Merge pull request #18 from SchisslerGroup/develop

passes Windows installation and checks

.Rbuildignore

@@ -13,3 +13,4 @@
 ^docs$
 ^pkgdown$
 ^\.github$
+^LICENSE\.md$

DESCRIPTION

@@ -1,20 +1,23 @@
 Package: bigsimr
-Title: An R Package for Generating High-Dimensional Random Vectors
-Version: 0.8.0
+Title: Fast Generation of High-Dimensional Random Vectors
+Version: 0.9.0
 Authors@R:
     c(person(given = "Alex",
              family = "Knudson",
              role = c("aut", "cre"),
              email = "[email protected]"),
       person(given = "Grant",
              family = "Schissler",
+             role = "aut"),
+      person(given = "Duc",
+             family = "Tran",
              role = "aut"))
 Maintainer: Alex Knudson <[email protected]>
 Description: Simulate multivariate data with arbitrary marginal distributions.
 URL: https://github.com/SchisslerGroup/bigsimr
 BugReports: https://github.com/SchisslerGroup/bigsimr/issues
-Depends: R (>= 3.5.0)
-License: GPL-3 + file LICENSE
+Depends: R (>= 3.6.0)
+License: GPL-3
 Encoding: UTF-8
 LazyData: true
 RoxygenNote: 7.1.1
@@ -31,13 +34,11 @@ Imports:
     pcaPP,
     Rcpp,
     RcppArmadillo,
-    foreach,
     parallel,
-    doParallel,
     rlang,
-    assertthat,
     reticulate,
-    rstudioapi
+    rstudioapi,
+    mvnfast
 LinkingTo:
     Rcpp,
     RcppArmadillo

R/cor_fast.R

@@ -14,7 +14,9 @@ cor_fast <- function(x, y = NULL, method = c("pearson", "kendall", "spearman"))
     x <- as.matrix(x)
   }
 
-  storage.mode(x) <- "double"
+  storage.mode(x) <- "numeric"
+  if (!is.null(y))
+    storage.mode(y) <- "numeric"
 
   if (method == "pearson") {
     if (is.null(y)) {
@@ -42,6 +44,7 @@ cor_fast <- function(x, y = NULL, method = c("pearson", "kendall", "spearman"))
 
 }
 
+
 fastrank_num_avg <- function(x) {
   .Call("fastrank_num_avg_", x, PACKAGE = "bigsimr")
 }

R/cor_rand.R

@@ -51,14 +51,14 @@ cor_randPD <- function(d, a=1.0) {
 #' This method is generally faster than \code{\link{cor_randPD}}
 #'
 #' @param d A positive integer number of dimensions
-#' @param k A tuning parameter
+#' @param k A tuning parameter between 1 and d
 #' @importFrom stats rnorm runif cov2cor
 #' @export
 cor_randPSD <- function(d, k=d) {
   if (d == 1) {
     return(matrix(1, 1, 1))
   }
-  assertthat::assert_that(1 <= k && k <= d)
+  stopifnot(1 <= k && k <= d)
   W  <- matrix(rnorm(d * k), d, k)
   S  <- tcrossprod(W, W) + diag(runif(d))
   S2 <- diag(1 / sqrt(diag(S)))

R/rand_vec.R

@@ -4,27 +4,24 @@
   eval(margin)
 }
 
+#' Simulate correlated multivariate data
+#'
 #' Creates \code{n} observations from a multivariate distribution with the
-#' given marginals and correlation matrix.
+#'   given marginals and correlation matrix. The correlation matrix is always
+#'   assumed to be a Pearson correlation matrix
 #'
 #' @param n The number random vectors to generate.
-#' @param rho The input correlation matrix.
+#' @param rho The input (Pearson) correlation matrix.
 #' @param margins The marginal distributions (Typically R's "quantile functions")
-#' @param type The type of correlation matrix that is being passed (Assumed to
-#'    be Pearson by default).
+#' @param cores The number of cores to run on.
 #' @param ensure_PSD Ensure that the converted correlation matrix is positive
-#'    semi-definite. More import if the input correlation type is Kendall or
-#'    Spearman.
-#' @param cores The number of cores to run on
+#'    semi-definite.
 #' @return A matrix of random vectors generated from the specified marginals
 #'   and parameters.
 #' @export
-rvec <- function(n, rho, margins, type = c("pearson", "kendall", "spearman"),
-                 ensure_PSD=FALSE, cores = 1L){
+rvec <- function(n, rho, margins, cores = 1L, ensure_PSD = FALSE) {
 
-  type <- match.arg(type)
-  rho  <- cor_convert(rho, from = type, to = "pearson")
-  d    <- length(margins)
+  d <- length(margins)
 
   if (ensure_PSD)
     rho <- cor_nearPSD(rho)
@@ -35,27 +32,29 @@ rvec <- function(n, rho, margins, type = c("pearson", "kendall", "spearman"),
   }
 
   # generate multivariate uniform distribution (via Z -> U)
-  U <- .rmvuu(n, rho)
+  if (getOption("use_jax", FALSE)) {
+    U <- .rmvuu(n, rho)
+  } else {
+    U <- stats::pnorm(mvnfast::rmvn(n, rep(0, d), rho, cores))
+  }
 
   # Apply the copula algorithm
   d <- nrow(rho)
 
   if (cores <= 1L) {
 
-    rv <- sapply(1:d, function(i){
+    rv <- vapply(1:d, function(i){
       .u2m(U[,i], margins[[i]])
-    })
+    }, numeric(n))
 
 
   } else {
 
-    `%dopar%` <- foreach::`%dopar%`
-    cl <- parallel::makeCluster(cores, type = "FORK")
-    doParallel::registerDoParallel(cl)
+    cl <- parallel::makeCluster(spec = cores, type = "FORK")
 
-    rv <- foreach::foreach(i = 1:d, .combine = 'cbind') %dopar% {
+    rv <- parallel::parSapply(cl = cl, 1:d, function(i) {
       .u2m(U[,i], margins[[i]])
-    }
+    }, simplify = TRUE)
 
     parallel::stopCluster(cl)
 

R/zzz.R

@@ -11,4 +11,10 @@ numpy <- NULL
                                delay_load = TRUE,
                                convert = FALSE)
 
+  if (.Platform$OS.type == "windows" || !have_jax()) {
+    options(use_jax = FALSE)
+  } else {
+    options(use_jax = getOption("use_jax", TRUE))
+  }
+
 }

README.Rmd

@@ -17,7 +17,7 @@ knitr::opts_chunk$set(
 
 ### bigsimr is an R package for simulating high-dimensional multivariate data with arbitrary marginal distributions
 
-bigsimr lets you simulate multivariate data given a correlation matrix and a list of distributions. The correlation matrix can be of type Pearson, Spearman, or Kendall, and we use a matching algorithm to ensure that the estimated correlation of the simulated data is the same as the input correlation.
+bigsimr lets you simulate multivariate data given a correlation matrix and a list of distributions. The correlation matrix can be of type Pearson, Spearman, or Kendall.
 
 ### See the [website](https://schisslergroup.github.io/bigsimr/) for more information, including [installation instructions](https://schisslergroup.github.io/bigsimr/articles/install-bigsimr.html), [tutorials](https://schisslergroup.github.io/bigsimr/articles/using-rvec.html), and [package documentation](https://schisslergroup.github.io/bigsimr/reference/index.html).
 
@@ -35,7 +35,7 @@ devtools::install_github("SchisslerGroup/bigsimr", ref="develop")
 
 This package depends on
 [reticulate](https://rstudio.github.io/reticulate/) to draw on the speed
-of Google’s [jax](https://github.com/google/jax) library. Please see the [bigsimr installation instructions](#) for more details.
+of Google’s [jax](https://github.com/google/jax) library. Please see the [bigsimr installation instructions](https://schisslergroup.github.io/bigsimr/articles/install-bigsimr.html) for more details.
 
 ---
 

README.md

@@ -7,9 +7,7 @@
 
 bigsimr lets you simulate multivariate data given a correlation matrix
 and a list of distributions. The correlation matrix can be of type
-Pearson, Spearman, or Kendall, and we use a matching algorithm to ensure
-that the estimated correlation of the simulated data is the same as the
-input correlation.
+Pearson, Spearman, or Kendall.
 
 ### See the [website](https://schisslergroup.github.io/bigsimr/) for more information, including [installation instructions](https://schisslergroup.github.io/bigsimr/articles/install-bigsimr.html), [tutorials](https://schisslergroup.github.io/bigsimr/articles/using-rvec.html), and [package documentation](https://schisslergroup.github.io/bigsimr/reference/index.html).
 
@@ -29,7 +27,9 @@ devtools::install_github("SchisslerGroup/bigsimr", ref="develop")
 This package depends on
 [reticulate](https://rstudio.github.io/reticulate/) to draw on the speed
 of Google’s [jax](https://github.com/google/jax) library. Please see the
-[bigsimr installation instructions](#) for more details.
+[bigsimr installation
+instructions](https://schisslergroup.github.io/bigsimr/articles/install-bigsimr.html)
+for more details.
 
 -----
 

pkgdown/_pkgdown.yml

@@ -9,8 +9,8 @@ authors:
     href: https://github.com/adknudson
   Grant Schissler:
     href: http://www.grantschissler.com/
-  Richard Foote:
-    href: https://github.com/rdlfoote
+  Duc Tran:
+    href: https://github.com/duct317
 
 reference:
   - title: Primary Functions
@@ -39,5 +39,6 @@ articles:
 - title: Articles
   navbar: Get Started
   contents:
+  - bigsimr
   - install-bigsimr
   - using-rvec

pkgdown/index.md

@@ -11,5 +11,30 @@ Status](https://travis-ci.com/SchisslerGroup/bigsimr.svg?branch=master)](https:/
 
 `bigsimr` is an R package for simulating high-dimensional multivariate data with arbitrary marginal distributions
 
-bigsimr lets you simulate multivariate data given a correlation matrix and a list of distributions. The correlation matrix can be of type Pearson, Spearman, or Kendall, and we use a matching algorithm to ensure that the estimated correlation of the simulated data is the same as the input correlation.
+bigsimr lets you simulate multivariate data given a correlation matrix and a list of distributions. The correlation matrix can be of type Pearson, Spearman, or Kendall.
 
+You can install the release version of the package from GitHub:
+
+```r
+devtools::install_github("SchisslerGroup/bigsimr")
+```
+
+To get a bug fix or to use a new feature, you can install the development version from GitHub:
+
+```r
+devtools::install_github("SchisslerGroup/bigsimr", ref="develop")
+```
+
+This package depends on
+[reticulate](https://rstudio.github.io/reticulate/) to draw on the speed
+of Google’s [jax](https://github.com/google/jax) library. Please see the [bigsimr installation instructions](https://schisslergroup.github.io/bigsimr/articles/install-bigsimr.html) for more details.
+
+If on Windows or a system without Python, then the package will default to alternative methods. The option can be toggled with
+
+```r
+options(use_jax = FALSE)
+```
+
+Currently Google's jax library does not have ready-to-use binaries for Windows. The recommendation is to use Windows Subsystem for Linux (WSL), otherwise you will need to build `jaxlib` from source ([see here](https://jax.readthedocs.io/en/latest/developer.html#additional-notes-for-building-jaxlib-from-source-on-windows)).
+
+Additionally, Windows does not allow for forked multiprocessing, so there will be no performance enhancement on a multicore Windows machine.

tests/testthat.R

@@ -1,5 +1,3 @@
-reticulate::py_config()
-
 library(testthat)
 library(bigsimr)
 

tests/testthat/test_cor_convert.R

@@ -1,4 +1,5 @@
 test_that("cor_convert works for scalars, vectors, and matrices", {
+
   S <- "spearman"
   K <- "kendall"
   P <- "pearson"
@@ -48,4 +49,5 @@ test_that("cor_convert works for scalars, vectors, and matrices", {
   expect_equal(cor_convert(-1.0, P, K), -1.0)
   expect_equal(cor_convert(-1.0, P, P), -1.0)
   expect_equal(cor_convert(-1.0, S, S), -1.0)
+
 })

tests/testthat/test_cor_nearPSD.R

@@ -1,4 +1,5 @@
 test_that("cor_nearPSD returns a positive semidefinite correlation matrix", {
+
   # define a negative definite correlation matrix
   p <- matrix(c(
     1.00, 0.82, 0.56, 0.44,
@@ -10,13 +11,14 @@ test_that("cor_nearPSD returns a positive semidefinite correlation matrix", {
   r <- cor_nearPSD(p)
   e <- eigen(r)
 
-  # Correlation matrices must be:
-  # 1. symmetric
+  # Correlation matrices must:
+  # 1. be symmetric
   # 2. have ones on diagonal
-  # 3. be positive semidefinite
+  # 3. be positive semi-definite
   # 4. have all values in the domain [-1, 1]
   expect_true(Matrix::isSymmetric(r))
   expect_true(all(diag(r) == 1))
   expect_true(all(e$values >= 0))
   expect_true(all((-1 <= r) & (r <= 1)))
+
 })

tests/testthat/test_cor_rand.R

@@ -1,4 +1,5 @@
 test_that("cor_randPD and cor_randPSD generate random correlation matrices", {
+
   # Correlation matrices must be:
   # 1. symmetric
   # 2. have ones on diagonal
@@ -19,4 +20,5 @@ test_that("cor_randPD and cor_randPSD generate random correlation matrices", {
   expect_true(all(diag(rho_PSD) == 1))
   expect_true(all(e$values >= 0))
   expect_true(all((-1 <= rho_PSD) & (rho_PSD <= 1)))
+
 })

tests/testthat/test_jax_rand.R

@@ -1,4 +1,7 @@
 test_that("jax_rmvn generates random multivariate normal data", {
+
+  skip_if_not(have_jax(), "jax is not available for testing")
+
   # Should fail for negative definite matrices
   nd_cor <- matrix(c(
     1.00, 0.82, 0.56, 0.44,