Why state_ in Sampler and SubsetGenerator?

[Peter]

Is it only for historical reasons there was a state in CrossSplitter? (now splitted into Sampler and SubsetGenerator?)? Instead of "next", "more" and "reset", wouldn't it be easier to just all member vectors as const references and users can use operator[] to get what they want.

[Markus Ringnér]

Markus thinks: Sampler should have no state but I do not like getting entire vectors (mostly vectors of vectors) out. Instead I have now removed the state but changed the Sampler interface so the access functions take an index as argument. For SubSetGenerator? I think state may be useful.

[Peter]

It is a bit strange though to have a container with a state. First, it implies a SubsetGenerator? object is not supposed to be shared because that would introduce very long correlations. Second, looking in our dear std containers are clearly separated from iterators, and the latter keep the state functionality. Third, the argument that the state is needed to avoid another parameter in SupervisedClassifier::make_class interface in no longer really valid. When the introduction of MatrixLookupWeighted? is completed make_class could take a DataLookup2D& instead (see #106) solving that issue

[Markus Ringnér]

I agree make_classifier should take a DataLookup2D. One reason to have a state is that SubsetGenerator? has tons of functions that typically are called close to one another and otherwise would take an index. Using a state would perhas minimize index errors. Is this a useful argument? Otherwise make it statless!

[Jari Häkkinen]

Removing type "idea" and changing this to a "request".

[Markus Ringnér]

(In [704]) Fixes #104. Also fixed inline bug in Averager.h