Download all datasets contained in all R-packages

[giuseppec]

We can do something like (ugly code) and then upload everything

# install all packages
# pkg = available.packages()
# for(i in 1:nrow(pkg)) install.packages(pkgs = pkg[i,"Package"])

# get table of all available data sets from all available packages
rdat = data(package = .packages(all.available = TRUE))
rdat = as.data.frame(rdat$results)
rdat$Package.Version = sapply(rdat$Package, function(x) as.character(packageVersion(x)))
# here we remove strange data set names
rdat = rdat[!grepl("\\(|\\)", rdat$Item),]

rdat.unique = unique(rdat[,c("Package", "Package.Version")])
ret = setNames(vector("list", nrow(rdat.unique)), 
  paste0(rdat.unique$Package, "_", rdat.unique$Package.Version))

for(i in seq_along(rdat.unique$Package)) {
  # get all data set names from package 'i'
  dat.names = as.character(subset(rdat, Package == rdat.unique$Package[i])[,"Item"])
  data(list = dat.names, package = as.character(rdat.unique$Package[i]))

  ret[[i]] = setNames(lapply(dat.names, function(dn) {
    return(tryCatch(get(dn), error = function(e) e, warning = function(w) w))
  }), dat.names)

  loaded.pkg = setdiff(loadedNamespaces(),
    c("stats", "graphics", "grDevices", "utils", "datasets", "methods", "base", "tools"))
  lapply(loaded.pkg, function(x) try(unloadNamespace(x)))

  cat(as.character(rdat.unique$Package[i]), ": all data sets downloaded", fill = TRUE)
}

[HeidiSeibold]

I asked on twitter if there are ways to do this without having to install the packages.

This is the best answer I got:
https://twitter.com/GaborCsardi/status/725776034910617600

Seems pretty promising 😃

[HeidiSeibold]

http://vincentarelbundock.github.io/Rdatasets/datasets.html

[jakobbossek]

Thanks Heidi!
Tried the approach proposed by Gabor. Pretty easy with his gh package and githubs's fantastic code search API to obtain a list of all rda files inside cran:

devtools::install_github("gaborcsardi/gh")
library(gh)
repos = gh("GET /search/code?q=user:cran+extension:rda")
catf("#Repos: %i", repos$total_count)

This way we can download the rda files only, e.g., via repos$items[[i]]$html_url. However, there is no easy way to access meta data for the data set, e.g., data set description, default target feature, citations etc. An possibility is to download the corresponding Rd docs as well an parse these. Uploading stuff to OpenML without at least a meaningful description seems useless to me.

Another point: why should we avoid downloading all packages by the crawler. Is it because of time and memory? We can simply download each package, extract the data sets, upload to OpenML and remove the package afterwards. The time aspect is unimportant. The crawler does not need to be fast.

[jakobbossek]

Started to work on a crawler which operates on the github cran repositories and reads 1) the data itself and 2) metadata from the corresponding Rd file. Works well so far. Just need to parallelize stuff and handle potential errors.

[cvitolo]

@jakobbossek I'd love to see the results of your crawler/experiment. Did you publish it?

[giuseppec]

A huge collection can be found here http://vincentarelbundock.github.io/Rdatasets/datasets.html