Backport PR #54881 on branch 2.1.x (REGR: read_csv raising when dtypes is specified with usecols) (#54926)

Author: phofl

doc/source/whatsnew/v2.1.1.rst

@@ -13,6 +13,7 @@ including other versions of pandas.
 
 Fixed regressions
 ~~~~~~~~~~~~~~~~~
+- Fixed regression in :func:`read_csv` when ``usecols`` is given and ``dtypes`` is a dict for ``engine="python"`` (:issue:`54868`)
 - Fixed regression in :meth:`DataFrame.__setitem__` raising ``AssertionError`` when setting a :class:`Series` with a partial :class:`MultiIndex` (:issue:`54875`)
 - Fixed regression when comparing a :class:`Series` with ``datetime64`` dtype with ``None`` (:issue:`54870`)
 

pandas/io/parsers/python_parser.py

@@ -1176,17 +1176,17 @@ def _set_no_thousand_columns(self) -> set[int]:
  )
  if self.columns and self.dtype:
  assert self._col_indices is not None
- for i in self._col_indices:
+ for i, col in zip(self._col_indices, self.columns):
  if not isinstance(self.dtype, dict) and not is_numeric_dtype(
  self.dtype
  ):
  no_thousands_columns.add(i)
  if (
  isinstance(self.dtype, dict)
- and self.columns[i] in self.dtype
+ and col in self.dtype
  and (
- not is_numeric_dtype(self.dtype[self.columns[i]])
- or is_bool_dtype(self.dtype[self.columns[i]])
+ not is_numeric_dtype(self.dtype[col])
+ or is_bool_dtype(self.dtype[col])
  )
  ):
  no_thousands_columns.add(i)

pandas/tests/io/parser/dtypes/test_dtypes_basic.py

@@ -558,3 +558,20 @@ def test_string_inference(all_parsers):
  columns=pd.Index(["a", "b"], dtype=dtype),
  )
  tm.assert_frame_equal(result, expected)
+
+
+def test_dtypes_with_usecols(all_parsers):
+ # GH#54868
+
+ parser = all_parsers
+ data = """a,b,c
+1,2,3
+4,5,6"""
+
+ result = parser.read_csv(StringIO(data), usecols=["a", "c"], dtype={"a": object})
+ if parser.engine == "pyarrow":
+ values = [1, 4]
+ else:
+ values = ["1", "4"]
+ expected = DataFrame({"a": pd.Series(values, dtype=object), "c": [3, 6]})
+ tm.assert_frame_equal(result, expected)