Nothing
R/glmnet.control.R
In glmnet: Lasso and Elastic-Net Regularized Generalized Linear Models
Defines functions
.resolve_control
#' Internal glmnet algorithm parameters
#'
#' View and/or change the factory default parameters that control
#' glmnet's numerical algorithms.
#'
#' If called with no arguments, \code{glmnet.control()} returns a list
#' with the current settings of these parameters. Any arguments
#' included in the call set those parameters to the new values, and
#' then silently return the updated settings. Values set via
#' \code{glmnet.control()} persist for the duration of the R session.
#' All parameters listed here can also be supplied per-call via the
#' \code{control = list(...)} argument of \code{\link{glmnet}()}, which
#' does \emph{not} mutate the session state; see that function's
#' documentation for the precedence rules.
#'
#' @section Parameter taxonomy:
#' glmnet has two execution paths, and not every control parameter is
#' consumed by both:
#'
#' \itemize{
#' \item \strong{Core-engine path} --- \code{family} passed as a
#' character string (\code{"gaussian"}, \code{"binomial"},
#' \code{"poisson"}, \code{"multinomial"}, \code{"mgaussian"},
#' \code{"cox"}) routes to purpose-built C++ glmnetpp kernels.
#' \item \strong{GLM-family (R-IRLS) path} --- \code{family} passed
#' as a \code{family()} object (or \code{glmnet.path()} /
#' \code{glmnet.fit()} called directly) runs an IRLS loop
#' implemented in R that uses the C++ \code{wls} kernel as its
#' inner solver.
#' }
#'
#' The table below shows the role and the scope of each parameter.
#' "Scope: engine" means the parameter is only meaningful on the
#' core-engine path; "Scope: R-IRLS" means only on the GLM-family path;
#' "Scope: both" means consulted by both.
#'
#' \tabular{llll}{
#' \strong{Parameter} \tab \strong{Role} \tab \strong{Scope} \tab \strong{Notes} \cr
#' \code{fdev} \tab path termination \tab both \tab C++ name \code{sml}; min fractional deviance change \cr
#' \code{devmax} \tab path termination \tab both \tab C++ name \code{rsqmax}; max explained-deviance ratio \cr
#' \code{mnlam} \tab path termination \tab both \tab min lambda count before early stop \cr
#' \code{eps} \tab path termination \tab both \tab lambda.min.ratio floor \cr
#' \code{thresh} \tab CD tolerance \tab both \tab inner coordinate-descent convergence \cr
#' \code{maxit} \tab CD budget \tab both \tab max CD passes across all lambdas \cr
#' \code{dfmax} \tab coefficient-count cap \tab both \tab default resolves to \code{nvars + 1} at call time \cr
#' \code{pmax} \tab coefficient-count cap \tab both \tab default resolves to \code{min(dfmax*2+20, nvars)} at call time \cr
#' \code{big} \tab numerical guard \tab both \tab \code{Inf} substitute for bounds limits \cr
#' \code{itrace} \tab progress \tab both \tab aliased with \code{trace.it} \cr
#' \code{trace.it} \tab progress \tab both \tab aliased with \code{itrace} \cr
#' \code{pmin} \tab probability floor \tab engine \tab logistic-family kernels only (\code{lognet}, \code{multnet}) \cr
#' \code{exmx} \tab exp() cap \tab engine \tab logistic-family kernels only \cr
#' \code{prec} \tab bounds-subsolver tolerance \tab engine \tab active when a coefficient has finite \code{lower} / \code{upper} \cr
#' \code{mxit} \tab bounds-subsolver budget \tab engine \tab paired with \code{prec} via C++ \code{chg_bnorm()} \cr
#' \code{epsnr} \tab Newton-Raphson tolerance \tab R-IRLS \tab \emph{inert on the core-engine path}; only \code{glmnet.fit()} reads it \cr
#' \code{mxitnr} \tab Newton-Raphson budget \tab R-IRLS \tab \emph{inert on the core-engine path}; only \code{glmnet.fit()} reads it \cr
#' }
#'
#' The naming zoo (three tolerances, three iteration budgets) reflects
#' the fact that glmnet contains three nested loops:
#' \itemize{
#' \item \strong{Outer path}: iterate over lambda values. Early
#' termination governed by \code{fdev}, \code{devmax}, \code{mnlam},
#' \code{eps}.
#' \item \strong{Middle Newton/IRLS loop} (only on the R-IRLS path):
#' convergence by \code{epsnr}, budget by \code{mxitnr}.
#' \item \strong{Inner coordinate descent}: convergence by
#' \code{thresh}, budget by \code{maxit}.
#' \item \strong{Bounds subsolver} (only for coefficients with finite
#' lower/upper): convergence by \code{prec}, budget by \code{mxit}.
#' }
#'
#' @param fdev Path-termination parameter: minimum fractional change
#' in deviance between consecutive lambdas for the path to continue.
#' Scope: both paths. Factory default \code{1.0e-5}.
#' @param devmax Path-termination parameter: maximum fraction of null
#' deviance the path is allowed to explain before stopping. Scope:
#' both paths. Factory default \code{0.999}.
#' @param eps Path-termination parameter: floor on
#' \code{lambda.min.ratio} (smallest allowed \eqn{\lambda_{\min} /
#' \lambda_{\max}}). Scope: both paths. Factory default
#' \code{1.0e-6}.
#' @param big Numerical guard: finite substitute for \code{Inf} in
#' \code{lower.limits} / \code{upper.limits}. Scope: both paths.
#' Factory default \code{9.9e35}.
#' @param mnlam Path-termination parameter: minimum number of lambda
#' values to compute before early stopping is allowed. Scope: both
#' paths. Factory default \code{5}.
#' @param pmin Numerical guard: floor on predicted probability in
#' logistic-family kernels (prevents \code{log(0)}). Implies a max
#' of \code{1 - pmin}. Scope: engine only (logistic-family kernels).
#' Factory default \code{1.0e-9}.
#' @param exmx Numerical guard: cap on \code{exp()} arguments in
#' logistic-family kernels (prevents overflow). Scope: engine only
#' (logistic-family kernels). Factory default \code{250}.
#' @param prec Bounds-subsolver convergence tolerance. Used by the
#' engine's \code{bnorm()} subroutine when a coefficient has a
#' finite \code{lower} or \code{upper} bound. Scope: engine only
#' (bounds-constrained fits). Factory default \code{1.0e-10}.
#' @param mxit Bounds-subsolver iteration budget. Paired with
#' \code{prec}. Scope: engine only (bounds-constrained fits).
#' Factory default \code{100}.
#' @param itrace Progress-bar flag. \code{1} enables the progress bar
#' in \code{glmnet()} and \code{cv.glmnet()}. Scope: both paths.
#' Aliased with \code{trace.it} (setting one sets the other).
#' Factory default \code{0}.
#' @param epsnr Newton-Raphson convergence tolerance for the R-level
#' IRLS loop in \code{glmnet.fit()}. \strong{Scope: R-IRLS path
#' only} --- the core-engine kernels do not read this parameter.
#' Factory default \code{1.0e-6}.
#' @param mxitnr Newton-Raphson iteration budget for the R-level IRLS
#' loop in \code{glmnet.fit()}. \strong{Scope: R-IRLS path only}.
#' Factory default \code{25}.
#' @param thresh Inner coordinate-descent convergence threshold. Each
#' inner CD loop continues until the maximum change in the objective
#' after any coefficient update is less than \code{thresh} times the
#' null deviance. Scope: both paths. Factory default \code{1e-7}.
#' Can also be set per-call via \code{control = list(thresh = ...)}
#' in \code{glmnet()}.
#' @param maxit Inner coordinate-descent iteration budget: maximum
#' total passes over the data across all lambda values. Scope: both
#' paths. Factory default \code{100000}. Can also be set per-call
#' via \code{control = list(maxit = ...)}.
#' @param dfmax Coefficient-count cap: limit the maximum number of
#' variables in the model at any lambda. Scope: both paths. Factory
#' default \code{NULL}, which resolves at call time to \code{nvars +
#' 1} (i.e., unconstrained). Can also be set per-call via
#' \code{control = list(dfmax = ...)}.
#' @param pmax Coefficient-count cap: limit the maximum number of
#' variables ever to be nonzero anywhere on the path. Scope: both
#' paths. Factory default \code{NULL}, which resolves at call time
#' to \code{min(dfmax * 2 + 20, nvars)}. Can also be set per-call
#' via \code{control = list(pmax = ...)}.
#' @param trace.it Progress-bar flag (alias of \code{itrace}). Scope:
#' both paths. Factory default \code{0}.
#' @param factory If \code{TRUE}, reset all parameters to their
#' factory defaults. Default is \code{FALSE}.
#' @return A list with named elements as in the argument list.
#' @author Jerome Friedman, Kenneth Tay, Trevor Hastie\cr
#' Maintainer: Trevor Hastie \email{hastie@@stanford.edu}
#' @seealso \code{\link{glmnet}}
#' @keywords models regression
#' @examples
#'
#' glmnet.control(fdev = 0) # continue along path even though not much changes
#' glmnet.control(thresh = 1e-8) # tighten CD convergence for session
#' glmnet.control() # view current settings
#' glmnet.control(factory = TRUE) # reset all the parameters to their default
#'
#' @export glmnet.control
## Factory defaults for all 17 algorithm-control parameters. Used by
## glmnet.control(factory = TRUE).
.GLMNET_CONTROL_FACTORY <- list(
fdev = 1e-5,
eps = 1e-6,
big = 9.9e35,
mnlam = 5L,
devmax = 0.999,
pmin = 1e-9,
exmx = 250,
itrace = 0L,
prec = 1e-10,
mxit = 100L,
epsnr = 1e-6,
mxitnr = 25L,
thresh = 1e-7,
maxit = 100000L,
dfmax = NULL,
pmax = NULL,
trace.it = 0L
)
glmnet.control <-
function (fdev = 1e-05, devmax = 0.999, eps = 1e-06, big = 9.9e+35,
mnlam = 5, pmin = 1e-09, exmx = 250, prec = 1e-10, mxit = 100,
itrace = 0, epsnr = 1e-06, mxitnr = 25,
thresh = 1e-7, maxit = 100000, dfmax = NULL,
pmax = NULL, trace.it = 0,
factory = FALSE)
{
inquiry <- !nargs()
if (factory) {
glmnet_control_set(.GLMNET_CONTROL_FACTORY)
} else {
## Build update list from supplied args only. Anything left at
## its formal default is left alone. Note this also fixes a
## pre-existing 5.0 bug: setting `prec` alone would silently
## reset `mxit` to 100 because `chg_bnorm` paired them.
updates <- list()
if (!missing(fdev)) updates$fdev <- as.double(fdev)
if (!missing(devmax)) updates$devmax <- as.double(devmax)
if (!missing(eps)) updates$eps <- as.double(eps)
if (!missing(big)) updates$big <- as.double(big)
if (!missing(mnlam)) updates$mnlam <- as.integer(mnlam)
if (!missing(pmin)) updates$pmin <- as.double(pmin)
if (!missing(exmx)) updates$exmx <- as.double(exmx)
if (!missing(prec)) updates$prec <- as.double(prec)
if (!missing(mxit)) updates$mxit <- as.integer(mxit)
if (!missing(itrace)) updates$itrace <- as.integer(itrace)
if (!missing(epsnr)) updates$epsnr <- as.double(epsnr)
if (!missing(mxitnr)) updates$mxitnr <- as.integer(mxitnr)
if (!missing(thresh)) updates$thresh <- as.double(thresh)
if (!missing(maxit)) updates$maxit <- as.integer(maxit)
if (!missing(dfmax)) updates$dfmax <- if (is.null(dfmax)) NULL else as.integer(dfmax)
if (!missing(pmax)) updates$pmax <- if (is.null(pmax)) NULL else as.integer(pmax)
if (!missing(trace.it)) updates$trace.it <- as.integer(trace.it)
if (length(updates)) glmnet_control_set(updates)
}
## Single read-back of the C++ state. dfmax/pmax come back as
## NA_integer_ when unset; map to NULL to preserve the historical
## R-side return shape. Use `["k"] <- list(NULL)` rather than
## `$k <- NULL` because the latter deletes the slot.
value <- glmnet_control_get()
if (is.na(value$dfmax)) value["dfmax"] <- list(NULL)
if (is.na(value$pmax)) value["pmax"] <- list(NULL)
if (inquiry) value else invisible(value)
}
## Set of deprecation messages already emitted this session, so
## .resolve_control() emits each at most once per arg name. (Without
## this, downstream packages that wrap glmnet() in tryCatch(warning =
## ...) -- notably stm -- treat each fit as a failure and grow nlambda
## without bound.)
.deprecation_emitted <- new.env(parent = emptyenv())
## Canonical list of valid control keys (the 17 names returned by
## glmnet.control()). Used by the test helper .control_with() in
## tests/testthat/helper-functions.R. The package itself no longer
## validates control= keys -- glmnet_control_set() silently drops
## unknown names by design.
.VALID_CONTROL_KEYS <- c(
"fdev", "eps", "big", "mnlam", "devmax", "pmin", "exmx", "itrace",
"prec", "mxit", "epsnr", "mxitnr",
"thresh", "maxit", "dfmax", "pmax", "trace.it"
)
## Resolve the layered algorithm-control configuration for one glmnet()
## invocation. Precedence (low to high): session defaults -> control=
## list -> deprecated top-level args.
##
## Arguments:
## control -- the user's control= list (may be empty/NULL).
## nvars -- number of predictors; needed to resolve dfmax/pmax
## data-dependent defaults to integers.
## deprecated -- named list of any deprecated top-level args the
## caller supplied (NULL entries mean "not supplied").
## Each non-NULL entry triggers a once-per-session
## .Deprecated() warning and is folded into control.
##
## Returns a list:
## control -- 17-key resolved list (dfmax/pmax as integers).
## restore -- niladic function to revert C++ session state, or NULL
## if no overrides were applied. Caller registers via
## on.exit().
.resolve_control <- function(control, nvars, deprecated = list()) {
## Fold deprecated top-level args into control= with one warning
## per arg per session. Deprecated args win over control= entries.
deprecated <- deprecated[!vapply(deprecated, is.null, logical(1))]
for (nm in names(deprecated)) {
if (!isTRUE(.deprecation_emitted[[nm]])) {
.deprecation_emitted[[nm]] <- TRUE
.Deprecated(msg = sprintf(
"Passing '%s' to glmnet() is deprecated. Use control = list(%s = ...) instead.",
nm, nm))
}
}
control <- modifyList(if (is.null(control)) list() else control, deprecated)
## Snapshot session state. If the caller supplied any overrides,
## push them to the C++ globals in one shot and register a single
## restore thunk. The C++ setter silently ignores unknown keys, so
## no R-side validation is needed.
gc <- glmnet.control()
if (length(control)) {
glmnet_control_set(control)
restore <- function() glmnet_control_set(gc)
## Build the post-override view by overlaying the known control
## keys onto gc. Avoids a second glmnet.control() round-trip.
valid <- intersect(names(control), .VALID_CONTROL_KEYS)
resolved <- modifyList(gc, control[valid])
} else {
restore <- NULL
resolved <- gc
}
## Resolve dfmax/pmax data-dependent defaults to integers.
resolved$dfmax <- if (is.null(resolved$dfmax)) as.integer(nvars + 1L)
else as.integer(resolved$dfmax)
resolved$pmax <- if (is.null(resolved$pmax)) as.integer(min(resolved$dfmax * 2L + 20L, nvars))
else as.integer(resolved$pmax)
list(control = resolved, restore = restore)
}
Try the glmnet package in your browser
Any scripts or data that you put into this service are public.
glmnet documentation built on May 4, 2026, 9:09 a.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Defines functions .resolve_control
#' Internal glmnet algorithm parameters
#'
#' View and/or change the factory default parameters that control
#' glmnet's numerical algorithms.
#'
#' If called with no arguments, \code{glmnet.control()} returns a list
#' with the current settings of these parameters. Any arguments
#' included in the call set those parameters to the new values, and
#' then silently return the updated settings. Values set via
#' \code{glmnet.control()} persist for the duration of the R session.
#' All parameters listed here can also be supplied per-call via the
#' \code{control = list(...)} argument of \code{\link{glmnet}()}, which
#' does \emph{not} mutate the session state; see that function's
#' documentation for the precedence rules.
#'
#' @section Parameter taxonomy:
#' glmnet has two execution paths, and not every control parameter is
#' consumed by both:
#'
#' \itemize{
#' \item \strong{Core-engine path} --- \code{family} passed as a
#' character string (\code{"gaussian"}, \code{"binomial"},
#' \code{"poisson"}, \code{"multinomial"}, \code{"mgaussian"},
#' \code{"cox"}) routes to purpose-built C++ glmnetpp kernels.
#' \item \strong{GLM-family (R-IRLS) path} --- \code{family} passed
#' as a \code{family()} object (or \code{glmnet.path()} /
#' \code{glmnet.fit()} called directly) runs an IRLS loop
#' implemented in R that uses the C++ \code{wls} kernel as its
#' inner solver.
#' }
#'
#' The table below shows the role and the scope of each parameter.
#' "Scope: engine" means the parameter is only meaningful on the
#' core-engine path; "Scope: R-IRLS" means only on the GLM-family path;
#' "Scope: both" means consulted by both.
#'
#' \tabular{llll}{
#' \strong{Parameter} \tab \strong{Role} \tab \strong{Scope} \tab \strong{Notes} \cr
#' \code{fdev} \tab path termination \tab both \tab C++ name \code{sml}; min fractional deviance change \cr
#' \code{devmax} \tab path termination \tab both \tab C++ name \code{rsqmax}; max explained-deviance ratio \cr
#' \code{mnlam} \tab path termination \tab both \tab min lambda count before early stop \cr
#' \code{eps} \tab path termination \tab both \tab lambda.min.ratio floor \cr
#' \code{thresh} \tab CD tolerance \tab both \tab inner coordinate-descent convergence \cr
#' \code{maxit} \tab CD budget \tab both \tab max CD passes across all lambdas \cr
#' \code{dfmax} \tab coefficient-count cap \tab both \tab default resolves to \code{nvars + 1} at call time \cr
#' \code{pmax} \tab coefficient-count cap \tab both \tab default resolves to \code{min(dfmax*2+20, nvars)} at call time \cr
#' \code{big} \tab numerical guard \tab both \tab \code{Inf} substitute for bounds limits \cr
#' \code{itrace} \tab progress \tab both \tab aliased with \code{trace.it} \cr
#' \code{trace.it} \tab progress \tab both \tab aliased with \code{itrace} \cr
#' \code{pmin} \tab probability floor \tab engine \tab logistic-family kernels only (\code{lognet}, \code{multnet}) \cr
#' \code{exmx} \tab exp() cap \tab engine \tab logistic-family kernels only \cr
#' \code{prec} \tab bounds-subsolver tolerance \tab engine \tab active when a coefficient has finite \code{lower} / \code{upper} \cr
#' \code{mxit} \tab bounds-subsolver budget \tab engine \tab paired with \code{prec} via C++ \code{chg_bnorm()} \cr
#' \code{epsnr} \tab Newton-Raphson tolerance \tab R-IRLS \tab \emph{inert on the core-engine path}; only \code{glmnet.fit()} reads it \cr
#' \code{mxitnr} \tab Newton-Raphson budget \tab R-IRLS \tab \emph{inert on the core-engine path}; only \code{glmnet.fit()} reads it \cr
#' }
#'
#' The naming zoo (three tolerances, three iteration budgets) reflects
#' the fact that glmnet contains three nested loops:
#' \itemize{
#' \item \strong{Outer path}: iterate over lambda values. Early
#' termination governed by \code{fdev}, \code{devmax}, \code{mnlam},
#' \code{eps}.
#' \item \strong{Middle Newton/IRLS loop} (only on the R-IRLS path):
#' convergence by \code{epsnr}, budget by \code{mxitnr}.
#' \item \strong{Inner coordinate descent}: convergence by
#' \code{thresh}, budget by \code{maxit}.
#' \item \strong{Bounds subsolver} (only for coefficients with finite
#' lower/upper): convergence by \code{prec}, budget by \code{mxit}.
#' }
#'
#' @param fdev Path-termination parameter: minimum fractional change
#' in deviance between consecutive lambdas for the path to continue.
#' Scope: both paths. Factory default \code{1.0e-5}.
#' @param devmax Path-termination parameter: maximum fraction of null
#' deviance the path is allowed to explain before stopping. Scope:
#' both paths. Factory default \code{0.999}.
#' @param eps Path-termination parameter: floor on
#' \code{lambda.min.ratio} (smallest allowed \eqn{\lambda_{\min} /
#' \lambda_{\max}}). Scope: both paths. Factory default
#' \code{1.0e-6}.
#' @param big Numerical guard: finite substitute for \code{Inf} in
#' \code{lower.limits} / \code{upper.limits}. Scope: both paths.
#' Factory default \code{9.9e35}.
#' @param mnlam Path-termination parameter: minimum number of lambda
#' values to compute before early stopping is allowed. Scope: both
#' paths. Factory default \code{5}.
#' @param pmin Numerical guard: floor on predicted probability in
#' logistic-family kernels (prevents \code{log(0)}). Implies a max
#' of \code{1 - pmin}. Scope: engine only (logistic-family kernels).
#' Factory default \code{1.0e-9}.
#' @param exmx Numerical guard: cap on \code{exp()} arguments in
#' logistic-family kernels (prevents overflow). Scope: engine only
#' (logistic-family kernels). Factory default \code{250}.
#' @param prec Bounds-subsolver convergence tolerance. Used by the
#' engine's \code{bnorm()} subroutine when a coefficient has a
#' finite \code{lower} or \code{upper} bound. Scope: engine only
#' (bounds-constrained fits). Factory default \code{1.0e-10}.
#' @param mxit Bounds-subsolver iteration budget. Paired with
#' \code{prec}. Scope: engine only (bounds-constrained fits).
#' Factory default \code{100}.
#' @param itrace Progress-bar flag. \code{1} enables the progress bar
#' in \code{glmnet()} and \code{cv.glmnet()}. Scope: both paths.
#' Aliased with \code{trace.it} (setting one sets the other).
#' Factory default \code{0}.
#' @param epsnr Newton-Raphson convergence tolerance for the R-level
#' IRLS loop in \code{glmnet.fit()}. \strong{Scope: R-IRLS path
#' only} --- the core-engine kernels do not read this parameter.
#' Factory default \code{1.0e-6}.
#' @param mxitnr Newton-Raphson iteration budget for the R-level IRLS
#' loop in \code{glmnet.fit()}. \strong{Scope: R-IRLS path only}.
#' Factory default \code{25}.
#' @param thresh Inner coordinate-descent convergence threshold. Each
#' inner CD loop continues until the maximum change in the objective
#' after any coefficient update is less than \code{thresh} times the
#' null deviance. Scope: both paths. Factory default \code{1e-7}.
#' Can also be set per-call via \code{control = list(thresh = ...)}
#' in \code{glmnet()}.
#' @param maxit Inner coordinate-descent iteration budget: maximum
#' total passes over the data across all lambda values. Scope: both
#' paths. Factory default \code{100000}. Can also be set per-call
#' via \code{control = list(maxit = ...)}.
#' @param dfmax Coefficient-count cap: limit the maximum number of
#' variables in the model at any lambda. Scope: both paths. Factory
#' default \code{NULL}, which resolves at call time to \code{nvars +
#' 1} (i.e., unconstrained). Can also be set per-call via
#' \code{control = list(dfmax = ...)}.
#' @param pmax Coefficient-count cap: limit the maximum number of
#' variables ever to be nonzero anywhere on the path. Scope: both
#' paths. Factory default \code{NULL}, which resolves at call time
#' to \code{min(dfmax * 2 + 20, nvars)}. Can also be set per-call
#' via \code{control = list(pmax = ...)}.
#' @param trace.it Progress-bar flag (alias of \code{itrace}). Scope:
#' both paths. Factory default \code{0}.
#' @param factory If \code{TRUE}, reset all parameters to their
#' factory defaults. Default is \code{FALSE}.
#' @return A list with named elements as in the argument list.
#' @author Jerome Friedman, Kenneth Tay, Trevor Hastie\cr
#' Maintainer: Trevor Hastie \email{hastie@@stanford.edu}
#' @seealso \code{\link{glmnet}}
#' @keywords models regression
#' @examples
#'
#' glmnet.control(fdev = 0) # continue along path even though not much changes
#' glmnet.control(thresh = 1e-8) # tighten CD convergence for session
#' glmnet.control() # view current settings
#' glmnet.control(factory = TRUE) # reset all the parameters to their default
#'
#' @export glmnet.control
## Factory defaults for all 17 algorithm-control parameters. Used by
## glmnet.control(factory = TRUE).
.GLMNET_CONTROL_FACTORY <- list(
fdev = 1e-5,
eps = 1e-6,
big = 9.9e35,
mnlam = 5L,
devmax = 0.999,
pmin = 1e-9,
exmx = 250,
itrace = 0L,
prec = 1e-10,
mxit = 100L,
epsnr = 1e-6,
mxitnr = 25L,
thresh = 1e-7,
maxit = 100000L,
dfmax = NULL,
pmax = NULL,
trace.it = 0L
)
glmnet.control <-
function (fdev = 1e-05, devmax = 0.999, eps = 1e-06, big = 9.9e+35,
mnlam = 5, pmin = 1e-09, exmx = 250, prec = 1e-10, mxit = 100,
itrace = 0, epsnr = 1e-06, mxitnr = 25,
thresh = 1e-7, maxit = 100000, dfmax = NULL,
pmax = NULL, trace.it = 0,
factory = FALSE)
{
inquiry <- !nargs()
if (factory) {
glmnet_control_set(.GLMNET_CONTROL_FACTORY)
} else {
## Build update list from supplied args only. Anything left at
## its formal default is left alone. Note this also fixes a
## pre-existing 5.0 bug: setting `prec` alone would silently
## reset `mxit` to 100 because `chg_bnorm` paired them.
updates <- list()
if (!missing(fdev)) updates$fdev <- as.double(fdev)
if (!missing(devmax)) updates$devmax <- as.double(devmax)
if (!missing(eps)) updates$eps <- as.double(eps)
if (!missing(big)) updates$big <- as.double(big)
if (!missing(mnlam)) updates$mnlam <- as.integer(mnlam)
if (!missing(pmin)) updates$pmin <- as.double(pmin)
if (!missing(exmx)) updates$exmx <- as.double(exmx)
if (!missing(prec)) updates$prec <- as.double(prec)
if (!missing(mxit)) updates$mxit <- as.integer(mxit)
if (!missing(itrace)) updates$itrace <- as.integer(itrace)
if (!missing(epsnr)) updates$epsnr <- as.double(epsnr)
if (!missing(mxitnr)) updates$mxitnr <- as.integer(mxitnr)
if (!missing(thresh)) updates$thresh <- as.double(thresh)
if (!missing(maxit)) updates$maxit <- as.integer(maxit)
if (!missing(dfmax)) updates$dfmax <- if (is.null(dfmax)) NULL else as.integer(dfmax)
if (!missing(pmax)) updates$pmax <- if (is.null(pmax)) NULL else as.integer(pmax)
if (!missing(trace.it)) updates$trace.it <- as.integer(trace.it)
if (length(updates)) glmnet_control_set(updates)
}
## Single read-back of the C++ state. dfmax/pmax come back as
## NA_integer_ when unset; map to NULL to preserve the historical
## R-side return shape. Use `["k"] <- list(NULL)` rather than
## `$k <- NULL` because the latter deletes the slot.
value <- glmnet_control_get()
if (is.na(value$dfmax)) value["dfmax"] <- list(NULL)
if (is.na(value$pmax)) value["pmax"] <- list(NULL)
if (inquiry) value else invisible(value)
}
## Set of deprecation messages already emitted this session, so
## .resolve_control() emits each at most once per arg name. (Without
## this, downstream packages that wrap glmnet() in tryCatch(warning =
## ...) -- notably stm -- treat each fit as a failure and grow nlambda
## without bound.)
.deprecation_emitted <- new.env(parent = emptyenv())
## Canonical list of valid control keys (the 17 names returned by
## glmnet.control()). Used by the test helper .control_with() in
## tests/testthat/helper-functions.R. The package itself no longer
## validates control= keys -- glmnet_control_set() silently drops
## unknown names by design.
.VALID_CONTROL_KEYS <- c(
"fdev", "eps", "big", "mnlam", "devmax", "pmin", "exmx", "itrace",
"prec", "mxit", "epsnr", "mxitnr",
"thresh", "maxit", "dfmax", "pmax", "trace.it"
)
## Resolve the layered algorithm-control configuration for one glmnet()
## invocation. Precedence (low to high): session defaults -> control=
## list -> deprecated top-level args.
##
## Arguments:
## control -- the user's control= list (may be empty/NULL).
## nvars -- number of predictors; needed to resolve dfmax/pmax
## data-dependent defaults to integers.
## deprecated -- named list of any deprecated top-level args the
## caller supplied (NULL entries mean "not supplied").
## Each non-NULL entry triggers a once-per-session
## .Deprecated() warning and is folded into control.
##
## Returns a list:
## control -- 17-key resolved list (dfmax/pmax as integers).
## restore -- niladic function to revert C++ session state, or NULL
## if no overrides were applied. Caller registers via
## on.exit().
.resolve_control <- function(control, nvars, deprecated = list()) {
## Fold deprecated top-level args into control= with one warning
## per arg per session. Deprecated args win over control= entries.
deprecated <- deprecated[!vapply(deprecated, is.null, logical(1))]
for (nm in names(deprecated)) {
if (!isTRUE(.deprecation_emitted[[nm]])) {
.deprecation_emitted[[nm]] <- TRUE
.Deprecated(msg = sprintf(
"Passing '%s' to glmnet() is deprecated. Use control = list(%s = ...) instead.",
nm, nm))
}
}
control <- modifyList(if (is.null(control)) list() else control, deprecated)
## Snapshot session state. If the caller supplied any overrides,
## push them to the C++ globals in one shot and register a single
## restore thunk. The C++ setter silently ignores unknown keys, so
## no R-side validation is needed.
gc <- glmnet.control()
if (length(control)) {
glmnet_control_set(control)
restore <- function() glmnet_control_set(gc)
## Build the post-override view by overlaying the known control
## keys onto gc. Avoids a second glmnet.control() round-trip.
valid <- intersect(names(control), .VALID_CONTROL_KEYS)
resolved <- modifyList(gc, control[valid])
} else {
restore <- NULL
resolved <- gc
}
## Resolve dfmax/pmax data-dependent defaults to integers.
resolved$dfmax <- if (is.null(resolved$dfmax)) as.integer(nvars + 1L)
else as.integer(resolved$dfmax)
resolved$pmax <- if (is.null(resolved$pmax)) as.integer(min(resolved$dfmax * 2L + 20L, nvars))
else as.integer(resolved$pmax)
list(control = resolved, restore = restore)
}
Try the glmnet package in your browser
Any scripts or data that you put into this service are public.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.