tests/testthat/test-tuning.R

# Tests for hyperparameter tuning

.gbsg2 <- function() {
  skip_if_not_installed("survival")
  skip_if_not_installed("TH.data")
  library("survival")
  data("GBSG2", package = "TH.data", envir = environment())
  GBSG2$surv <- with(GBSG2, Surv(time, cens))
  GBSG2
}

test_that("cross validation", {
  GBSG2 <- .gbsg2()
  set.seed(241068)
  folds <- sample(rep(1:2, length.out = nrow(GBSG2)))
  ## default interface
  X <- 1 * matrix(GBSG2$horTh == "yes", ncol = 1)
  colnames(X) <- "horThyes"
  m <- Coxph(surv ~ 1, data = GBSG2, log_first = TRUE)
  mt <- suppressWarnings(tramnet(model = m, x = X, lambda = 0, alpha = 0))
  cv <- suppressMessages(suppressWarnings(
    cvl_tramnet(mt, lambda = c(0, 1), alpha = c(0, 1), folds = folds)
  ))
  expect_s3_class(cv, "cvl_tramnet")
  expect_equal(nrow(cv$logLik_tab), 4L)
  ## formula interface: scaling within folds
  mf <- Coxphnet(surv ~ horTh + age,
    data = GBSG2, lambda = 0, alpha = 0,
    log_first = TRUE
  )
  cvf <- suppressMessages(suppressWarnings(
    cvl_tramnet(mf,
      lambda = c(0, 5), alpha = 1, folds = folds,
      fit_opt = TRUE
    )
  ))
  expect_true(all(is.finite(cvf$logLik_tab$sum_logLik)))
  expect_equal(
    colnames(cvf$coefficients[[1]]),
    c("lambda", "alpha", "horThyes", "age")
  )
  expect_s3_class(cvf$full_fit, "tramnet")
  ## out-of-sample log-likelihood by hand
  trn <- GBSG2[folds != 1, ]
  tst <- GBSG2[folds == 1, ]
  m1 <- Coxphnet(surv ~ horTh + age,
    data = trn, lambda = 5, alpha = 1,
    log_first = TRUE
  )
  expect_equal(cvf$logLik_tab$logLik_fold_1[2], c(logLik(m1, newdata = tst)),
    tolerance = 1e-6
  )
})

test_that("profiling", {
  GBSG2 <- .gbsg2()
  mf <- Coxphnet(surv ~ horTh + age + tsize,
    data = GBSG2, lambda = 1,
    alpha = 1, log_first = TRUE
  )
  pfl <- suppressMessages(prof_lambda(mf, nprof = 3))
  expect_s3_class(pfl, "prof_lambda")
  expect_equal(dim(pfl$cfx), c(3L, 3L))
  pfa <- suppressMessages(prof_alpha(mf, nprof = 3))
  expect_s3_class(pfa, "prof_alpha")
  pdf(NULL)
  on.exit(dev.off())
  expect_no_error(plot_path(pfl, plot_logLik = TRUE))
  expect_no_error(plot_path(pfa))
  ## default interface
  X <- scale(model.matrix(~ 0 + age + tsize, data = GBSG2))
  m0 <- Coxph(surv ~ 1, data = GBSG2, log_first = TRUE)
  mt <- tramnet(m0, x = X, lambda = 1, alpha = 1)
  pfd <- suppressMessages(prof_lambda(mt, nprof = 2))
  expect_equal(colnames(pfd$cfx), c("age", "tsize"))
  ## Lm
  ml <- Lmnet(dist ~ speed, data = cars, lambda = 0, alpha = 0)
  pfl <- suppressMessages(prof_lambda(ml, nprof = 3, as.lm = TRUE))
  expect_equal(colnames(pfl$cfx), "speed")
})

test_that("model based optimization", {
  skip_on_cran()
  skip_if_not_installed("DiceKriging")
  GBSG2 <- .gbsg2()
  set.seed(241068)
  mf <- Coxphnet(surv ~ horTh + age + tsize,
    data = GBSG2, lambda = 0,
    alpha = 0, log_first = TRUE
  )
  for (type in c("lasso", "ridge", "elnet")) {
    tmbo <- suppressWarnings(mbo_tramnet(mf,
      obj_type = type, fold = 2,
      n_design = 4, n_iter = 2,
      verbose = FALSE
    ))
    expect_equal(nrow(tmbo$instance$archive$data), 6L)
    expect_true(tmbo$x$lmb >= 0 & tmbo$x$lmb <= 16)
    rec <- mbo_recommended(tmbo, mf)
    expect_s3_class(rec, "tramnet")
    expect_equal(rec$tuning_parm[["lambda"]], tmbo$x$lmb)
    alp <- switch(type,
      "lasso" = 1,
      "ridge" = 0,
      tmbo$x$alp
    )
    expect_equal(rec$tuning_parm[["alpha"]], alp)
  }
  ## default interface and objective functions
  X <- scale(model.matrix(~ 0 + age + tsize, data = GBSG2))
  m0 <- Coxph(surv ~ 1, data = GBSG2, log_first = TRUE)
  mt <- tramnet(m0, x = X, lambda = 0, alpha = 0)
  tmbo <- suppressWarnings(mbo_tramnet(mt,
    obj_type = "lasso", n_design = 3,
    n_iter = 1, verbose = FALSE,
    opt_crit = "cb"
  ))
  expect_s3_class(mbo_recommended(tmbo, m0, X), "tramnet")
  ## user-specified folds, noisy objective and learner object
  lrn_km <- mlr3learners::LearnerRegrKM$new()
  lrn_km$predict_type <- "se"
  lrn_km$param_set$set_values(
    nugget.estim = TRUE,
    control = list(trace = FALSE)
  )
  tmbo <- suppressWarnings(mbo_tramnet(mt,
    obj_type = "ridge", n_design = 3,
    n_iter = 1, verbose = FALSE,
    noisy = TRUE, learner = lrn_km
  ))
  expect_equal(nrow(tmbo$instance$archive$data), 4L)
  tmbo <- suppressWarnings(mbo_tramnet(
    mt,
    obj_type = "lasso", n_design = 3, n_iter = 1, verbose = FALSE,
    folds = rep(1:3, length.out = nrow(GBSG2)), noisy = FALSE
  ))
  expect_equal(nrow(tmbo$instance$archive$data), 4L)
  expect_warning(
    suppressMessages(
      mbo_tramnet(mf, n_design = 3, n_iter = 1, verbose = FALSE, foo = 1)
    ),
    "ignored"
  )
  obj <- elnet_obj(mt, folds = rep(1:2, length.out = nrow(GBSG2)), fold = 2)
  expect_s3_class(obj, "ObjectiveRFun")
  expect_true(is.finite(suppressWarnings(
    suppressMessages(obj$eval(list(lmb = 1, alp = 0.5)))$y
  )))
})

Try the tramnet package in your browser

Any scripts or data that you put into this service are public.

tramnet documentation built on Oct. 16, 2026, 5:07 p.m.