Handling the case where converting empty categorical to 'pyarrow' dtype_backend results in error. Since conversion of non-empty categorical returns categorical of 'numpy_nullable' dtype_backend, now, instead of raising an error, we ensure empty categorical is returned as well.

Author: veljanin

pandas/core/dtypes/cast.py

@@ -1112,7 +1112,7 @@ def convert_dtypes(
 
     else:
         inferred_dtype = input_array.dtype
-
+    
     if dtype_backend == "pyarrow":
         from pandas.core.arrays.arrow.array import to_pyarrow_type
         from pandas.core.arrays.string_ import StringDtype
@@ -1145,12 +1145,16 @@ def convert_dtypes(
                 and isna(input_array).all()
             ):
                 import pyarrow as pa
-
+                
                 pa_type = pa.null()
             else:
                 pa_type = to_pyarrow_type(base_dtype)
             if pa_type is not None:
-                inferred_dtype = ArrowDtype(pa_type)
+                if isna(input_array).all() and hasattr(input_array, 'categories'):
+                    inferred_dtype = input_array.dtype
+                else:
+                    inferred_dtype = ArrowDtype(pa_type)
+
     elif dtype_backend == "numpy_nullable" and isinstance(inferred_dtype, ArrowDtype):
         # GH 53648
         inferred_dtype = _arrow_dtype_mapping()[inferred_dtype.pyarrow_dtype]

pandas/tests/frame/methods/test_convert_dtypes.py

@@ -34,6 +34,22 @@ def test_convert_empty(self):
         # Empty DataFrame can pass convert_dtypes, see GH#40393
         empty_df = pd.DataFrame()
         tm.assert_frame_equal(empty_df, empty_df.convert_dtypes())
+    
+    def test_convert_empty_categorical_to_pyarrow(self):
+        df = pd.DataFrame(
+            {
+                "A": pd.Series(pd.Categorical([None] * 5)),
+                "B": pd.Series(pd.Categorical([None] * 5, categories=["B1", "B2"])),
+             }
+        )
+        converted = df.convert_dtypes(dtype_backend="pyarrow")
+        expected = df
+        tm.assert_frame_equal(converted, expected)
+        
+        assert df.A.dtype == "category", "Dtype in column A is not 'category'"
+        assert df.B.dtype == "category", "Dtype in column B is not 'category'"
+        assert df.A.cat.categories.empty, "Categories in column A are not empty"
+        assert (df.B.cat.categories == ["B1", "B2"]).all(), "Categories in column A are not empty"
 
     def test_convert_dtypes_retain_column_names(self):
         # GH#41435

pandas/tests/series/methods/test_convert_dtypes.py

@@ -297,3 +297,16 @@ def test_convert_dtypes_pyarrow_null(self):
         result = ser.convert_dtypes(dtype_backend="pyarrow")
         expected = pd.Series([None, None], dtype=pd.ArrowDtype(pa.null()))
         tm.assert_series_equal(result, expected)
+    
+    def test_convert_empty_categorical_to_pyarrow(self):
+        ser = pd.Series(pd.Series(pd.Categorical([None] * 5)))
+        
+        converted = ser.convert_dtypes(dtype_backend="pyarrow")
+        expected = ser
+        tm.assert_series_equal(converted, expected)
+        
+        assert ser.dtype == "category", "Series dtype is not 'category'"
+        assert ser.cat.categories.empty, "Series categories are not empty"
+        
+        ser2 = pd.Series(pd.Series(pd.Categorical([None] * 5, categories=["S1", "S2"])))
+        assert (ser2.cat.categories == ["S1", "S2"]).all(), "Series categories are not empty"