We have currently the issue that we load at the startup time libraires such as TMVA because we inject free functions like TMVA.Experimental.AsRTensor. Same holds true for ROOT.VecOps.AsRVec. Since these loads slow down the startup of PyROOT, we try to get rid of them.
This comes together with making the names of our utilities for interoperability between python/numpy and ROOT more consistent. ROOT has the functions ROOT.RDataFrame.AsNumpy and the two mentioned above. We propose to move the free functions as members of the corresponding C++ class, for example move ROOT.VecOps.AsRVec to ROOT.VecOps.RVec.FromNumpy. So each conversion from a C++ object to numpy array would be called AsNumpy and the conversion from the numpy world to the C++ world FromNumpy.