[BUG] iloc indexing with a Categorical not possible

[wence-]

Describe the bug

import cudf
import pandas as pd

f = pd.Series([0, 1, 2])
index1 = pd.Categorical([1])
f.iloc[index1] # => Length 1 series with the second row
index2 = pd.Categorical([0, 2])
f.iloc[index2] # => Length 2 series with first and third rows

cf = cudf.from_pandas(f)
cf.iloc[index1] # => TypeError: int() argument must be a string, a bytes-like object or a real number, not 'Categorical'
cf.iloc[index2] # => NotImplementedError: Unknown indexer <class 'cudf.core.column.categorical.CategoricalColumn'>

The former error is because _is_scalar_or_zero_d_array treats categoricals of length one as scalars (prior to vyasr@6251122 all categoricals were treated as scalars). Which, as observed above, is in conflict with pandas.

The latter is an error because casting into a an integer array doesn't work, and so the indexer ends up not knowing what to do.

Expected behavior

I'm not sure, the implicit casting to integers for iloc-based indexing of an object that might well not contain integers feels like an accident of pandas history. The pandas documentation says (of iloc indexing):

The .iloc attribute is the primary access method. The following are valid inputs:
[...]

• A list or array of integers [4, 3, 0].

I think a categorical is not a list of integers, so this should be a TypeError, but it just happens to work because np.asarray(whatever).astype(int) works as long as whatever is coercible to int (this is broad e.g. f.iloc[[1.5]] works!?).

@mroeschke do you happen to know if iloc-based indexing with categoricals is "definitely supported" or "accident of history"?

[mroeschke]

do you happen to know if iloc-based indexing with categoricals is "definitely supported" or "accident of history"?

pd.Categorical is "array-like" so I would expect a pd.Categorical of integers to work in iloc in pandas

[wence-]

Thanks!

[wence-]

pd.Categorical is "array-like" so I would expect a pd.Categorical of integers to work in iloc in pandas

What about a categorical of floats?

import pandas as pd
s = pd.Series([0, 1, 2])
s.iloc[pd.Categorical([1.5])]
# => 1 1

[mroeschke]

What about a categorical of floats?

Yeah I would expect this to raise an exception ( I think pandas has an existing issue of passing an array-like of floats to iloc with an integer index )