tests/testthat/baselines/README.md

In nlmixr2est: Nonlinear Mixed Effects Models in Population PK/PD, Estimation Routines

Test Baselines - Static Reference Data

Checked-in golden/reference .rds files that tests compare against:

• qrpem-baseline-ref.rds -- pre-QRPEM impmap baseline used by test-qrpem-slow.R; requires re-running the historical code to regenerate.
• vae-encoder-golden.rds -- torch autograd oracle for the C++ LSTM encoder, used by test-vae-encoder.R; regenerated by tools/vaeGenEncoderFixtures.R (needs torch).
• vi-stan-golden.rds -- Stan ADVI (rstan::vb()) posterior means for the linear random-intercept cross-check of est="fbvi", used by test-vi-stan.R; regenerated by tools/viGenStanFixture.R (needs rstan + a C++ toolchain). Stores the simulated DV vector as a fingerprint, so the test skips rather than asserts if the data drifts.
• gradref-*.rds -- cached numeric references for the analytic-derivative tests: central differences of the objective (analytic outer gradient) and covMethod="r" standard errors (analytic covariance). Each is a property of the model, data and theta point, not of the implementation under test, so they are frozen here rather than recomputed -- rebuilding all of them costs ~178 fits. Written and read by .numRef() in ../helper-gradref.R; regenerate deliberately with NLMIXR2EST_REGEN_GRADREF=TRUE when the objective itself legitimately changes.

Unlike ../fixtures/ (a regenerable fit cache that is gitignored and cleared by .nlmixr2estbuild() during devtools::document()), files here must never be deleted by cache clearing. (gradref-*.rds are produced by the test suite, but only on an explicit regeneration run -- never as a side effect of an ordinary run.)

nlmixr2est documentation built on Aug. 5, 2026, 1:11 a.m.