Segmentor3IsBack - rstats-gsoc/gsoc2018 GitHub Wiki

The Segmentor3IsBack package provides a fast functional pruning algorithm for computing the optimal changepoints using several likelihood models (normal heteroscedastic, normal homoscedastic, Poisson, etc).

However there are several issues with the current code:

• it ERRORs on Solaris.
• it has no tests.
• packages which use it have been known to crash on windows. for example penaltyLearning uses Segmentor in its vignette, which crashes when re-building on windows.

The changepoint package provides cpt.mean for the normal homoscedastic model, and cpt.var for for the normal heteroscedastic model, but does not provide a solver for the other models (e.g. Poisson).

The jointseg package provides the Fpsn function which computes the normal homoscedastic model.

Get Segmentor3IsBack passing all CRAN checks on solaris/windows.

• Setup a github repo for Segmentor3IsBack, with TravisCI for GNU/Linux testing, Appveyor for windows testing and Coveralls for code coverage.
• Add some trivial edge cases as tests, for example Segmentor(c(1,2,2), Kmax=3, model=1) should return a valid 3-segment model, but it currently does not.
• Write some extensive test cases using library(testthat) and library(neuroblastoma). Goal: 100% coverage in both R and C++ code by the end of summer.
• Can test on windows via win-builder.
• Compile package with GCC 6, and fix warnings report from CRAN repository.
• Figure out a way to access a solaris machine for testing.

This project will increase the portability and test coverage of the Segmentor3IsBack package.

• Alice Cleynen <[email protected]>, one of the authors of Segmentor3IsBack who knows its R/C++ code.
• Toby Dylan Hocking <[email protected]> is a user of Segmentor3IsBack, is familiar with Travis/Coveralls, and can suggest some tests.

MENTORS: please think of some tests for prospective students.

• Easy: create an Rmd web page in which you demonstrate how Segmentor can be used to find change points in different types of data (e.g. normal model for real-valued data, Poisson model for count data). For some example data sets, see data(neuroblastoma, package=”neuroblastoma”) for real-valued data, and data(chr11ChIPseq, package=”PeakSegDP”) for count data. Plot the data and the segmentation models.
• Medium: copy the Segmentor3IsBack package to one of your GitHub repos. Setup Travis, Appveyor and Coveralls, and create a README with badges.
• Hard: Write a test that fails on windows. Show that the package fails on both Appveyor and win-builder.

Students, please post a link to your test results here.