REF: Make DataFrame.from_pandas process by column (#14483)

Also encountered a bug where `cudf.Index.from_pandas` would return an `cudf.Index[int64]` from a `pandas.RangeIndex`

Authors:
  - Matthew Roeschke (https://github.com/mroeschke)

Approvers:
  - Lawrence Mitchell (https://github.com/wence-)

URL: #14483

python/cudf/cudf/core/_base_index.py

@@ -1873,10 +1873,18 @@ def from_pandas(cls, index, nan_as_null=no_default):
 
         if not isinstance(index, pd.Index):
             raise TypeError("not a pandas.Index")
-
-        ind = cudf.Index(column.as_column(index, nan_as_null=nan_as_null))
-        ind.name = index.name
-        return ind
+        if isinstance(index, pd.RangeIndex):
+            return cudf.RangeIndex(
+                start=index.start,
+                stop=index.stop,
+                step=index.step,
+                name=index.name,
+            )
+        else:
+            return cudf.Index(
+                column.as_column(index, nan_as_null=nan_as_null),
+                name=index.name,
+            )
 
     @property
     def _constructor_expanddim(self):

python/cudf/cudf/core/dataframe.py

@@ -5245,30 +5245,20 @@ def from_pandas(cls, dataframe, nan_as_null=no_default):
             if not dataframe.columns.is_unique:
                 raise ValueError("Duplicate column names are not allowed")
 
-            # Set columns
-            data = {}
-            for col_name, col_value in dataframe.items():
-                # necessary because multi-index can return multiple
-                # columns for a single key
-                if len(col_value.shape) == 1:
-                    data[col_name] = column.as_column(
-                        col_value.array, nan_as_null=nan_as_null
-                    )
-                else:
-                    vals = col_value.values.T
-                    if vals.shape[0] == 1:
-                        data[col_name] = column.as_column(
-                            vals.flatten(), nan_as_null=nan_as_null
-                        )
-                    else:
-                        if isinstance(col_name, tuple):
-                            col_name = str(col_name)
-                        for idx in range(len(vals.shape)):
-                            data[col_name] = column.as_column(
-                                vals[idx], nan_as_null=nan_as_null
-                            )
-
-            index = cudf.from_pandas(dataframe.index, nan_as_null=nan_as_null)
+            data = {
+                col_name: column.as_column(
+                    col_value.array, nan_as_null=nan_as_null
+                )
+                for col_name, col_value in dataframe.items()
+            }
+            if isinstance(dataframe.index, pd.MultiIndex):
+                index = cudf.MultiIndex.from_pandas(
+                    dataframe.index, nan_as_null=nan_as_null
+                )
+            else:
+                index = cudf.Index.from_pandas(
+                    dataframe.index, nan_as_null=nan_as_null
+                )
             df = cls._from_data(data, index)
             df._data._level_names = tuple(dataframe.columns.names)
 
@@ -5279,13 +5269,14 @@ def from_pandas(cls, dataframe, nan_as_null=no_default):
                 df.columns = dataframe.columns
 
             return df
+        elif hasattr(dataframe, "__dataframe__"):
+            # TODO: Probably should be handled in the constructor as
+            # this isn't pandas specific
+            return from_dataframe(dataframe, allow_copy=True)
         else:
-            try:
-                return from_dataframe(dataframe, allow_copy=True)
-            except Exception:
-                raise TypeError(
-                    f"Could not construct DataFrame from {type(dataframe)}"
-                )
+            raise TypeError(
+                f"Could not construct DataFrame from {type(dataframe)}"
+            )
 
     @classmethod
     @_cudf_nvtx_annotate
@@ -7915,10 +7906,6 @@ def from_pandas(obj, nan_as_null=no_default):
         return ret
     elif isinstance(obj, pd.MultiIndex):
         return MultiIndex.from_pandas(obj, nan_as_null=nan_as_null)
-    elif isinstance(obj, pd.RangeIndex):
-        return cudf.core.index.RangeIndex(
-            start=obj.start, stop=obj.stop, step=obj.step, name=obj.name
-        )
     elif isinstance(obj, pd.Index):
         return cudf.Index.from_pandas(obj, nan_as_null=nan_as_null)
     elif isinstance(obj, pd.CategoricalDtype):

python/cudf/cudf/tests/test_index.py

@@ -2954,6 +2954,13 @@ def test_index_getitem_from_nonint_raises(idx):
         cudf.Index([1, 2])[idx]
 
 
+def test_from_pandas_rangeindex_return_rangeindex():
+    pidx = pd.RangeIndex(start=3, stop=9, step=3, name="a")
+    result = cudf.Index.from_pandas(pidx)
+    expected = cudf.RangeIndex(start=3, stop=9, step=3, name="a")
+    assert_eq(result, expected, exact=True)
+
+
 @pytest.mark.parametrize(
     "idx",
     [