[ArrayManager] Indexing - implement iset (pandas-dev#39734)

Author: jorisvandenbossche

.github/workflows/ci.yml

@@ -153,3 +153,9 @@ jobs:
       run: |
         source activate pandas-dev
         pytest pandas/tests/frame/methods --array-manager
+
+        # indexing iset related (temporary since other tests don't pass yet)
+        pytest pandas/tests/frame/indexing/test_indexing.py::TestDataFrameIndexing::test_setitem_multi_index --array-manager
+        pytest pandas/tests/frame/indexing/test_setitem.py::TestDataFrameSetItem::test_setitem_listlike_indexer_duplicate_columns --array-manager
+        pytest pandas/tests/indexing/multiindex/test_setitem.py::TestMultiIndexSetItem::test_astype_assignment_with_dups --array-manager
+        pytest pandas/tests/indexing/multiindex/test_setitem.py::TestMultiIndexSetItem::test_frame_setitem_multi_column --array-manager

pandas/core/frame.py

@@ -9723,12 +9723,3 @@ def _reindex_for_setitem(value: FrameOrSeriesUnion, index: Index) -> ArrayLike:
             "incompatible index of inserted column with frame index"
         ) from err
     return reindexed_value
-
-
-def _maybe_atleast_2d(value):
-    # TODO(EA2D): not needed with 2D EAs
-
-    if is_extension_array_dtype(value):
-        return value
-
-    return np.atleast_2d(np.asarray(value))

pandas/core/internals/array_manager.py

@@ -659,24 +659,53 @@ def idelete(self, indexer):
 
     def iset(self, loc: Union[int, slice, np.ndarray], value):
         """
-        Set new item in-place. Does not consolidate. Adds new Block if not
-        contained in the current set of items
+        Set new column(s).
+
+        This changes the ArrayManager in-place, but replaces (an) existing
+        column(s), not changing column values in-place).
+
+        Parameters
+        ----------
+        loc : integer, slice or boolean mask
+            Positional location (already bounds checked)
+        value : array-like
         """
+        # single column -> single integer index
         if lib.is_integer(loc):
-            # TODO normalize array -> this should in theory not be needed?
+            # TODO the extract array should in theory not be needed?
             value = extract_array(value, extract_numpy=True)
+
+            # TODO can we avoid needing to unpack this here? That means converting
+            # DataFrame into 1D array when loc is an integer
             if isinstance(value, np.ndarray) and value.ndim == 2:
+                assert value.shape[1] == 1
                 value = value[0, :]
 
             assert isinstance(value, (np.ndarray, ExtensionArray))
-            # value = np.asarray(value)
-            # assert isinstance(value, np.ndarray)
+            assert value.ndim == 1
             assert len(value) == len(self._axes[0])
             self.arrays[loc] = value
             return
 
-        # TODO
-        raise Exception
+        # multiple columns -> convert slice or array to integer indices
+        elif isinstance(loc, slice):
+            indices = range(
+                loc.start if loc.start is not None else 0,
+                loc.stop if loc.stop is not None else self.shape_proper[1],
+                loc.step if loc.step is not None else 1,
+            )
+        else:
+            assert isinstance(loc, np.ndarray)
+            assert loc.dtype == "bool"
+            indices = np.nonzero(loc)[0]
+
+        assert value.ndim == 2
+        assert value.shape[0] == len(self._axes[0])
+
+        for value_idx, mgr_idx in enumerate(indices):
+            value_arr = value[:, value_idx]
+            self.arrays[mgr_idx] = value_arr
+        return
 
     def insert(self, loc: int, item: Hashable, value, allow_duplicates: bool = False):
         """

pandas/core/internals/managers.py

@@ -1140,8 +1140,7 @@ def insert(self, loc: int, item: Hashable, value, allow_duplicates: bool = False
 
         if value.ndim == 2:
             value = value.T
-
-        if value.ndim == self.ndim - 1 and not is_extension_array_dtype(value.dtype):
+        elif value.ndim == self.ndim - 1 and not is_extension_array_dtype(value.dtype):
             # TODO(EA2D): special case not needed with 2D EAs
             value = safe_reshape(value, (1,) + value.shape)