sag: Stochastic Average Gradient with warm-starting

In IshmaelBelghazi/bigoptim: Efficient Large Scale Convex Finite Sums Optimization

Usage

sag(X, y, lambdas, maxiter = NULL, w = NULL, alpha = NULL,
  stepSizeType = 1, Li = NULL, Lmax = NULL, increasing = TRUE,
  d = NULL, g = NULL, covered = NULL, standardize = FALSE,
  tol = 0.001, family = "binomial", fit_alg = "constant",
  user_loss_function = NULL, ...)

Arguments

X	Matrix, possibly sparse of features.
y	Matrix of targets.
lambdas	Vector. Vector of L2 regularization parameters.
maxiter	Maximum number of iterations.
w	Matrix of weights.
alpha	constant step-size. Used only when fit_alg == "constant"
stepSizeType	scalar default is 1 to use 1/L, set to 2 to use 2/(L + n*myu). Only used when fit_alg="linesearch"
Li	Scalar or Matrix.Initial individual Lipschitz approximation.
Lmax	Initial global Lipschitz approximation.
increasing	Boolean. TRUE allows increase of Lipschitz coeffecient. False allows only decrease.
d	Initial approximation of cost function gradient.
g	Initial approximation of individual losses gradient.
covered	Matrix of covered samples.
standardize	Boolean. Scales the data if True
tol	Real. Miminal required approximate gradient norm before convergence.
family	One of "binomial", "gaussian", "exponential" or "poisson"
fit_alg	One of "constant", "linesearch" (default), or "adaptive".
user_loss_function	User supplied R or C loss and gradient functions
...	Any other pass-through parameters.

Value

object of class SAG

IshmaelBelghazi/bigoptim documentation built on May 7, 2019, 6:44 a.m.