Pandas (and other array package) improved compatibility

[henryiii]

boost-histogram currently doesn't work very well with Pandas DataFrames, requiring a np.asarary to work (and sometimes an explicit cast to a NumPy string datatype).

• My suggestion would be to do the np.asarray wrapping inside the Python fill wrapper, so that these can be simplified when using non-NumPy based arrays.

• You also can't setup categories from all iterables, like sets, but only from true lists, which is restrictive.

Actual usage:

hist = bh.Histogram(
    bh.axis.StrCategory(list(set(skhep.file_project))),
    bh.axis.Integer(2018, 2021, underflow=False, overflow=False),
    bh.axis.Integer(2, 4, underflow=False, overflow=False),
    storage=bh.storage.Int64()
)

hist.fill(np.asarray(skhep.file_project, dtype=str),
          np.asarray(skhep.timestamp.dt.year),
          np.asarray(skhep.details_python.str[0].astype(int))
)

Ideal usage:

hist = bh.Histogram(
    bh.axis.StrCategory(set(skhep.file_project)),
    bh.axis.Integer(2018, 2021, underflow=False, overflow=False),
    bh.axis.Integer(2, 4, underflow=False, overflow=False),
    storage=bh.storage.Int64()
)

hist.fill(skhep.file_project,
          skhep.timestamp.dt.year,
          skhep.details_python.str[0].astype(int)
)

[HDembinski]

I agree, fill should accept any array-like sequence and use np.asarray to convert it.

Edit: Perhaps it should be np.ascontiguousarray to make sure everything works correctly. https://stackoverflow.com/questions/14415741/what-is-the-difference-between-numpys-array-and-asarray-functions

[HDembinski]

I also agree on Category accepting any sequence.

[henryiii]

We already forcecast in PyBind11, so np.ascontiguousarray shouldn't be necessary, but probably wouldn't hurt. It just moves where the conversion happens - if a copy is made, it should (I think) become contiguous at the same time.

[henryiii]

This is not supported still by weights and probably samples - it should be, though. Reported by @meliache on Scikit-HEP Gitter.