| Type: | Package |
| Title: | Parallel GLM |
| Version: | 0.1.7 |
| Description: | Provides a parallel estimation method for generalized linear models without compiling with a multithreaded LAPACK or BLAS. |
| License: | GPL-2 |
| Encoding: | UTF-8 |
| URL: | https://github.com/boennecd/parglm |
| BugReports: | https://github.com/boennecd/parglm/issues |
| LinkingTo: | Rcpp, RcppArmadillo |
| Imports: | Rcpp, Matrix |
| SystemRequirements: | C++11 |
| Suggests: | testthat, SuppDists, knitr, rmarkdown, speedglm,microbenchmark, R.rsp |
| RoxygenNote: | 6.1.1 |
| VignetteBuilder: | R.rsp |
| NeedsCompilation: | yes |
| Packaged: | 2021-10-14 14:55:16 UTC; boennecd |
| Author: | Benjamin Christoffersen [cre, aut], Anthony Williams [cph], Boost developers [cph] |
| Maintainer: | Benjamin Christoffersen <boennecd@gmail.com> |
| Repository: | CRAN |
| Date/Publication: | 2021-10-14 15:10:02 UTC |
Fitting Generalized Linear Models in Parallel
Description
Function likeglm which can make the computationin parallel. The function supports most families listed infamily.See "vignette("parglm", "parglm")" for run time examples.
Usage
parglm(formula, family = gaussian, data, weights, subset, na.action, start = NULL, offset, control = list(...), contrasts = NULL, model = TRUE, x = FALSE, y = TRUE, ...)parglm.fit(x, y, weights = rep(1, NROW(x)), start = NULL, etastart = NULL, mustart = NULL, offset = rep(0, NROW(x)), family = gaussian(), control = list(), intercept = TRUE, ...)Arguments
formula | an object of class |
family | a |
data | an optional data frame, list or environment containing the variablesin the model. |
weights | an optional vector of 'prior weights' to be used in the fitting process. Shouldbe |
subset | an optional vector specifying a subset of observations to be used inthe fitting process. |
na.action | a function which indicates what should happen when the data contain |
start | starting values for the parameters in the linear predictor. |
offset | this can be used to specify an a priori known component to beincluded in the linear predictor during fitting. |
control | a list of parameters for controlling the fitting process.For parglm.fit this is passed to |
contrasts | an optional list. See the |
model | a logical value indicating whether model frame should be includedas a component of the returned value. |
x,y | For For |
... | For For |
etastart | starting values for the linear predictor. Not supported. |
mustart | starting values for the vector of means. Not supported. |
intercept | logical. Should an intercept be included in the null model? |
Details
The current implementation usesmin(as.integer(n / p), nthreads)threads wheren is the number observations,p is thenumber of covariates, andnthreads is thenthreads element ofthe listreturned byparglm.control. Thus, there is likely little (ifany) reduction in computation time ifp is almost equal ton.The current implementation cannot handlep > n.
Value
glm object as returned byglm but differs mainly by theqrelement. Theqr element in the object returned byparglm(.fit) only has theRmatrix from the QR decomposition.
Examples
# small example from `help('glm')`. Fitting this model in parallel does# not matter as the data set is smallclotting <- data.frame( u = c(5,10,15,20,30,40,60,80,100), lot1 = c(118,58,42,35,27,25,21,19,18), lot2 = c(69,35,26,21,18,16,13,12,12))f1 <- glm (lot1 ~ log(u), data = clotting, family = Gamma)f2 <- parglm(lot1 ~ log(u), data = clotting, family = Gamma, control = parglm.control(nthreads = 2L))all.equal(coef(f1), coef(f2))Auxiliary for Controlling GLM Fitting in Parallel
Description
Auxiliary function forparglm fitting.
Usage
parglm.control(epsilon = 1e-08, maxit = 25, trace = FALSE, nthreads = 1L, block_size = NULL, method = "LINPACK")Arguments
epsilon | positive convergence tolerance. |
maxit | integer giving the maximal number of IWLS iterations. |
trace | logical indicating if output should be produced doing estimation. |
nthreads | number of cores to use. You may get the best performance byusing your number of physical cores if your data set is sufficiently large.Using the number of physical CPUs/cores may yield the best performance(check your number e.g., by calling |
block_size | number of observation to include in each parallel block. |
method | string specifying which method to use. Either |
Details
TheLINPACK method uses the same QR method asglm.fit for the final QR decomposition.This is thedqrdc2 method described inqr. All other QRdecompositions but the last are made withDGEQP3 fromLAPACK.See Wood, Goude, and Shaw (2015) for details on the QR method.
TheFAST method computes the Fisher information and then solves the normalequation. This is faster but less numerically stable.
Value
A list with components named as the arguments.
References
Wood, S.N., Goude, Y. & Shaw S. (2015) Generalized additive models forlarge datasets. Journal of the Royal Statistical Society, Series C64(1): 139-155.
Examples
# use one coreclotting <- data.frame( u = c(5,10,15,20,30,40,60,80,100), lot1 = c(118,58,42,35,27,25,21,19,18), lot2 = c(69,35,26,21,18,16,13,12,12))f1 <- parglm(lot1 ~ log(u), data = clotting, family = Gamma, control = parglm.control(nthreads = 1L))# use two coresf2 <- parglm(lot1 ~ log(u), data = clotting, family = Gamma, control = parglm.control(nthreads = 2L))all.equal(coef(f1), coef(f2))