Mlpack: Port to MLJ.jl

[azev77]

Hi and thank you for this package!
Is it possible to port your models to MLJ.jl?

[rcurtin]

Hi @azev77, what would need to be done for that? I agree it would definitely be cool, the only problem is that the interface we can provide to Julia is somewhat limited because it's automatically generated. The main drawback is that models are "black boxes"; i.e., if I train a logistic regression model in Julia, I can predict new points with it, but I can't look at the coefficients of the model. This is restrictive but it's a requirement, unfortunately, because we can't provide arbitrary introspection to C++ types to all of the languages we bind to via our automatic binding generation system.

Sorry if the paragraph above is a little dense, but in any case, if we provide the equivalent of a "fit" and "predict" function, would it then be possible to wrap easily into MLJ? (I don't think we could generate that automatically, but a "downstream" package that wraps mlpack.jl into MLJ.jl could be manually maintained, perhaps.)

[mlpack-bot[bot]]

This issue has been automatically marked as stale because it has not had any recent activity. It will be closed in 7 days if no further activity occurs. Thank you for your contributions! :+1:

[azev77]

@rcurtin porting a package to MLJ usually involves adding one script to the repo.
For example: ParallelKmeans.jl does this in mlj_interface.jl.

[rcurtin]

@azev77 thanks for the clarification. That does seem not too hard, although our automatic binding generation system would make it tricky, so it may be a while until this can be done. However I will mark this keep-open, because at least for the work I am doing for my company I think it would be useful, and I also think it would be good overall to provide an MLJ interface. :+1:

[azev77]

@tlienart @ablaom
FYI: mlpack can be accessed via mlpack.jl, w/ Pkg.add("mlpack").
Here is a link to the Julia quickstart.