru_compat_pre_2_14_enable off and drops those re-exports.

◆Recent moves

  1. 23d ago

    Predictor linearisation rewritten; broom tidiers, truncated families

    The predictor evaluation and linearisation rewrite continues the internal-machinery overhaul that 2.14.0 staged, while the broom methods and the nz* observation families widen the ecosystem and model surface without changing how bru() is called. Consistent with the arc, it also drops the plyr dependency and moves stats out of Depends.

    View source ↗
  2. 3mo ago

    Bugfix release: factor contrasts, raster extraction, error classes

    A CRAN-targeted patch: factor contrast levels, SpatRaster factor extraction, and error-class checks widened for R > 4.6. The one structural item, moving fmesher from Depends to Imports, is deliberately softened by keeping fm_int and fm_pixels re-exported, so nothing user-facing breaks yet.

    View source ↗
  3. 5mo ago

    New mappers, standardised cgeneric support, bru_obs storage refactor

    The substantive move is standardising where a pre-computed mapper lives on inla.cgeneric and inla.rgeneric objects, so external packages no longer need their own bru_get_mapper() methods. The bru_obs/bru_comp storage refactor is staged behind bru_compat_pre_2_14_enable, which is the pattern this package now uses to spread a break across releases.

    View source ↗
  4. 1y ago

    Mapper classes shortened to bm_*, experimental predictor aggregation

    Two renames of core public classes — bru_like to bru_obs and bru_mapper_<type> to bm_<type> — both aliased for back-compatibility, plus an experimental aggregation step for predictors and automatic detection of purely additive models. The pre-evaluated component inputs, described as halving preprocessing time on large spatio-temporal models, are the concrete user-visible gain.

    View source ↗
  5. 1y ago

    Drops sp and ggmap for an sf-native spatial stack

    ⚡ SPARK

    This is the release the rest of the visible history executes against: once the spatial data layer became sf-native, the mapper renames, namespace pruning and extension-point standardisation in 2.13 and 2.14 follow as cleanup. It is also the first appearance of the package's habit of shipping a removal alongside an addition.

    View source ↗