Fix type of empty Index and raise warning in Series constructor
#14116
Conversation
galipremsagar
added
Python
| @@ -2840,7 +2841,7 @@ def test_series_all_null(num_elements, null_type): | |||
| @pytest.mark.parametrize("num_elements", [0, 2, 10, 100]) | |||
| def test_series_all_valid_nan(num_elements): | |||
| data = [np.nan] * num_elements | |||
| sr = cudf.Series(data, nan_as_null=False) | |||
| sr = _create_cudf_series(data, nan_as_null=False) | |||
There was a problem hiding this comment.
Why is this change needed?
For num_elements == 0, the null count should still be zero even if it interprets the dtype as string/object (so the test is still valid, though it might warn?).
For num_elements > 0, the data should be non-empty and thus be correctly interpreted as a floating type.
There was a problem hiding this comment.
For
num_elements == 0, the null count should still be zero even if it interprets the dtype as string/object (so the test is still valid, though it might warn?).
Yup, this is the reason for this change.
| @@ -4073,28 +4074,28 @@ def test_empty_dataframe_describe(): | |||
|
|
|||
|
|
|||
| def test_as_column_types(): | |||
| col = column.as_column(cudf.Series([])) | |||
| col = column.as_column(_create_cudf_series([])) | |||
There was a problem hiding this comment.
Same here and below -- I'm not sure if I understand why these helper functions are needed. It seems like the test construction might need subtle changes but I don't think this casting helper function is the right answer. It's not announcing what it is actually doing -- the differences between cudf.Series and _create_cudf_series should be in the name, if this is needed (like _create_cudf_series_float64_default).
There was a problem hiding this comment.
Agreed, reverted the change here and passed explicit dtypes. Renamed the helper to _create_cudf_series_float64_default and _create_pandas_series_float64_default. These will be removed in pandas-2.0 upgrade.
|
/merge |
galipremsagar
added
5 - Ready to Merge
Calling `cudf.Index([])` results in `str` dtype `Index`. This PR fixes an issue with a pytorch related pytest by explicitly passing a `float64` dtype. xref: #14116 Authors: - GALI PREM SAGAR (https://github.com/galipremsagar) Approvers: - https://github.com/brandon-b-miller URL: #14198
Description
Fixes: #14091
This PR fixes empty inputs dtype in
Indexto default tostrinstead offloat64. Another change is there is a deprecation warning forSeriesconstructor to match pandas.Checklist