Tight integration with REP

[arogozhnikov]

In the nearest future, hep_ml will be cleaned and integrated with REP.

Prospect

hep_ml will be an extension of scikit-learn, which will follow it module structure, and interface (with some corrections).

Sklearn doesn't contain any tools to print reports on classification, since they are very specific and hep_ml won't too - current reports module will be moved to REP.

This will allow to focus on classifiers.
There are currently many experiments with boosting and its variations. In my plans extending this with neural networks, providing some general modifications of boosting and decision/regressing trees.

Current parallelization will probably be removed:

• simple parallelization of train / prediction is built into REP
• the only thing that cannot be done directly in REP is inner parallelization of uBoost, but in current implementation it is hardly usable if one wants to use several classifiers, not only uBoost. I'll probably use python's multiprocessing there.

[arogozhnikov]

Grid search was moved to REP

[eyadsibai]

when will this happen :)
I am trying to use REP with new scikit-learn and python 3 ... I did the changes needed in rep but will this be integrated soon or should I do a pull request for both?

this package now will be only needed if someone is using tmva.
eyadsibai/rep@014e5a6

[arogozhnikov]

Hi, Eyad.
I was sure that I will end this in May, but haven't yet.
hep_ml is not wrapper around TMVA (if I understood your point right)

Current status

At this moment I am working over rewriting hep_ml to minimalistic version and remove duplicate code and functions from rep (this version is quite nice and probably very soon I'll be ready to publish it).

From practical point, current version of hep_ml is completely usable inside rep, you can use SklearnClassifier as wrapper around classifiers from hep_ml (since those are sklearn-compatible).

Uniformity metrics can be used inside REP reports too.

There is minor problem, because you'll have to set features, train_variables and uniform_variables, but it isn't something crucial.