[RF] RDataFrame to RooDataSet/RooDataHist conversion

[hageboeck]

Is your feature request related to a problem? Please describe.

You could call it a "problem", yes.

What we want basically is to drop out an RooDataSet directly from RDataFrame. There's two ways to do it:

Easy way:

• Create a loopable collection of events that drops out of RDF. Could be a std::vector, a snapshot, etc
• One by one, fill those events into RooDataSet à la https://root.cern/doc/master/rf402__datahandling_8C.html

   // Fill d with dummy values
   Int_t i;
   for (i = 0; i < 1000; i++) {
      x = i / 50 - 10;
      y = sqrt(1.0 * i);
      c.setLabel((i % 2) ? "Plus" : "Minus");
 
      // We must explicitly refer to x,y,c here to pass the values because
      // d is not linked to them (as explained above)
      d.add(RooArgSet(x, y, c));
   }

This is going to be slow, and not so elegant.

Requires more development, but better:

• RooDataSet should accept collections when you construct it
• Since data is stored in a std::vector internally, it could even std::swap with an external vector!
• RDataFrame should directly spit out such a collection or it should be easy to connect the two.
  • It is clear that MT runs are not trivial, but maybe not a priority.
  • @eguiraud what is the easiest to spit out from RDF, which is loopable, maybe even adopt the memory?

[eguiraud]

I think you want a dedicated custom action that users can Book.

In general you can also just Take std::vectors (or other collections) out of a RDataFrame.

[hageboeck]

Book is good-ish, it would fill the dataset one-by-one, but RF can do better.

I think Take is exactly what's needed!

Eventually, this could look like:

auto rooDataSet = rdf.RooDataSet("datasetname", {"x", "y", "z"});
pdf.fitTo(*rooDataSet);

Internally, this would do something like:

RooRealVar x(...);
RooRealVar y(...);
RooRealVar z(...);

auto xData = rdf.Take("x");
auto yData = rdf.Take("y");
auto zData = rdf.Take("z");

return RooDataSet(datasetname, RooArgSet(x, y, z), {std::move(*xData), std::move(*yData), std::move(*zData)}); // This constructor doesn't exist

There is lots of details missing, but that's what it could do eventually. Maybe the RooRealVars have to be passed into whatever it is I wrote above, because they have ranges and binnings and labels ...
In general, the thing is just supposed to connect those RooRealVar objects with the data that's taken from RDF.

Possible?

[eguiraud]

Book is good-ish, it would fill the dataset one-by-one, but RF can do better.

Uhm, Book exposes the exact same mechanics that RDF uses for its own actions like rdf.RooDataSet, so anything you would like to do in there, you can do in the helper type you pass to Book.

Then it's a small step to go from df.Book(RooDataSetMaker(...), {"x", "y", "z"}) to df.MakeRooDataSet, basically just syntactic sugar.

[hageboeck]

Ok, for single value filling, this may be achievable in a few days, nice.

[hageboeck]

Reopening, because weights should be tested (and supported). What's needed:

• Creating and filling bare dataset / datahist
• A way to select the a weight column in RDataFrame, and tell the RooDataHist/RooDataSet instances about what's the weight.
• Passing of the weight in the add call.
• Test of weighted filling.