Cleanup MLJ interface - Githubissues

davidbp / MulticlassPerceptron.jl

MulticlassPerceptron.jl

GNU General Public License v3.0

3 stars 2 forks source link

Cleanup MLJ interface #4

Closed ablaom closed 3 years ago

ablaom commented 3 years ago

This PR:

dumps some redundant dependencies: DataFrames, RDatasets
updates some [compat] entries to get pkg working with latest MLJBase (in particular bumps Tables [compat] which was very old)
some tidy up, including adding _reformat method to address table/matrix distinction using type dispatch
extends input_scitype to allow MLJ user to supply matrix instead of table. However, we note in passing that such a table is expected to features-as-rows, which could be confusing, since tables in MLJ are always features-as-columns. Moreover, it is possible MLJ will make features-as-columns a firm requirement for matrix input as well. I suggest a remedy in #5

ablaom commented 3 years ago

I've also add a github action for continuous integration. Travis is no longer free. Possibly it won't work until this gets merged onto master. If you don't want this integration, rm .github/workflows/ci.yml .

This branch passes tests locally for me.

davidbp commented 3 years ago

Thank you for your update, I see we still need to explicitly call MLJBase.fit in the tests. It would be more elegant to just use a fit and it seems it should be possible because the input types of the fit change. One is a MulticlassPerceptronCore and the other one is a MulticlassPerceptronClassifier.