Skip to content

sperrorest v2.0.0 (12-Jun-2017)

Compare
Choose a tag to compare
@pat-s pat-s released this 13 Jun 19:28

Major:

  • integration of parsperrorest() into sperrorest().
  • by default, sperrorest() now runs in parallel using all available cores.
  • runfolds() and runreps() are now doing the heavy lifting in the background. All modes are now running on the same code base. Before, all parallel modes were running on different code implementations.
  • function and argument name changes to 'snake_case'

Features:

  • new (parallel) modes:
    • apply: calls pbmclapply() on Unix and pbapply() on Windows.
    • future: calls future_lapply() with various future options (multiprocess, multicore, etc.).
    • foreach: foreach() with various future options (multiprocess, multicore, etc.). Default option to cluster. This is also the overall default mode for sperrorest().
    • sequential: sequential execution using future backend.
  • RMSE instead of MSE as error measure
  • You can now pass also single values to repetition argument of sperrorest(). Specifying a range like repetition = 1:10 will also stay valid.
  • New vignette sperrorest::parallel-modes comparing the various parallel modes.
  • New vignette sperrorest::custom-pred-and-model-functions explaining why and how custom defined model and predict functions are needed for some model setups.

Misc:

  • Limit workers to number of repetitions if number of cores > number of repetitions. This ensures that no unnecessary workers are started and increases robustness of parallel execuction.
  • documentation improvements.
  • do_try argument has been removed.
  • error.fold, error.rep and err.train arguments have been removed because they are all calculated by default now.

Bugfixes:

  • partial matching of arguments
  • account for factor levels only present in test data but missing in training data. Previously, sperrorest errored during the predict step when this case occured. Now, this is accounted for and an informative message is given.