如何减小回归对象的大小适合保存到.rds时大小爆炸的函数

Question

我正在使用 rstanarm 将 stan_glm 模型拟合到函数中。遇到一个问题，即保存的 stanfit 对象的大小在保存到 .rds 时会爆炸，但仅当模型适合函数时。问题似乎是 stanfit 对象正在存储本地环境的副本，然后使用 write_rds 将其保存到磁盘？手动删除函数内的大对象或多或少地解决了这个问题，但这是一个相当笨拙的解决方案，所以想知道是否有人建议以更优雅的方式解决这个问题？下面的玩具 reprex（警告这会将一些 .rds 文件写入磁盘，我在示例结束时将它们删除但请注意）

library(readr)
library(rstanarm)
#> Loading required package: Rcpp
#> rstanarm (Version 2.19.3, packaged: 2020-02-11 05:16:41 UTC)
#> - Do not expect the default priors to remain the same in future rstanarm versions.
#> Thus, R scripts should specify priors explicitly, even if they are just the defaults.
#> - For execution on a local, multicore CPU with excess RAM we recommend calling
#> options(mc.cores = parallel::detectCores())
#> - bayesplot theme set to bayesplot::theme_default()
#>    * Does _not_ affect other ggplot2 plots
#>    * See ?bayesplot_theme_set for details on theme setting
library(gapminder)

# create a largeish object
test <- matrix(data = rnorm(10000), nrow = 10000/2, ncol = 10000/2)

# fit model in the global environment

a = stan_glm(lifeExp ~ gdpPercap, data = gapminder, refresh =0)

print(object.size(a), unit = "Mb")
#> 1.4 Mb


# fit model inside function , passing but not using largeish object
memfoo <- function(gap, testy, clean = FALSE){

  d <- testy

  if (clean){
    rm(d,testy)
  }

  a <- stan_glm(lifeExp ~ gdpPercap, data = gap, refresh = 0)

}

b <- memfoo(gapminder, test)

# fit model again, but removing large obects from the environment before running
d <- memfoo(gapminder, test, clean = TRUE)

print(object.size(a), unit = "Mb")
#> 1.4 Mb

print(object.size(b), unit = "Mb")
#> 1.4 Mb

print(object.size(d), unit = "Mb")
#> 1.4 Mb

# all same size in memory

# write to .rds

write_rds(a,"a.rds")

write_rds(b,"b.rds")

write_rds(d,"d.rds")

# rstan object run in function with largeish object in environment is 45 times bigger than same regression 
# fit outside function!
file.size("a.rds")
#> [1] 9026456

file.size("b.rds")
#> [1] 410011317

file.size("d.rds")
#> [1] 10011197

file.size("b.rds") / file.size("a.rds")
#> [1] 45.42329

file.remove(c("a.rds", "b.rds", "d.rds"))
#> [1] TRUE TRUE TRUE

^{由reprex package (v0.3.0) 于 2020-03-11 创建}

Answer 1

结果与here 发布的解决方案相同，为函数内部的回归指定一个新环境就可以了。

library(tidyverse)
library(rstanarm)
#> Loading required package: Rcpp
#> rstanarm (Version 2.19.3, packaged: 2020-02-11 05:16:41 UTC)
#> - Do not expect the default priors to remain the same in future rstanarm versions.
#> Thus, R scripts should specify priors explicitly, even if they are just the defaults.
#> - For execution on a local, multicore CPU with excess RAM we recommend calling
#> options(mc.cores = parallel::detectCores())
#> - bayesplot theme set to bayesplot::theme_default()
#>    * Does _not_ affect other ggplot2 plots
#>    * See ?bayesplot_theme_set for details on theme setting
library(gapminder)
library(butcher)

# create a largeish object
test <- matrix(data = rnorm(10000), nrow = 10000/2, ncol = 10000/2)

# fit model in the global environment

a = stan_glm(lifeExp ~ gdpPercap, data = gapminder, refresh =0)

# fit model inside function , passing but not using largeish object
memfoo <- function(gap, testy, clean = FALSE){

  d <- testy

  if (clean){


  env <- new.env(parent = .GlobalEnv)

  env$gap <- gap

  a <- with(env,{stan_glm(lifeExp ~ gdpPercap, data = gap, refresh = 0)})

  } else {

    a <- stan_glm(lifeExp ~ gdpPercap, data = gap, refresh = 0)
  }
}

b <- memfoo(gapminder, test)

d <- memfoo(gapminder, test, clean = TRUE)

butcher::weigh(a)
#> # A tibble: 50 x 2
#>    object              size
#>    <chr>              <dbl>
#>  1 stanfit           3.87  
#>  2 residuals         0.123 
#>  3 data.country      0.017 
#>  4 fitted.values     0.0143
#>  5 linear.predictors 0.0143
#>  6 y                 0.0143
#>  7 model.lifeExp     0.0137
#>  8 model.gdpPercap   0.0137
#>  9 data.lifeExp      0.0137
#> 10 data.gdpPercap    0.0137
#> # … with 40 more rows

butcher::weigh(b)
#> # A tibble: 50 x 2
#>    object                size
#>    <chr>                <dbl>
#>  1 terms             204.    
#>  2 formula           204.    
#>  3 stanfit             3.87  
#>  4 residuals           0.123 
#>  5 data.country        0.017 
#>  6 fitted.values       0.0143
#>  7 linear.predictors   0.0143
#>  8 y                   0.0143
#>  9 model.lifeExp       0.0137
#> 10 model.gdpPercap     0.0137
#> # … with 40 more rows

butcher::weigh(d)
#> # A tibble: 50 x 2
#>    object              size
#>    <chr>              <dbl>
#>  1 stanfit           3.87  
#>  2 residuals         0.123 
#>  3 terms             0.0700
#>  4 formula           0.0677
#>  5 data.country      0.017 
#>  6 fitted.values     0.0143
#>  7 linear.predictors 0.0143
#>  8 y                 0.0143
#>  9 model.lifeExp     0.0137
#> 10 model.gdpPercap   0.0137
#> # … with 40 more rows

write_rds(a, path = "a.rds")

write_rds(b, path = "b.rds")

write_rds(d, path = "d.rds")

file.size("a.rds")
#> [1] 9110178

file.size("b.rds")
#> [1] 410171588

file.size("d.rds")
#> [1] 9167588

file.size("d.rds") / file.size("a.rds")
#> [1] 1.006302

file.remove(c("a.rds", "b.rds", "d.rds"))
#> [1] TRUE TRUE TRUE

^{由reprex package (v0.3.0) 于 2020 年 3 月 12 日创建}