How to optimize memory usage in dplyr + purrr

Question

I have an issue where after replicating data for a training and testing set, I'm showing a large amount of memory allocated to my user in Rstudio, but not being used in my R session. I've created a small example to reproduce my situation :)

This code runs a bunch of model, based on different formulas, algorithms, and parameter sets that I give it. It is a function, but I've created a simple script for reprex.

library(dplyr)
library(purrr)
library(modelr)
library(tidyr)
library(pryr)

# set my inputs
data <- mtcars
formulas <- c(test1 = mpg ~ cyl + wt + hp,
              test2 = mpg ~ cyl + wt)
params = list()
methods <- "lm"

n <- 20 # num of cv splits
mult <- 10 # number of times I want to replicate some of the data
frac <- .25 # how much I want to cut down other data (fractional)

### the next few chunks get the unique combos of the inputs.
if (length(params) != 0) {
  cross_params <- params %>% 
    map(cross) %>% 
    map_df(enframe, name = "param_set", .id = "method") %>% 
    list
} else cross_params <- NULL

methods_df <- tibble(method = methods) %>% 
  list %>% 
  append(cross_params)  %>% 
  reduce(left_join, by = "method") %>% 
  split(1:nrow(.))

# wrangle formulas into a split dataframe
formulas_df <- tibble(formula = formulas,
                      name = names(formulas)) %>% 
  split(.$name)

# split out the data into n random train-test combos
cv_data <- data %>% 
  crossv_kfold(n) %>% # rsample?
  mutate_at(vars(train:test), ~map(.x, as_tibble))

# sample out if needed
cv_data_samp <- cv_data %>%
  mutate(train = modify(train, 
                        ~ .x %>% 
                          split(.$gear == 4) %>% 
                          # take a sample of the non-vo data
                          modify_at("FALSE", sample_frac, frac) %>% 
                          # multiply out the vo-on data
                          modify_at("TRUE", function(.df) {
                            map_df(seq_along(1:mult), ~ .df) 
                          }) %>% 
                          bind_rows))

# get all unique combos of formula and method
model_combos <- list(cv = list(cv_data_samp), 
                     form = formulas_df, 
                     meth = methods_df) %>% 
  cross %>%
  map_df(~ bind_cols(nest(.x$cv), .x$form, .x$meth)) %>% 
  unnest(data, .preserve = matches("formula|param|value")) %>% 
  {if ("value" %in% names(.)) . else mutate(., value = list(NULL))} 

# run the models
model_combos %>% 
  # put all arguments into a single params column
  mutate(params = pmap(list(formula = formula, data = train), list)) %>%
  mutate(params = map2(params, value, ~ append(.x, .y))) %>%
  mutate(params = modify(params, discard, is.null)) %>%
  # run the models
  mutate(model = invoke_map(method, params))  

mem_change(rm(data, cv_data, cv_data_samp))
mem_used()

Now after I do this, my mem_used comes out to 77.3mb, but I'm seeing roughly double that (160Mb) allocated to my R user. This really explodes when my data is 3 Gb, which is my real-life case. I end up using 100Gb and tying up a whole server :(.

What is going on and how can I optimize?

Any help appreciated!!!

Answer 1

I figured this out! The issue was that I was converting my series of modelr resample objects to tibbles and that was exploding the memory even though I was subsequently sampling them down. The solution? Write methods for dealing with resample objects so that I never have to convert resample objects to tibble. These looked like:

# this function just samples the indexes instead of the data
sample_frac.resample <- function(data, frac) {
  data$idx <- sample(data$idx, frac * length(data$idx))
  data
}

# this function replicates the indexes. I should probably call it something else.
augment.resample <- function(data, n) {
  data$idx <- unlist(map(seq_along(1:n), ~ data$idx))
  data
}

# This function does simple splitting (logical only) of resample obejcts
split.resample <- function(data, .p) {
  pos <- list(data = data$data, idx = which(.p, 1:nrow(data$data)))
  neg <- list(data = data$data, idx = which(!.p, 1:nrow(data$data)))
  class(pos) <- "resample"
  class(neg) <- "resample"
  list("TRUE" = pos,
       "FALSE" = neg)
}

# This function takes the equivalent of a `bind_rows` for resample objects.
# Since bind rows does not call `useMethod` I had to call it something else
bind <- function(data) {
  out <- list(data = data[[1]]$data, idx = unlist(map(data, pluck, "idx")))
  class(out) <- "resample"
  out
}

Then I just converted to a tibble in the same purrr closure in which my model is run for that CV. Problem solved! My memory usage is VERY low now.