rropenscidrake-r-package

How to create a plan target to cross over the results of previous map targets and a new variable?

From multiple target (a) created with map I have 2 other targets (b and d) that iterate over the first target. Now I would like to use the results of these targets in another target. In addition I would like to cross with another variable (model).

I pasted a reprex below, but for some context in my case a describe different subset of a dataset, b and d pre-compute some stuff, e apply different models on each subset using the pre-computed data.

I tried different combination of map cross (like e below) without success. I tried to add all the targets names I want to use after in fn4 but it creates unnecessary crosses.

library(drake)
drake_plan(
  a = target(
    fn1(arg1, arg2),
    transform = map(
      arg1 = !!c("arg11", "arg12"),
      arg2 = !!c("arg21", "arg22")
    )
  ),
  b = target(
    fn2(arg1),
    transform = map(arg1)
  ),
  d = target(
    fn3(arg1),
    transform = map(arg1)
  ),
  e = target(
    fn4(b, d, model, arg1),
    transform = cross(
      b,
      d,
      model = !!c("x", "y", "z"),
      .by = arg1,
      .id = c(arg1, model)
    )
  ),
  trace = TRUE
)
#> # A tibble: 18 x 10
#>    target   command     arg1    arg2   a      b     d     model .by   e    
#>    <chr>    <expr>      <chr>   <chr>  <chr>  <chr> <chr> <chr> <chr> <chr>
#>  1 a_arg11… fn1("arg11… "\"arg… "\"ar… a_arg… <NA>  <NA>  <NA>  <NA>  <NA> 
#>  2 a_arg12… fn1("arg12… "\"arg… "\"ar… a_arg… <NA>  <NA>  <NA>  <NA>  <NA> 
#>  3 b_arg11  fn2("arg11… "\"arg… "\"ar… a_arg… b_ar… <NA>  <NA>  <NA>  <NA> 
#>  4 b_arg12  fn2("arg12… "\"arg… "\"ar… a_arg… b_ar… <NA>  <NA>  <NA>  <NA> 
#>  5 d_arg11  fn3("arg11… "\"arg… "\"ar… a_arg… <NA>  d_ar… <NA>  <NA>  <NA> 
#>  6 d_arg12  fn3("arg12… "\"arg… "\"ar… a_arg… <NA>  d_ar… <NA>  <NA>  <NA> 
#>  7 e_NA_x   fn4(b_arg1… <NA>    <NA>   <NA>   b_ar… d_ar… "\"x… arg1  e_NA…
#>  8 e_NA_y   fn4(b_arg1… <NA>    <NA>   <NA>   b_ar… d_ar… "\"y… arg1  e_NA…
#>  9 e_NA_z   fn4(b_arg1… <NA>    <NA>   <NA>   b_ar… d_ar… "\"z… arg1  e_NA…
#> 10 e_NA_x_2 fn4(b_arg1… <NA>    <NA>   <NA>   b_ar… d_ar… "\"x… arg1  e_NA…
#> 11 e_NA_y_2 fn4(b_arg1… <NA>    <NA>   <NA>   b_ar… d_ar… "\"y… arg1  e_NA…
#> 12 e_NA_z_2 fn4(b_arg1… <NA>    <NA>   <NA>   b_ar… d_ar… "\"z… arg1  e_NA…
#> 13 e_NA_x_3 fn4(b_arg1… <NA>    <NA>   <NA>   b_ar… d_ar… "\"x… arg1  e_NA…
#> 14 e_NA_y_3 fn4(b_arg1… <NA>    <NA>   <NA>   b_ar… d_ar… "\"y… arg1  e_NA…
#> 15 e_NA_z_3 fn4(b_arg1… <NA>    <NA>   <NA>   b_ar… d_ar… "\"z… arg1  e_NA…
#> 16 e_NA_x_4 fn4(b_arg1… <NA>    <NA>   <NA>   b_ar… d_ar… "\"x… arg1  e_NA…
#> 17 e_NA_y_4 fn4(b_arg1… <NA>    <NA>   <NA>   b_ar… d_ar… "\"y… arg1  e_NA…
#> 18 e_NA_z_4 fn4(b_arg1… <NA>    <NA>   <NA>   b_ar… d_ar… "\"z… arg1  e_NA…

Created on 2019-07-15 by the reprex package (v0.3.0)

It seems to work, but the arg1 and arg2 are not carried over and are not usable in fn4 and following targets. Should I split this step in 2 steps, if so how? (map then cross, cross then map?) I tried to cross earlier, after a, but I don't wont to recompute identical b and d multiple times, it may take a lot of time and memory.

Edit: A more realistic example

because many targets use the same data that need to be saved as file for the run function (call to an external binaries), so to prevent re-computing the same thing multiple times and to save multiple times the same thing in different files (taht can be huge) I seperated all these tasks in Drake.


library(drake)
library(tibble)
library(dplyr)
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union

path_data <- c("path/data_1.csv", "path/data_2.csv")
countries <- c("1", "2")
analysis_dir <- "path"
substudies_1 <- tribble(
  ~substudy, ~adjust, ~sex,
  "sub1", "no", "male/female",
  "sub2", "yes", "male/female"
)
models <- c("x", "y")

plan <- drake_plan(
  data = target(
    get_data(file_in(path)),
    transform = map(path = !!path_data, country = !!countries, .id = country)
  ),
  SNP = target(
    get_SNP_data_country(SNP_gene, data),
    transform = map(data, .id = country)
  ),
  map = target(
    # actually write file and save path
    write_snp_map(SNP, file.path(analysis_dir, country, "SNP_map.txt")),
    transform = map(SNP, .id = country)
  ),
  ref = target(
    # actually write file and save path
    write_snp_ref(SNP, file.path(analysis_dir, country, "SNP_ref.txt")),
    transform = map(SNP, .id = country)
  ),
  # data_2 is managed in another target because it has a different set of substudies,
  # this maybe can be tidied up, a problem for another day...
  population_1 = target(
    extract_population(data, sex, adjust),
    transform = map(
      data = data_1,
      country = "1",
      .data = !!substudies_1,
      .id = c(substudy)
    ),
  ),
  pedigree_1 = target(
    extract_pedigree(data_1, population_1),
    transform = map(
      population_1,
      .id = substudy
    )
  ),
  covariable_1 = target(
    extract_covariable(data_1, population_1, adjust, sex),
    transform = map(
      population_1,
      .id = substudy
    )
  ),
  # run_1 = target(
  #   run_fn(map_1, ref_1, pedigree_1, covariable_1, substudy, model, adjust, sex),
  #   transform = cross(population_1, model = !!models)
  # ),
  trace = TRUE
)

# the desired plan for the run target
run_plan <- tibble(
  target = c("run_1_x_population_1_sub1", "run_1_y_population_1_sub1", "run_1_x_population_1_sub2", "run_1_y_population_1_sub2"),
  command = list(
    expr(run(map_1, ref_1, pedigree_1_sub1, covariable_1_sub1, "x", "sub1", "no")),
    expr(run(map_1, ref_1, pedigree_1_sub1, covariable_1_sub1, "y", "sub1", "no")),
    expr(run(map_1, ref_1, pedigree_1_sub2, covariable_1_sub2, "x", "sub2", "yes")),
    expr(run(map_1, ref_1, pedigree_1_sub2, covariable_1_sub2, "y", "sub2", "yes"))
  ),
  path = NA_character_,
  country = "1",
  population_1 = c(rep("population_1_sub1", 2), rep("population_1_sub2", 2)),
  substudy = c(rep("sub1", 2), rep("sub2", 2)),
  adjust = c(rep("no", 2), rep("yes", 2)),
  sex = c(rep("male/female", 4)),
  pedigree_1 = c(rep("pedigree_1_sub1", 2), rep("pedigree_1_sub2", 2)),
  covariable_1 =  c(rep("covariable_1_sub1", 2), rep("covariable_1_sub2", 2)),
  model = c("x", "y", "x", "y"),
  SNP = "SNP_1",
  map = "map_1",
  ref = "ref_1"
)

config <- drake_config(bind_rows(plan, run_plan))
vis_drake_graph(config, targets_only = TRUE)

Created on 2019-07-15 by the reprex package (v0.3.0)

plan: i.imgur.com/MyqoKJi.png

Edit 2:

I now use the .data parameter in a map transform using a dataframe with previous target names (using rlang::syms) it works fine except that it doesn't work with drake::drake_plan's max_expand parameter. This solution is not optimal also because crafting a grid for .data is very verbose.

Solution

  • Would you mind explicitly posting the plan you want without any transforms? drake_plan_source() can help.

    One note: only combine() understands .by. Maybe another approach is to use transform = map(.data = !!your_grid_of_combinations): https://ropenscilabs.github.io/drake-manual/plans.html#map.

    Does the plan you want look something like this?

    library(drake)
plan <- drake_plan(
  a = target(
    fn1(arg1, arg2),
    transform = map(
      arg1 = !!c("arg11", "arg12"),
      arg2 = !!c("arg21", "arg22")
    )
  ),
  b = target(
    fn2(arg1),
    transform = map(arg1)
  ),
  d = target(
    fn3(arg1),
    transform = map(arg1)
  ),
  e = target(
    fn4(b, d, model, arg1),
    transform = cross(
      b,
      d,
      model = c("x", "y", "z"),
      arg1,
      .id = c(arg1, model)
    )
  )
)

config <- drake_config(plan)
vis_drake_graph(config)

    Created on 2019-07-15 by the reprex package (v0.3.0)