BUG: .replace coerces incorrect dtype

doc/source/whatsnew/v0.20.0.txt

@@ -823,6 +823,7 @@ Bug Fixes
 
 
 - Bug in the display of ``.info()`` where a qualifier (+) would always be displayed with a ``MultiIndex`` that contains only non-strings (:issue:`15245`)
+- Bug in ``.replace()`` may result in incorrect dtypes. (:issue:`12747`)
 
 - Bug in ``.asfreq()``, where frequency was not set for empty ``Series`` (:issue:`14320`)
 

pandas/core/internals.py

@@ -1894,8 +1894,11 @@ def convert(self, *args, **kwargs):
                 blocks.append(newb)
 
         else:
-            values = fn(
-                self.values.ravel(), **fn_kwargs).reshape(self.values.shape)
+            values = fn(self.values.ravel(), **fn_kwargs)
+            try:
+                values = values.reshape(self.values.shape)
+            except NotImplementedError:
+                pass
             blocks.append(make_block(values, ndim=self.ndim,
                                      placement=self.mgr_locs))
 
@@ -3238,6 +3241,16 @@ def comp(s):
             return _possibly_compare(values, getattr(s, 'asm8', s),
                                      operator.eq)
 
+        def _cast_scalar(block, scalar):
+            dtype, val = _infer_dtype_from_scalar(scalar, pandas_dtype=True)
+            if not is_dtype_equal(block.dtype, dtype):
+                dtype = _find_common_type([block.dtype, dtype])
+                block = block.astype(dtype)
+                # use original value
+                val = scalar
+
+            return block, val
+
         masks = [comp(s) for i, s in enumerate(src_list)]
 
         result_blocks = []
@@ -3260,7 +3273,8 @@ def comp(s):
                         # particular block
                         m = masks[i][b.mgr_locs.indexer]
                         if m.any():
-                            new_rb.extend(b.putmask(m, d, inplace=True))
+                            b, val = _cast_scalar(b, d)
+                            new_rb.extend(b.putmask(m, val, inplace=True))
                         else:
                             new_rb.append(b)
                 rb = new_rb

pandas/tests/indexing/test_coercion.py

@@ -1153,12 +1153,27 @@ def setUp(self):
         self.rep['float64'] = [1.1, 2.2]
         self.rep['complex128'] = [1 + 1j, 2 + 2j]
         self.rep['bool'] = [True, False]
+        self.rep['datetime64[ns]'] = [pd.Timestamp('2011-01-01'),
+                                      pd.Timestamp('2011-01-03')]
+
+        for tz in ['UTC', 'US/Eastern']:
+            # to test tz => different tz replacement
+            key = 'datetime64[ns, {0}]'.format(tz)
+            self.rep[key] = [pd.Timestamp('2011-01-01', tz=tz),
+                             pd.Timestamp('2011-01-03', tz=tz)]
+
+        self.rep['timedelta64[ns]'] = [pd.Timedelta('1 day'),
+                                       pd.Timedelta('2 day')]
 
     def _assert_replace_conversion(self, from_key, to_key, how):
         index = pd.Index([3, 4], name='xxx')
         obj = pd.Series(self.rep[from_key], index=index, name='yyy')
         self.assertEqual(obj.dtype, from_key)
 
+        if (from_key.startswith('datetime') and to_key.startswith('datetime')):
+            # different tz, currently mask_missing raises SystemError
+            return
+
         if how == 'dict':
             replacer = dict(zip(self.rep[from_key], self.rep[to_key]))
         elif how == 'series':
@@ -1175,17 +1190,12 @@ def _assert_replace_conversion(self, from_key, to_key, how):
             pytest.skip("windows platform buggy: {0} -> {1}".format
                         (from_key, to_key))
 
-        if ((from_key == 'float64' and
-             to_key in ('bool', 'int64')) or
-
+        if ((from_key == 'float64' and to_key in ('bool', 'int64')) or
             (from_key == 'complex128' and
              to_key in ('bool', 'int64', 'float64')) or
 
-            (from_key == 'int64' and
-             to_key in ('bool')) or
-
             # TODO_GH12747 The result must be int?
-                (from_key == 'bool' and to_key == 'int64')):
+            (from_key == 'int64' and to_key in ('bool'))):
 
             # buggy on 32-bit
             if tm.is_platform_32bit():
@@ -1248,13 +1258,31 @@ def test_replace_series_bool(self):
             self._assert_replace_conversion(from_key, to_key, how='series')
 
     def test_replace_series_datetime64(self):
-        pass
+        from_key = 'datetime64[ns]'
+        for to_key in self.rep:
+            self._assert_replace_conversion(from_key, to_key, how='dict')
+
+        from_key = 'datetime64[ns]'
+        for to_key in self.rep:
+            self._assert_replace_conversion(from_key, to_key, how='series')
 
     def test_replace_series_datetime64tz(self):
-        pass
+        from_key = 'datetime64[ns, US/Eastern]'
+        for to_key in self.rep:
+            self._assert_replace_conversion(from_key, to_key, how='dict')
+
+        from_key = 'datetime64[ns, US/Eastern]'
+        for to_key in self.rep:
+            self._assert_replace_conversion(from_key, to_key, how='series')
 
     def test_replace_series_timedelta64(self):
-        pass
+        from_key = 'timedelta64[ns]'
+        for to_key in self.rep:
+            self._assert_replace_conversion(from_key, to_key, how='dict')
+
+        from_key = 'timedelta64[ns]'
+        for to_key in self.rep:
+            self._assert_replace_conversion(from_key, to_key, how='series')
 
     def test_replace_series_period(self):
         pass

pandas/tests/series/test_replace.py

@@ -132,8 +132,8 @@ def check_replace(to_rep, val, expected):
             tm.assert_series_equal(expected, r)
             tm.assert_series_equal(expected, sc)
 
-        # should NOT upcast to float
-        e = pd.Series([0, 1, 2, 3, 4])
+        # MUST upcast to float
+        e = pd.Series([0., 1., 2., 3., 4.])
         tr, v = [3], [3.0]
         check_replace(tr, v, e)
 

pandas/types/cast.py

@@ -21,7 +21,7 @@
                      _ensure_int32, _ensure_int64,
                      _NS_DTYPE, _TD_DTYPE, _INT64_DTYPE,
                      _POSSIBLY_CAST_DTYPES)
-from .dtypes import ExtensionDtype
+from .dtypes import ExtensionDtype, DatetimeTZDtype, PeriodDtype
 from .generic import ABCDatetimeIndex, ABCPeriodIndex, ABCSeries
 from .missing import isnull, notnull
 from .inference import is_list_like
@@ -312,8 +312,17 @@ def _maybe_promote(dtype, fill_value=np.nan):
     return dtype, fill_value
 
 
-def _infer_dtype_from_scalar(val):
-    """ interpret the dtype from a scalar """
+def _infer_dtype_from_scalar(val, pandas_dtype=False):
+    """
+    interpret the dtype from a scalar
+
+    Parameters
+    ----------
+    pandas_dtype : bool, default False
+        whether to infer dtype including pandas extension types.
+        If False, scalar belongs to pandas extension types is inferred as
+        object
+    """
 
     dtype = np.object_
 
@@ -336,13 +345,20 @@ def _infer_dtype_from_scalar(val):
 
         dtype = np.object_
 
-    elif isinstance(val, (np.datetime64,
-                          datetime)) and getattr(val, 'tzinfo', None) is None:
-        val = lib.Timestamp(val).value
-        dtype = np.dtype('M8[ns]')
+    elif isinstance(val, (np.datetime64, datetime)):
+        val = tslib.Timestamp(val)
+        if val is tslib.NaT or val.tz is None:
+            dtype = np.dtype('M8[ns]')
+        else:
+            if pandas_dtype:
+                dtype = DatetimeTZDtype(unit='ns', tz=val.tz)
+            else:
+                # return datetimetz as object
+                return np.object_, val
+        val = val.value
 
     elif isinstance(val, (np.timedelta64, timedelta)):
-        val = lib.Timedelta(val).value
+        val = tslib.Timedelta(val).value
         dtype = np.dtype('m8[ns]')
 
     elif is_bool(val):
@@ -363,6 +379,11 @@ def _infer_dtype_from_scalar(val):
     elif is_complex(val):
         dtype = np.complex_
 
+    elif pandas_dtype:
+        if lib.is_period(val):
+            dtype = PeriodDtype(freq=val.freq)
+            val = val.ordinal
+
     return dtype, val