From bd0369fffc2667ecf9876952ecf458aa88c598e1 Mon Sep 17 00:00:00 2001
From: jreback <jeff@reback.net>
Date: Wed, 11 Sep 2013 19:59:29 -0400
Subject: [PATCH 1/4] ENH: support timedelta64[ns] as a serialization type in
 HDFStore for query and append (GH3577)

---
 doc/source/io.rst                | 16 +++++++++++
 doc/source/release.rst           |  1 +
 doc/source/v0.13.0.txt           |  2 +-
 pandas/core/common.py            | 21 ++++++++++++--
 pandas/io/pytables.py            | 39 ++++++++++++++++++++++++--
 pandas/io/tests/test_pytables.py | 48 +++++++++++++++++++++++++++++++-
 pandas/tslib.pyx                 |  2 +-
 7 files changed, 121 insertions(+), 8 deletions(-)

diff --git a/doc/source/io.rst b/doc/source/io.rst
index 1d3980e216587..94c8343bfec1d 100644
--- a/doc/source/io.rst
+++ b/doc/source/io.rst
@@ -2009,6 +2009,22 @@ space. These are in terms of the total number of rows in a table.
                        Term('minor_axis', '=', ['A','B']) ],
                 start=0, stop=10)
 
+**Using timedelta64[ns]**
+
+.. versionadded:: 0.13
+
+Beginning in 0.13.0, you can store and query using the ``timedelta64[ns]`` type. Terms can be
+specified in the format: ``<float>(<unit>)``, where float may be signed (and fractional), and unit can be
+``D,s,ms,us,ns`` for the timedelta. Here's an example:
+
+.. ipython:: python
+
+   from datetime import timedelta
+   dftd = DataFrame(dict(A = Timestamp('20130101'), B = [ Timestamp('20130101') + timedelta(days=i,seconds=10) for i in range(10) ]))
+   dftd['C'] = dftd['A']-dftd['B']
+   dftd
+   store.append('dftd',dftd,data_columns=True)
+   store.select('dftd',Term("C","<","-3.5D"))
 
 Indexing
 ~~~~~~~~
diff --git a/doc/source/release.rst b/doc/source/release.rst
index 087d2880511d2..d50438cd08058 100644
--- a/doc/source/release.rst
+++ b/doc/source/release.rst
@@ -156,6 +156,7 @@ API Changes
     - a column multi-index will be recreated properly (:issue:`4710`); raise on trying to use a multi-index
       with data_columns on the same axis
     - ``select_as_coordinates`` will now return an ``Int64Index`` of the resultant selection set
+    - support ``timedelta64[ns]`` as a serialization type (:issue:`3577`)
   - ``JSON``
 
     - added ``date_unit`` parameter to specify resolution of timestamps. Options
diff --git a/doc/source/v0.13.0.txt b/doc/source/v0.13.0.txt
index caf218747bdfb..536b0badeec00 100644
--- a/doc/source/v0.13.0.txt
+++ b/doc/source/v0.13.0.txt
@@ -80,7 +80,7 @@ API changes
       See :ref:`here<io.hdf5-selecting_coordinates>` for an example.
     - allow a passed locations array or mask as a ``where`` condition (:issue:`4467`).
       See :ref:`here<io.hdf5-where_mask>` for an example.
-
+    - support ``timedelta64[ns]`` as a serialization type (:issue:`3577`)
     - the ``format`` keyword now replaces the ``table`` keyword; allowed values are ``fixed(f)`` or ``table(t)``
       the same defaults as prior < 0.13.0 remain, e.g. ``put`` implies 'fixed` or 'f' (Fixed) format
       and ``append`` imples 'table' or 't' (Table) format
diff --git a/pandas/core/common.py b/pandas/core/common.py
index ba7c6cc511933..2f153e88f4dc6 100644
--- a/pandas/core/common.py
+++ b/pandas/core/common.py
@@ -19,7 +19,6 @@
 from pandas.core.config import get_option
 from pandas.core import array as pa
 
-
 # XXX: HACK for NumPy 1.5.1 to suppress warnings
 try:
     np.seterr(all='ignore')
@@ -704,13 +703,29 @@ def diff(arr, n, axis=0):
 
     return out_arr
 
+timedelta_search = re.compile(
+    "^(?P<value>-?\d*\.?\d*)(?P<unit>D|s|ms|us|ns)?$")
 
-def _coerce_scalar_to_timedelta_type(r):
+def _coerce_scalar_to_timedelta_type(r, unit='ns'):
     # kludgy here until we have a timedelta scalar
     # handle the numpy < 1.7 case
 
+    if isinstance(r, compat.string_types):
+        m = timedelta_search.search(r)
+        if m:
+            r = float(m.groupdict()['value'])
+            u = m.groupdict().get('unit')
+            if u is not None:
+                unit = u
+        else:
+            raise ValueError("cannot convert timedelta scalar value!")
+
+        r = tslib.cast_from_unit(unit, r)
+        r = timedelta(microseconds=int(r)/1000)
+
     if is_integer(r):
-        r = timedelta(microseconds=r/1000)
+        r = tslib.cast_from_unit(unit, r)
+        r = timedelta(microseconds=int(r)/1000)
 
     if _np_version_under1p7:
         if not isinstance(r, timedelta):
diff --git a/pandas/io/pytables.py b/pandas/io/pytables.py
index 6759e07ed7935..bf45fbf96c0e5 100644
--- a/pandas/io/pytables.py
+++ b/pandas/io/pytables.py
@@ -25,7 +25,7 @@
 from pandas.core.common import adjoin, is_list_like, pprint_thing
 from pandas.core.algorithms import match, unique
 from pandas.core.categorical import Categorical
-from pandas.core.common import _asarray_tuplesafe
+from pandas.core.common import _asarray_tuplesafe, _np_version_under1p7
 from pandas.core.internals import BlockManager, make_block
 from pandas.core.reshape import block2d_to_blocknd, factor_indexer
 from pandas.core.index import _ensure_index
@@ -1527,6 +1527,8 @@ def set_kind(self):
                 self.kind = 'integer'
             elif dtype.startswith(u('date')):
                 self.kind = 'datetime'
+            elif dtype.startswith(u('timedelta')):
+                self.kind = 'timedelta'
             elif dtype.startswith(u('bool')):
                 self.kind = 'bool'
             else:
@@ -1547,6 +1549,11 @@ def set_atom(self, block, existing_col, min_itemsize, nan_rep, info, encoding=No
 
         if inferred_type == 'datetime64':
             self.set_atom_datetime64(block)
+        elif dtype == 'timedelta64[ns]':
+            if _np_version_under1p7:
+                raise TypeError(
+                    "timdelta64 is not supported under under numpy < 1.7")
+            self.set_atom_timedelta64(block)
         elif inferred_type == 'date':
             raise TypeError(
                 "[date] is not implemented as a table column")
@@ -1667,6 +1674,16 @@ def set_atom_datetime64(self, block, values=None):
             values = block.values.view('i8')
         self.set_data(values, 'datetime64')
 
+    def get_atom_timedelta64(self, block):
+        return _tables().Int64Col(shape=block.shape[0])
+
+    def set_atom_timedelta64(self, block, values=None):
+        self.kind = 'timedelta64'
+        self.typ = self.get_atom_timedelta64(block)
+        if values is None:
+            values = block.values.view('i8')
+        self.set_data(values, 'timedelta64')
+
     @property
     def shape(self):
         return getattr(self.data, 'shape', None)
@@ -1719,6 +1736,8 @@ def convert(self, values, nan_rep, encoding):
                 else:
                     self.data = np.asarray(self.data, dtype='M8[ns]')
 
+            elif dtype == u('timedelta64'):
+                self.data = np.asarray(self.data, dtype='m8[ns]')
             elif dtype == u('date'):
                 self.data = np.array(
                     [date.fromtimestamp(v) for v in self.data], dtype=object)
@@ -1767,6 +1786,9 @@ def get_atom_data(self, block):
     def get_atom_datetime64(self, block):
         return _tables().Int64Col()
 
+    def get_atom_timedelta64(self, block):
+        return _tables().Int64Col()
+
 
 class GenericDataIndexableCol(DataIndexableCol):
 
@@ -2007,6 +2029,11 @@ def read_array(self, key):
 
             if dtype == u('datetime64'):
                 ret = np.array(ret, dtype='M8[ns]')
+            elif dtype == u('timedelta64'):
+                if _np_version_under1p7:
+                    raise TypeError(
+                        "timedelta64 is not supported under under numpy < 1.7")
+                ret = np.array(ret, dtype='m8[ns]')
 
         if transposed:
             return ret.T
@@ -2214,6 +2241,9 @@ def write_array(self, key, value, items=None):
         elif value.dtype.type == np.datetime64:
             self._handle.createArray(self.group, key, value.view('i8'))
             getattr(self.group, key)._v_attrs.value_type = 'datetime64'
+        elif value.dtype.type == np.timedelta64:
+            self._handle.createArray(self.group, key, value.view('i8'))
+            getattr(self.group, key)._v_attrs.value_type = 'timedelta64'
         else:
             if empty_array:
                 self.write_array_empty(key, value)
@@ -4000,7 +4030,9 @@ def eval(self):
         """ set the numexpr expression for this term """
 
         if not self.is_valid:
-            raise ValueError("query term is not valid [%s]" % str(self))
+            raise ValueError("query term is not valid [{0}]\n"
+                             "  all queries terms must include a reference to\n"
+                             "  either an axis (e.g. index or column), or a data_columns\n".format(str(self)))
 
         # convert values if we are in the table
         if self.is_in_table:
@@ -4060,6 +4092,9 @@ def stringify(value):
             if v.tz is not None:
                 v = v.tz_convert('UTC')
             return TermValue(v, v.value, kind)
+        elif kind == u('timedelta64') or kind == u('timedelta'):
+            v = com._coerce_scalar_to_timedelta_type(v,unit='s').item()
+            return TermValue(int(v), v, kind)
         elif (isinstance(v, datetime) or hasattr(v, 'timetuple')
               or kind == u('date')):
             v = time.mktime(v.timetuple())
diff --git a/pandas/io/tests/test_pytables.py b/pandas/io/tests/test_pytables.py
index 7e5c3f9fff061..3667dff994232 100644
--- a/pandas/io/tests/test_pytables.py
+++ b/pandas/io/tests/test_pytables.py
@@ -23,6 +23,7 @@
                                  assert_series_equal)
 from pandas import concat, Timestamp
 from pandas import compat
+from pandas.core import common as com
 
 from numpy.testing.decorators import slow
 
@@ -1732,7 +1733,7 @@ def test_unimplemented_dtypes_table_columns(self):
             # this fails because we have a date in the object block......
             self.assertRaises(TypeError, store.append, 'df_unimplemented', df)
 
-    def test_table_append_with_timezones(self):
+    def test_append_with_timezones(self):
 
         from datetime import timedelta
 
@@ -1798,6 +1799,51 @@ def compare(a,b):
             result = store.select('df')
             assert_frame_equal(result,df)
 
+    def test_append_with_timedelta(self):
+        if com._np_version_under1p7:
+            raise nose.SkipTest("requires numpy >= 1.7")
+
+        # GH 3577
+        # append timedelta
+
+        from datetime import timedelta
+        df = DataFrame(dict(A = Timestamp('20130101'), B = [ Timestamp('20130101') + timedelta(days=i,seconds=10) for i in range(10) ]))
+        df['C'] = df['A']-df['B']
+        df.ix[3:5,'C'] = np.nan
+
+        with ensure_clean(self.path) as store:
+
+            # table
+            _maybe_remove(store, 'df')
+            store.append('df',df,data_columns=True)
+            result = store.select('df')
+            assert_frame_equal(result,df)
+
+            result = store.select('df',Term("C<100000"))
+            assert_frame_equal(result,df)
+
+            result = store.select('df',Term("C","<",-3*86400))
+            assert_frame_equal(result,df.iloc[3:])
+
+            result = store.select('df',Term("C","<",'-3D'))
+            assert_frame_equal(result,df.iloc[3:])
+
+            # a bit hacky here as we don't really deal with the NaT properly
+
+            result = store.select('df',Term("C","<",'-500000s'))
+            result = result.dropna(subset=['C'])
+            assert_frame_equal(result,df.iloc[6:])
+
+            result = store.select('df',Term("C","<",'-3.5D'))
+            result = result.iloc[1:]
+            assert_frame_equal(result,df.iloc[4:])
+
+            # fixed
+            _maybe_remove(store, 'df2')
+            store.put('df2',df)
+            result = store.select('df2')
+            assert_frame_equal(result,df)
+
     def test_remove(self):
 
         with ensure_clean(self.path) as store:
diff --git a/pandas/tslib.pyx b/pandas/tslib.pyx
index 983d3385e8f85..b145bd8fba507 100644
--- a/pandas/tslib.pyx
+++ b/pandas/tslib.pyx
@@ -1276,7 +1276,7 @@ cdef inline _get_datetime64_nanos(object val):
     else:
         return ival
 
-cdef inline int64_t cast_from_unit(object unit, object ts) except -1:
+cpdef inline int64_t cast_from_unit(object unit, object ts) except -1:
     """ return a casting of the unit represented to nanoseconds
         round the fractional part of a float to our precision, p """
     if unit == 'D':

From d8280c12a3449b66beaf0459bf86e71e99a44b2b Mon Sep 17 00:00:00 2001
From: jreback <jeff@reback.net>
Date: Thu, 12 Sep 2013 07:41:47 -0400
Subject: [PATCH 2/4] CLN: refactored locations of timedeltas to
 core/tseries/timedeltas (from a series of functions in core/common)

---
 pandas/__init__.py               |  13 ++++
 pandas/core/common.py            | 119 ++-----------------------------
 pandas/core/generic.py           |   4 +-
 pandas/core/series.py            |  16 +++--
 pandas/io/pytables.py            |   7 +-
 pandas/io/tests/test_pytables.py |   4 +-
 pandas/tests/test_frame.py       |   5 +-
 pandas/tests/test_series.py      |  14 ++--
 pandas/tseries/index.py          |   3 +-
 pandas/tseries/timedeltas.py     | 111 ++++++++++++++++++++++++++++
 10 files changed, 156 insertions(+), 140 deletions(-)
 create mode 100644 pandas/tseries/timedeltas.py

diff --git a/pandas/__init__.py b/pandas/__init__.py
index a0edb397c28c1..03681d3fa5a3f 100644
--- a/pandas/__init__.py
+++ b/pandas/__init__.py
@@ -18,6 +18,19 @@
 from datetime import datetime
 import numpy as np
 
+# XXX: HACK for NumPy 1.5.1 to suppress warnings
+try:
+    np.seterr(all='ignore')
+    # np.set_printoptions(suppress=True)
+except Exception:  # pragma: no cover
+    pass
+
+# numpy versioning
+from distutils.version import LooseVersion
+_np_version = np.version.short_version
+_np_version_under1p6 = LooseVersion(_np_version) < '1.6'
+_np_version_under1p7 = LooseVersion(_np_version) < '1.7'
+
 from pandas.version import version as __version__
 from pandas.info import __doc__
 
diff --git a/pandas/core/common.py b/pandas/core/common.py
index 2f153e88f4dc6..b58bd92a4fd1f 100644
--- a/pandas/core/common.py
+++ b/pandas/core/common.py
@@ -11,7 +11,6 @@
 import pandas.algos as algos
 import pandas.lib as lib
 import pandas.tslib as tslib
-from distutils.version import LooseVersion
 from pandas import compat
 from pandas.compat import StringIO, BytesIO, range, long, u, zip, map
 from datetime import timedelta
@@ -19,14 +18,6 @@
 from pandas.core.config import get_option
 from pandas.core import array as pa
 
-# XXX: HACK for NumPy 1.5.1 to suppress warnings
-try:
-    np.seterr(all='ignore')
-    # np.set_printoptions(suppress=True)
-except Exception:  # pragma: no cover
-    pass
-
-
 class PandasError(Exception):
     pass
 
@@ -34,11 +25,6 @@ class PandasError(Exception):
 class AmbiguousIndexError(PandasError, KeyError):
     pass
 
-# versioning
-_np_version = np.version.short_version
-_np_version_under1p6 = LooseVersion(_np_version) < '1.6'
-_np_version_under1p7 = LooseVersion(_np_version) < '1.7'
-
 _POSSIBLY_CAST_DTYPES = set([np.dtype(t)
                             for t in ['M8[ns]', 'm8[ns]', 'O', 'int8', 'uint8', 'int16', 'uint16', 'int32', 'uint32', 'int64', 'uint64']])
 
@@ -703,50 +689,13 @@ def diff(arr, n, axis=0):
 
     return out_arr
 
-timedelta_search = re.compile(
-    "^(?P<value>-?\d*\.?\d*)(?P<unit>D|s|ms|us|ns)?$")
-
-def _coerce_scalar_to_timedelta_type(r, unit='ns'):
-    # kludgy here until we have a timedelta scalar
-    # handle the numpy < 1.7 case
-
-    if isinstance(r, compat.string_types):
-        m = timedelta_search.search(r)
-        if m:
-            r = float(m.groupdict()['value'])
-            u = m.groupdict().get('unit')
-            if u is not None:
-                unit = u
-        else:
-            raise ValueError("cannot convert timedelta scalar value!")
-
-        r = tslib.cast_from_unit(unit, r)
-        r = timedelta(microseconds=int(r)/1000)
-
-    if is_integer(r):
-        r = tslib.cast_from_unit(unit, r)
-        r = timedelta(microseconds=int(r)/1000)
-
-    if _np_version_under1p7:
-        if not isinstance(r, timedelta):
-            raise AssertionError("Invalid type for timedelta scalar: %s" % type(r))
-        if compat.PY3:
-            # convert to microseconds in timedelta64
-            r = np.timedelta64(int(r.total_seconds()*1e9 + r.microseconds*1000))
-        else:
-            return r
-
-    if isinstance(r, timedelta):
-        r = np.timedelta64(r)
-    elif not isinstance(r, np.timedelta64):
-        raise AssertionError("Invalid type for timedelta scalar: %s" % type(r))
-    return r.astype('timedelta64[ns]')
-
 def _coerce_to_dtypes(result, dtypes):
     """ given a dtypes and a result set, coerce the result elements to the dtypes """
     if len(result) != len(dtypes):
         raise AssertionError("_coerce_to_dtypes requires equal len arrays")
 
+    from pandas.tseries.timedeltas import _coerce_scalar_to_timedelta_type
+
     def conv(r,dtype):
         try:
             if isnull(r):
@@ -1339,68 +1288,6 @@ def _possibly_convert_platform(values):
 
     return values
 
-
-def _possibly_cast_to_timedelta(value, coerce=True):
-    """ try to cast to timedelta64, if already a timedeltalike, then make
-        sure that we are [ns] (as numpy 1.6.2 is very buggy in this regards,
-        don't force the conversion unless coerce is True
-
-        if coerce='compat' force a compatibilty coercerion (to timedeltas) if needeed
-        """
-
-    # coercion compatability
-    if coerce == 'compat' and _np_version_under1p7:
-
-        def convert(td, dtype):
-
-            # we have an array with a non-object dtype
-            if hasattr(td,'item'):
-                td = td.astype(np.int64).item()
-                if td == tslib.iNaT:
-                    return td
-                if dtype == 'm8[us]':
-                    td *= 1000
-                return td
-
-            if td == tslib.compat_NaT:
-                return tslib.iNaT
-
-            # convert td value to a nanosecond value
-            d = td.days
-            s = td.seconds
-            us = td.microseconds
-
-            if dtype == 'object' or dtype == 'm8[ns]':
-                td = 1000*us + (s + d * 24 * 3600) * 10 ** 9
-            else:
-                raise ValueError("invalid conversion of dtype in np < 1.7 [%s]" % dtype)
-
-            return td
-
-        # < 1.7 coercion
-        if not is_list_like(value):
-            value = np.array([ value ])
-
-        dtype = value.dtype
-        return np.array([ convert(v,dtype) for v in value ], dtype='m8[ns]')
-
-    # deal with numpy not being able to handle certain timedelta operations
-    if isinstance(value, (ABCSeries, np.ndarray)) and value.dtype.kind == 'm':
-        if value.dtype != 'timedelta64[ns]':
-            value = value.astype('timedelta64[ns]')
-        return value
-
-    # we don't have a timedelta, but we want to try to convert to one (but
-    # don't force it)
-    if coerce:
-        new_value = tslib.array_to_timedelta64(
-            _values_from_object(value).astype(object), coerce=False)
-        if new_value.dtype == 'i8':
-            value = np.array(new_value, dtype='timedelta64[ns]')
-
-    return value
-
-
 def _possibly_cast_to_datetime(value, dtype, coerce=False):
     """ try to cast the array/value to a datetimelike dtype, converting float nan to iNaT """
 
@@ -1438,6 +1325,7 @@ def _possibly_cast_to_datetime(value, dtype, coerce=False):
                             from pandas.tseries.tools import to_datetime
                             value = to_datetime(value, coerce=coerce).values
                         elif is_timedelta64:
+                            from pandas.tseries.timedeltas import _possibly_cast_to_timedelta
                             value = _possibly_cast_to_timedelta(value)
                     except:
                         pass
@@ -1463,6 +1351,7 @@ def _possibly_cast_to_datetime(value, dtype, coerce=False):
                     except:
                         pass
                 elif inferred_type in ['timedelta', 'timedelta64']:
+                    from pandas.tseries.timedeltas import _possibly_cast_to_timedelta
                     value = _possibly_cast_to_timedelta(value)
 
     return value
diff --git a/pandas/core/generic.py b/pandas/core/generic.py
index 212e2bad563b6..b9ffe788d183d 100644
--- a/pandas/core/generic.py
+++ b/pandas/core/generic.py
@@ -13,7 +13,7 @@
 from pandas.tseries.index import DatetimeIndex
 from pandas.core.internals import BlockManager
 import pandas.core.common as com
-from pandas import compat
+from pandas import compat, _np_version_under1p7
 from pandas.compat import map, zip, lrange
 from pandas.core.common import (isnull, notnull, is_list_like,
                                 _values_from_object,
@@ -1908,7 +1908,7 @@ def abs(self):
         obj = np.abs(self)
 
         # suprimo numpy 1.6 hacking
-        if com._np_version_under1p7:
+        if _np_version_under1p7:
             if self.ndim == 1:
                 if obj.dtype == 'm8[us]':
                     obj = obj.astype('m8[ns]')
diff --git a/pandas/core/series.py b/pandas/core/series.py
index 4516fcfbaee8e..8d6591c3acd60 100644
--- a/pandas/core/series.py
+++ b/pandas/core/series.py
@@ -19,6 +19,7 @@
                                 _asarray_tuplesafe, is_integer_dtype,
                                 _NS_DTYPE, _TD_DTYPE,
                                 _infer_dtype_from_scalar, is_list_like, _values_from_object,
+                                _possibly_cast_to_datetime, _possibly_castable, _possibly_convert_platform,
                                 ABCSparseArray)
 from pandas.core.index import (Index, MultiIndex, InvalidIndexError,
                                _ensure_index, _handle_legacy_indexes)
@@ -32,6 +33,7 @@
 from pandas.tseries.index import DatetimeIndex
 from pandas.tseries.period import PeriodIndex, Period
 from pandas.tseries.offsets import DateOffset
+from pandas.tseries.timedeltas import _possibly_cast_to_timedelta
 from pandas import compat
 from pandas.util.terminal import get_terminal_size
 from pandas.compat import zip, lzip, u, OrderedDict
@@ -142,7 +144,7 @@ def _convert_to_array(self, values, name=None):
                 values = values.to_series()
         elif inferred_type in ('timedelta', 'timedelta64'):
             # have a timedelta, convert to to ns here
-            values = com._possibly_cast_to_timedelta(values, coerce=coerce)
+            values = _possibly_cast_to_timedelta(values, coerce=coerce)
         elif inferred_type == 'integer':
             # py3 compat where dtype is 'm' but is an integer
             if values.dtype.kind == 'm':
@@ -160,7 +162,7 @@ def _convert_to_array(self, values, name=None):
                 raise TypeError("cannot use a non-absolute DateOffset in "
                                 "datetime/timedelta operations [{0}]".format(
                                     ','.join([ com.pprint_thing(v) for v in values[mask] ])))
-            values = com._possibly_cast_to_timedelta(os, coerce=coerce)
+            values = _possibly_cast_to_timedelta(os, coerce=coerce)
         else:
             raise TypeError("incompatible type [{0}] for a datetime/timedelta operation".format(pa.array(values).dtype))
 
@@ -3215,11 +3217,11 @@ def _try_cast(arr, take_fast_path):
 
         # perf shortcut as this is the most common case
         if take_fast_path:
-            if com._possibly_castable(arr) and not copy and dtype is None:
+            if _possibly_castable(arr) and not copy and dtype is None:
                 return arr
 
         try:
-            arr = com._possibly_cast_to_datetime(arr, dtype)
+            arr = _possibly_cast_to_datetime(arr, dtype)
             subarr = pa.array(arr, dtype=dtype, copy=copy)
         except (ValueError, TypeError):
             if dtype is not None and raise_cast_failure:
@@ -3266,9 +3268,9 @@ def _try_cast(arr, take_fast_path):
                 subarr = lib.maybe_convert_objects(subarr)
 
         else:
-            subarr = com._possibly_convert_platform(data)
+            subarr = _possibly_convert_platform(data)
 
-        subarr = com._possibly_cast_to_datetime(subarr, dtype)
+        subarr = _possibly_cast_to_datetime(subarr, dtype)
 
     else:
         subarr = _try_cast(data, False)
@@ -3285,7 +3287,7 @@ def _try_cast(arr, take_fast_path):
                 dtype, value = _infer_dtype_from_scalar(value)
             else:
                 # need to possibly convert the value here
-                value = com._possibly_cast_to_datetime(value, dtype)
+                value = _possibly_cast_to_datetime(value, dtype)
 
             subarr = pa.empty(len(index), dtype=dtype)
             subarr.fill(value)
diff --git a/pandas/io/pytables.py b/pandas/io/pytables.py
index bf45fbf96c0e5..9b6a230f6a551 100644
--- a/pandas/io/pytables.py
+++ b/pandas/io/pytables.py
@@ -17,7 +17,7 @@
 import numpy as np
 import pandas
 from pandas import (Series, TimeSeries, DataFrame, Panel, Panel4D, Index,
-                    MultiIndex, Int64Index, Timestamp)
+                    MultiIndex, Int64Index, Timestamp, _np_version_under1p7)
 from pandas.sparse.api import SparseSeries, SparseDataFrame, SparsePanel
 from pandas.sparse.array import BlockIndex, IntIndex
 from pandas.tseries.api import PeriodIndex, DatetimeIndex
@@ -25,10 +25,11 @@
 from pandas.core.common import adjoin, is_list_like, pprint_thing
 from pandas.core.algorithms import match, unique
 from pandas.core.categorical import Categorical
-from pandas.core.common import _asarray_tuplesafe, _np_version_under1p7
+from pandas.core.common import _asarray_tuplesafe
 from pandas.core.internals import BlockManager, make_block
 from pandas.core.reshape import block2d_to_blocknd, factor_indexer
 from pandas.core.index import _ensure_index
+from pandas.tseries.timedeltas import _coerce_scalar_to_timedelta_type
 import pandas.core.common as com
 from pandas.tools.merge import concat
 from pandas import compat
@@ -4093,7 +4094,7 @@ def stringify(value):
                 v = v.tz_convert('UTC')
             return TermValue(v, v.value, kind)
         elif kind == u('timedelta64') or kind == u('timedelta'):
-            v = com._coerce_scalar_to_timedelta_type(v,unit='s').item()
+            v = _coerce_scalar_to_timedelta_type(v,unit='s').item()
             return TermValue(int(v), v, kind)
         elif (isinstance(v, datetime) or hasattr(v, 'timetuple')
               or kind == u('date')):
diff --git a/pandas/io/tests/test_pytables.py b/pandas/io/tests/test_pytables.py
index 3667dff994232..3f4ce72198215 100644
--- a/pandas/io/tests/test_pytables.py
+++ b/pandas/io/tests/test_pytables.py
@@ -22,7 +22,7 @@
                                  assert_frame_equal,
                                  assert_series_equal)
 from pandas import concat, Timestamp
-from pandas import compat
+from pandas import compat, _np_version_under1p7
 from pandas.core import common as com
 
 from numpy.testing.decorators import slow
@@ -1800,7 +1800,7 @@ def compare(a,b):
             assert_frame_equal(result,df)
 
     def test_append_with_timedelta(self):
-        if com._np_version_under1p7:
+        if _np_version_under1p7:
             raise nose.SkipTest("requires numpy >= 1.7")
 
         # GH 3577
diff --git a/pandas/tests/test_frame.py b/pandas/tests/test_frame.py
index 723810a19d140..c5af0b0d4d5c8 100644
--- a/pandas/tests/test_frame.py
+++ b/pandas/tests/test_frame.py
@@ -3248,9 +3248,10 @@ def test_operators_timedelta64(self):
         mixed['F'] = Timestamp('20130101')
 
         # results in an object array
+        from pandas.tseries.timedeltas import _coerce_scalar_to_timedelta_type
         result = mixed.min()
-        expected = Series([com._coerce_scalar_to_timedelta_type(timedelta(seconds=5*60+5)),
-                           com._coerce_scalar_to_timedelta_type(timedelta(days=-1)),
+        expected = Series([_coerce_scalar_to_timedelta_type(timedelta(seconds=5*60+5)),
+                           _coerce_scalar_to_timedelta_type(timedelta(days=-1)),
                            'foo',
                            1,
                            1.0,
diff --git a/pandas/tests/test_series.py b/pandas/tests/test_series.py
index 1f008354756bc..7a993cbcf07f4 100644
--- a/pandas/tests/test_series.py
+++ b/pandas/tests/test_series.py
@@ -14,7 +14,7 @@
 import pandas as pd
 
 from pandas import (Index, Series, DataFrame, isnull, notnull,
-                    bdate_range, date_range)
+                    bdate_range, date_range, _np_version_under1p7)
 from pandas.core.index import MultiIndex
 from pandas.tseries.index import Timestamp, DatetimeIndex
 import pandas.core.config as cf
@@ -2188,7 +2188,7 @@ def test_timedeltas_with_DateOffset(self):
             [Timestamp('20130101 9:06:00.005'), Timestamp('20130101 9:07:00.005')])
         assert_series_equal(result, expected)
 
-        if not com._np_version_under1p7:
+        if not _np_version_under1p7:
 
             # operate with np.timedelta64 correctly
             result = s + np.timedelta64(1, 's')
@@ -2292,7 +2292,7 @@ def test_timedelta64_operations_with_integers(self):
                 self.assertRaises(TypeError, sop, s2.values)
 
     def test_timedelta64_conversions(self):
-        if com._np_version_under1p7:
+        if _np_version_under1p7:
             raise nose.SkipTest("cannot use 2 argument form of timedelta64 conversions with numpy < 1.7")
 
         startdate = Series(date_range('2013-01-01', '2013-01-03'))
@@ -2317,7 +2317,7 @@ def test_timedelta64_equal_timedelta_supported_ops(self):
                           'm': 60 * 1000000, 's': 1000000, 'us': 1}
 
         def timedelta64(*args):
-            if com._np_version_under1p7:
+            if _np_version_under1p7:
                 coeffs = np.array(args)
                 terms = np.array([npy16_mappings[interval]
                                   for interval in intervals])
@@ -2426,7 +2426,7 @@ def test_timedelta64_functions(self):
         assert_series_equal(result, expected)
 
     def test_timedelta_fillna(self):
-        if com._np_version_under1p7:
+        if _np_version_under1p7:
             raise nose.SkipTest("timedelta broken in np 1.6.1")
 
         #GH 3371
@@ -2498,12 +2498,12 @@ def test_datetime64_fillna(self):
         assert_series_equal(result,expected)
 
     def test_sub_of_datetime_from_TimeSeries(self):
-        from pandas.core import common as com
+        from pandas.tseries.timedeltas import _possibly_cast_to_timedelta
         from datetime import datetime
         a = Timestamp(datetime(1993, 0o1, 0o7, 13, 30, 00))
         b = datetime(1993, 6, 22, 13, 30)
         a = Series([a])
-        result = com._possibly_cast_to_timedelta(np.abs(a - b))
+        result = _possibly_cast_to_timedelta(np.abs(a - b))
         self.assert_(result.dtype == 'timedelta64[ns]')
 
     def test_datetime64_with_index(self):
diff --git a/pandas/tseries/index.py b/pandas/tseries/index.py
index e91cad62e7dce..1572ca481d8a4 100644
--- a/pandas/tseries/index.py
+++ b/pandas/tseries/index.py
@@ -7,8 +7,7 @@
 import numpy as np
 
 from pandas.core.common import (isnull, _NS_DTYPE, _INT64_DTYPE,
-                                is_list_like,_possibly_cast_to_timedelta,
-                                _values_from_object, _maybe_box)
+                                is_list_like,_values_from_object, _maybe_box)
 from pandas.core.index import Index, Int64Index
 import pandas.compat as compat
 from pandas.compat import u
diff --git a/pandas/tseries/timedeltas.py b/pandas/tseries/timedeltas.py
new file mode 100644
index 0000000000000..990dd9c351e27
--- /dev/null
+++ b/pandas/tseries/timedeltas.py
@@ -0,0 +1,111 @@
+"""
+timedelta support tools
+"""
+
+import re
+from datetime import timedelta
+
+import numpy as np
+import pandas.tslib as tslib
+from pandas import compat, _np_version_under1p7
+from pandas.core.common import ABCSeries, is_integer, _values_from_object
+
+timedelta_search = re.compile(
+    "^(?P<value>-?\d*\.?\d*)(?P<unit>D|s|ms|us|ns)?$")
+
+def _coerce_scalar_to_timedelta_type(r, unit='ns'):
+    # kludgy here until we have a timedelta scalar
+    # handle the numpy < 1.7 case
+
+    if isinstance(r, compat.string_types):
+        m = timedelta_search.search(r)
+        if m:
+            r = float(m.groupdict()['value'])
+            u = m.groupdict().get('unit')
+            if u is not None:
+                unit = u
+        else:
+            raise ValueError("cannot convert timedelta scalar value!")
+
+        r = tslib.cast_from_unit(unit, r)
+        r = timedelta(microseconds=int(r)/1000)
+
+    if is_integer(r):
+        r = tslib.cast_from_unit(unit, r)
+        r = timedelta(microseconds=int(r)/1000)
+
+    if _np_version_under1p7:
+        if not isinstance(r, timedelta):
+            raise AssertionError("Invalid type for timedelta scalar: %s" % type(r))
+        if compat.PY3:
+            # convert to microseconds in timedelta64
+            r = np.timedelta64(int(r.total_seconds()*1e9 + r.microseconds*1000))
+        else:
+            return r
+
+    if isinstance(r, timedelta):
+        r = np.timedelta64(r)
+    elif not isinstance(r, np.timedelta64):
+        raise AssertionError("Invalid type for timedelta scalar: %s" % type(r))
+    return r.astype('timedelta64[ns]')
+
+def _possibly_cast_to_timedelta(value, coerce=True):
+    """ try to cast to timedelta64, if already a timedeltalike, then make
+        sure that we are [ns] (as numpy 1.6.2 is very buggy in this regards,
+        don't force the conversion unless coerce is True
+
+        if coerce='compat' force a compatibilty coercerion (to timedeltas) if needeed
+        """
+
+    # coercion compatability
+    if coerce == 'compat' and _np_version_under1p7:
+
+        def convert(td, dtype):
+
+            # we have an array with a non-object dtype
+            if hasattr(td,'item'):
+                td = td.astype(np.int64).item()
+                if td == tslib.iNaT:
+                    return td
+                if dtype == 'm8[us]':
+                    td *= 1000
+                return td
+
+            if td == tslib.compat_NaT:
+                return tslib.iNaT
+
+            # convert td value to a nanosecond value
+            d = td.days
+            s = td.seconds
+            us = td.microseconds
+
+            if dtype == 'object' or dtype == 'm8[ns]':
+                td = 1000*us + (s + d * 24 * 3600) * 10 ** 9
+            else:
+                raise ValueError("invalid conversion of dtype in np < 1.7 [%s]" % dtype)
+
+            return td
+
+        # < 1.7 coercion
+        if not is_list_like(value):
+            value = np.array([ value ])
+
+        dtype = value.dtype
+        return np.array([ convert(v,dtype) for v in value ], dtype='m8[ns]')
+
+    # deal with numpy not being able to handle certain timedelta operations
+    if isinstance(value, (ABCSeries, np.ndarray)) and value.dtype.kind == 'm':
+        if value.dtype != 'timedelta64[ns]':
+            value = value.astype('timedelta64[ns]')
+        return value
+
+    # we don't have a timedelta, but we want to try to convert to one (but
+    # don't force it)
+    if coerce:
+        new_value = tslib.array_to_timedelta64(
+            _values_from_object(value).astype(object), coerce=False)
+        if new_value.dtype == 'i8':
+            value = np.array(new_value, dtype='timedelta64[ns]')
+
+    return value
+

From 30d91c75919ae6cd9ecbe821ceb3c3b4036688d8 Mon Sep 17 00:00:00 2001
From: jreback <jeff@reback.net>
Date: Thu, 12 Sep 2013 08:52:16 -0400
Subject: [PATCH 3/4] TST: add pandas/tseries/tests/test_timedeltas.py API: add
 full timedelta parsing and conversion to np.timedelta64[ns]

---
 doc/source/release.rst                  |   1 +
 pandas/tseries/tests/test_timedeltas.py | 128 ++++++++++++++++++++++++
 pandas/tseries/timedeltas.py            |  78 ++++++++++++---
 pandas/tslib.pyx                        |  12 +--
 4 files changed, 197 insertions(+), 22 deletions(-)
 create mode 100644 pandas/tseries/tests/test_timedeltas.py

diff --git a/doc/source/release.rst b/doc/source/release.rst
index d50438cd08058..2ed866e1fdeda 100644
--- a/doc/source/release.rst
+++ b/doc/source/release.rst
@@ -191,6 +191,7 @@ API Changes
   - provide automatic dtype conversions on _reduce operations (:issue:`3371`)
   - exclude non-numerics if mixed types with datelike in _reduce operations (:issue:`3371`)
   - default for ``tupleize_cols`` is now ``False`` for both ``to_csv`` and ``read_csv``. Fair warning in 0.12 (:issue:`3604`)
+  - moved timedeltas support to pandas.tseries.timedeltas.py; add timedeltas string parsing
 
 Internal Refactoring
 ~~~~~~~~~~~~~~~~~~~~
diff --git a/pandas/tseries/tests/test_timedeltas.py b/pandas/tseries/tests/test_timedeltas.py
new file mode 100644
index 0000000000000..8b88da1d7ef28
--- /dev/null
+++ b/pandas/tseries/tests/test_timedeltas.py
@@ -0,0 +1,128 @@
+# pylint: disable-msg=E1101,W0612
+
+from datetime import datetime, timedelta
+import nose
+import unittest
+
+import numpy as np
+import pandas as pd
+
+from pandas import (Index, Series, DataFrame, isnull, notnull,
+                    bdate_range, date_range, _np_version_under1p7)
+import pandas.core.common as com
+from pandas.compat import StringIO, lrange, range, zip, u, OrderedDict, long
+from pandas import compat
+from pandas.tseries.timedeltas import _coerce_scalar_to_timedelta_type as ct
+from pandas.util.testing import (assert_series_equal,
+                                 assert_frame_equal,
+                                 assert_almost_equal,
+                                 ensure_clean)
+import pandas.util.testing as tm
+
+def _skip_if_numpy_not_friendly():
+    # not friendly for < 1.7
+    if _np_version_under1p7:
+        raise nose.SkipTest("numpy < 1.7")
+
+class TestTimedeltas(unittest.TestCase):
+    _multiprocess_can_split_ = True
+
+    def setUp(self):
+        pass
+
+    def test_numeric_conversions(self):
+        _skip_if_numpy_not_friendly()
+
+        # ns not converted properly
+        self.assert_(ct(0) == np.timedelta64(0,'ns'))
+        self.assert_(ct(10) == np.timedelta64(0,'ns'))
+        self.assert_(ct(10,unit='ns') == np.timedelta64(0,'ns').astype('m8[ns]'))
+
+        self.assert_(ct(10,unit='us') == np.timedelta64(10,'us').astype('m8[ns]'))
+        self.assert_(ct(10,unit='ms') == np.timedelta64(10,'ms').astype('m8[ns]'))
+        self.assert_(ct(10,unit='s') == np.timedelta64(10,'s').astype('m8[ns]'))
+        self.assert_(ct(10,unit='d') == np.timedelta64(10,'D').astype('m8[ns]'))
+
+    def test_timedelta_conversions(self):
+        _skip_if_numpy_not_friendly()
+
+        self.assert_(ct(timedelta(seconds=1)) == np.timedelta64(1,'s').astype('m8[ns]'))
+        self.assert_(ct(timedelta(microseconds=1)) == np.timedelta64(1,'us').astype('m8[ns]'))
+        self.assert_(ct(timedelta(days=1)) == np.timedelta64(1,'D').astype('m8[ns]'))
+
+    def test_short_format_converters(self):
+        _skip_if_numpy_not_friendly()
+
+        def conv(v):
+            return v.astype('m8[ns]')
+
+        # ns not converted properly
+        self.assert_(ct('10') == np.timedelta64(0,'ns'))
+        self.assert_(ct('10ns') == np.timedelta64(0,'ns'))
+        self.assert_(ct('100') == np.timedelta64(0,'ns'))
+        self.assert_(ct('100ns') == np.timedelta64(0,'ns'))
+
+        self.assert_(ct('1000') == np.timedelta64(1000,'ns'))
+        self.assert_(ct('1000ns') == np.timedelta64(1000,'ns'))
+        self.assert_(ct('1000NS') == np.timedelta64(1000,'ns'))
+
+        self.assert_(ct('10us') == np.timedelta64(10000,'ns'))
+        self.assert_(ct('100us') == np.timedelta64(100000,'ns'))
+        self.assert_(ct('1000us') == np.timedelta64(1000000,'ns'))
+        self.assert_(ct('1000Us') == np.timedelta64(1000000,'ns'))
+        self.assert_(ct('1000uS') == np.timedelta64(1000000,'ns'))
+
+        self.assert_(ct('1ms') == np.timedelta64(1000000,'ns'))
+        self.assert_(ct('10ms') == np.timedelta64(10000000,'ns'))
+        self.assert_(ct('100ms') == np.timedelta64(100000000,'ns'))
+        self.assert_(ct('1000ms') == np.timedelta64(1000000000,'ns'))
+
+        self.assert_(ct('-1s') == -np.timedelta64(1000000000,'ns'))
+        self.assert_(ct('1s') == np.timedelta64(1000000000,'ns'))
+        self.assert_(ct('10s') == np.timedelta64(10000000000,'ns'))
+        self.assert_(ct('100s') == np.timedelta64(100000000000,'ns'))
+        self.assert_(ct('1000s') == np.timedelta64(1000000000000,'ns'))
+
+        self.assert_(ct('1d') == conv(np.timedelta64(1,'D')))
+        self.assert_(ct('-1d') == -conv(np.timedelta64(1,'D')))
+        self.assert_(ct('1D') == conv(np.timedelta64(1,'D')))
+        self.assert_(ct('10D') == conv(np.timedelta64(10,'D')))
+        self.assert_(ct('100D') == conv(np.timedelta64(100,'D')))
+        self.assert_(ct('1000D') == conv(np.timedelta64(1000,'D')))
+        self.assert_(ct('10000D') == conv(np.timedelta64(10000,'D')))
+
+        # space
+        self.assert_(ct(' 10000D ') == conv(np.timedelta64(10000,'D')))
+        self.assert_(ct(' - 10000D ') == -conv(np.timedelta64(10000,'D')))
+
+        # invalid
+        self.assertRaises(ValueError, ct, '1foo')
+        self.assertRaises(ValueError, ct, 'foo')
+
+    def test_full_format_converters(self):
+        _skip_if_numpy_not_friendly()
+
+        def conv(v):
+            return v.astype('m8[ns]')
+        d1 = np.timedelta64(1,'D')
+
+        self.assert_(ct('1days') == conv(d1))
+        self.assert_(ct('1days,') == conv(d1))
+        self.assert_(ct('- 1days,') == -conv(d1))
+
+        self.assert_(ct('00:00:01') == conv(np.timedelta64(1,'s')))
+        self.assert_(ct('06:00:01') == conv(np.timedelta64(6*3600+1,'s')))
+        self.assert_(ct('06:00:01.0') == conv(np.timedelta64(6*3600+1,'s')))
+        self.assert_(ct('06:00:01.01') == conv(np.timedelta64(1000*(6*3600+1)+10,'ms')))
+
+        self.assert_(ct('- 1days, 00:00:01') == -conv(d1+np.timedelta64(1,'s')))
+        self.assert_(ct('1days, 06:00:01') == conv(d1+np.timedelta64(6*3600+1,'s')))
+        self.assert_(ct('1days, 06:00:01.01') == conv(d1+np.timedelta64(1000*(6*3600+1)+10,'ms')))
+
+        # invalid
+        self.assertRaises(ValueError, ct, '- 1days, 00')
+
+
+if __name__ == '__main__':
+    nose.runmodule(argv=[__file__, '-vvs', '-x', '--pdb', '--pdb-failure'],
+                   exit=False)
diff --git a/pandas/tseries/timedeltas.py b/pandas/tseries/timedeltas.py
index 990dd9c351e27..d12791755859e 100644
--- a/pandas/tseries/timedeltas.py
+++ b/pandas/tseries/timedeltas.py
@@ -8,31 +8,29 @@
 import numpy as np
 import pandas.tslib as tslib
 from pandas import compat, _np_version_under1p7
-from pandas.core.common import ABCSeries, is_integer, _values_from_object
+from pandas.core.common import (ABCSeries, is_integer,
+                                _values_from_object, is_list_like)
 
-timedelta_search = re.compile(
-    "^(?P<value>-?\d*\.?\d*)(?P<unit>D|s|ms|us|ns)?$")
+repr_timedelta = tslib.repr_timedelta64
+repr_timedelta64 = tslib.repr_timedelta64
+
+_short_search = re.compile(
+    "^\s*(?P<neg>-?)\s*(?P<value>\d*\.?\d*)\s*(?P<unit>d|s|ms|us|ns)?\s*$",re.IGNORECASE)
+_full_search = re.compile(
+    "^\s*(?P<neg>-?)\s*(?P<days>\d+)?\s*(days|d)?,?\s*(?P<time>\d{2}:\d{2}:\d{2})?(?P<frac>\.\d+)?\s*$",re.IGNORECASE)
 
 def _coerce_scalar_to_timedelta_type(r, unit='ns'):
     # kludgy here until we have a timedelta scalar
     # handle the numpy < 1.7 case
 
     if isinstance(r, compat.string_types):
-        m = timedelta_search.search(r)
-        if m:
-            r = float(m.groupdict()['value'])
-            u = m.groupdict().get('unit')
-            if u is not None:
-                unit = u
-        else:
-            raise ValueError("cannot convert timedelta scalar value!")
-
-        r = tslib.cast_from_unit(unit, r)
-        r = timedelta(microseconds=int(r)/1000)
+        converter = _get_string_converter(r, unit=unit)
+        r = converter()
+        r = timedelta(microseconds=r/1000.0)
 
     if is_integer(r):
-        r = tslib.cast_from_unit(unit, r)
-        r = timedelta(microseconds=int(r)/1000)
+        r = tslib.cast_from_unit(r, unit)
+        r = timedelta(microseconds=r/1000.0)
 
     if _np_version_under1p7:
         if not isinstance(r, timedelta):
@@ -49,6 +47,54 @@ def _coerce_scalar_to_timedelta_type(r, unit='ns'):
         raise AssertionError("Invalid type for timedelta scalar: %s" % type(r))
     return r.astype('timedelta64[ns]')
 
+def _get_string_converter(r, unit='ns'):
+    """ return a string converter for r to process the timedelta format """
+
+    m = _short_search.search(r)
+    if m:
+        def convert(r=None, unit=unit, m=m):
+            if r is not None:
+                m = _short_search.search(r)
+
+            gd = m.groupdict()
+
+            r = float(gd['value'])
+            u = gd.get('unit')
+            if u is not None:
+                unit = u.lower()
+            if gd['neg']:
+                r *= -1
+            return tslib.cast_from_unit(r, unit)
+        return convert
+
+    m = _full_search.search(r)
+    if m:
+        def convert(r=None, unit=None, m=m):
+            if r is not None:
+                m = _full_search.search(r)
+
+            gd = m.groupdict()
+
+            # convert to seconds
+            value = float(gd['days'] or 0) * 86400
+
+            time = gd['time']
+            if time:
+                (hh,mm,ss) = time.split(':')
+                value += float(hh)*3600 + float(mm)*60 + float(ss)
+
+            frac = gd['frac']
+            if frac:
+                value += float(frac)
+
+            if gd['neg']:
+                value *= -1
+            return tslib.cast_from_unit(value, 's')
+        return convert
+
+    # no converter
+    raise ValueError("cannot create timedelta string converter")
+
 def _possibly_cast_to_timedelta(value, coerce=True):
     """ try to cast to timedelta64, if already a timedeltalike, then make
         sure that we are [ns] (as numpy 1.6.2 is very buggy in this regards,
diff --git a/pandas/tslib.pyx b/pandas/tslib.pyx
index b145bd8fba507..8117ff7d3539b 100644
--- a/pandas/tslib.pyx
+++ b/pandas/tslib.pyx
@@ -665,14 +665,14 @@ cdef convert_to_tsobject(object ts, object tz, object unit):
         if ts == NPY_NAT:
             obj.value = NPY_NAT
         else:
-            ts = ts * cast_from_unit(unit,None)
+            ts = ts * cast_from_unit(None,unit)
             obj.value = ts
             pandas_datetime_to_datetimestruct(ts, PANDAS_FR_ns, &obj.dts)
     elif util.is_float_object(ts):
         if ts != ts or ts == NPY_NAT:
             obj.value = NPY_NAT
         else:
-            ts = cast_from_unit(unit,ts)
+            ts = cast_from_unit(ts,unit)
             obj.value = ts
             pandas_datetime_to_datetimestruct(ts, PANDAS_FR_ns, &obj.dts)
     elif util.is_string_object(ts):
@@ -852,7 +852,7 @@ def array_to_datetime(ndarray[object] values, raise_=False, dayfirst=False,
         pandas_datetimestruct dts
         bint utc_convert = bool(utc)
         _TSObject _ts
-        int64_t m = cast_from_unit(unit,None)
+        int64_t m = cast_from_unit(None,unit)
 
     try:
         result = np.empty(n, dtype='M8[ns]')
@@ -892,7 +892,7 @@ def array_to_datetime(ndarray[object] values, raise_=False, dayfirst=False,
                 if val != val or val == iNaT:
                     iresult[i] = iNaT
                 else:
-                    iresult[i] = cast_from_unit(unit,val)
+                    iresult[i] = cast_from_unit(val,unit)
             else:
                 try:
                     if len(val) == 0:
@@ -1276,10 +1276,10 @@ cdef inline _get_datetime64_nanos(object val):
     else:
         return ival
 
-cpdef inline int64_t cast_from_unit(object unit, object ts) except -1:
+cpdef inline int64_t cast_from_unit(object ts, object unit) except -1:
     """ return a casting of the unit represented to nanoseconds
         round the fractional part of a float to our precision, p """
-    if unit == 'D':
+    if unit == 'D' or unit == 'd':
         m = 1000000000L * 86400
         p = 6
     elif unit == 's':

From ef2cfb1f9d4f89c27ca759687fc1e13a4c44f1d0 Mon Sep 17 00:00:00 2001
From: jreback <jeff@reback.net>
Date: Thu, 12 Sep 2013 10:56:43 -0400
Subject: [PATCH 4/4] ENH: add top-level to_timedelta to convert string/integer
 based to timedeltas

---
 doc/source/io.rst                       |  4 ++
 doc/source/release.rst                  |  3 +-
 doc/source/timeseries.rst               | 20 +++++++
 doc/source/v0.13.0.txt                  | 15 +++++
 pandas/tseries/api.py                   |  1 +
 pandas/tseries/tests/test_timedeltas.py | 58 +++++++++++++++---
 pandas/tseries/timedeltas.py            | 79 +++++++++++++++++++++++--
 pandas/tslib.pyx                        |  4 +-
 8 files changed, 168 insertions(+), 16 deletions(-)

diff --git a/doc/source/io.rst b/doc/source/io.rst
index 94c8343bfec1d..da611c0375789 100644
--- a/doc/source/io.rst
+++ b/doc/source/io.rst
@@ -2017,6 +2017,10 @@ Beginning in 0.13.0, you can store and query using the ``timedelta64[ns]`` type.
 specified in the format: ``<float>(<unit>)``, where float may be signed (and fractional), and unit can be
 ``D,s,ms,us,ns`` for the timedelta. Here's an example:
 
+.. warning::
+
+   This requires ``numpy >= 1.7``
+
 .. ipython:: python
 
    from datetime import timedelta
diff --git a/doc/source/release.rst b/doc/source/release.rst
index 2ed866e1fdeda..75194f6877a6e 100644
--- a/doc/source/release.rst
+++ b/doc/source/release.rst
@@ -191,7 +191,8 @@ API Changes
   - provide automatic dtype conversions on _reduce operations (:issue:`3371`)
   - exclude non-numerics if mixed types with datelike in _reduce operations (:issue:`3371`)
   - default for ``tupleize_cols`` is now ``False`` for both ``to_csv`` and ``read_csv``. Fair warning in 0.12 (:issue:`3604`)
-  - moved timedeltas support to pandas.tseries.timedeltas.py; add timedeltas string parsing
+  - moved timedeltas support to pandas.tseries.timedeltas.py; add timedeltas string parsing,
+    add top-level ``to_timedelta`` function
 
 Internal Refactoring
 ~~~~~~~~~~~~~~~~~~~~
diff --git a/doc/source/timeseries.rst b/doc/source/timeseries.rst
index 11f4ac9f487c2..5dbf1ce77bad8 100644
--- a/doc/source/timeseries.rst
+++ b/doc/source/timeseries.rst
@@ -1211,6 +1211,26 @@ Time Deltas & Conversions
 
 .. versionadded:: 0.13
 
+**string/integer conversion**
+
+Using the top-level ``to_timedelta``, you can convert a scalar or array from the standard
+timedelta format (produced by ``to_csv``) into a timedelta type (``np.timedelta64`` in ``nanoseconds``).
+It can also construct Series.
+
+.. warning::
+
+   This requires ``numpy >= 1.7``
+
+.. ipython:: python
+
+   to_timedelta('1 days 06:05:01.00003')
+   to_timedelta('15.5us')
+   to_timedelta(['1 days 06:05:01.00003','15.5us','nan'])
+   to_timedelta(np.arange(5),unit='s')
+   to_timedelta(np.arange(5),unit='d')
+
+**frequency conversion**
+
 Timedeltas can be converted to other 'frequencies' by dividing by another timedelta.
 These operations yield ``float64`` dtyped Series.
 
diff --git a/doc/source/v0.13.0.txt b/doc/source/v0.13.0.txt
index 536b0badeec00..f0a23b46373e9 100644
--- a/doc/source/v0.13.0.txt
+++ b/doc/source/v0.13.0.txt
@@ -208,6 +208,21 @@ Enhancements
 
   - ``timedelta64[ns]`` operations
 
+    - Using the new top-level ``to_timedelta``, you can convert a scalar or array from the standard
+      timedelta format (produced by ``to_csv``) into a timedelta type (``np.timedelta64`` in ``nanoseconds``).
+
+      .. warning::
+
+         This requires ``numpy >= 1.7``
+
+      .. ipython:: python
+
+         to_timedelta('1 days 06:05:01.00003')
+         to_timedelta('15.5us')
+         to_timedelta(['1 days 06:05:01.00003','15.5us','nan'])
+         to_timedelta(np.arange(5),unit='s')
+         to_timedelta(np.arange(5),unit='d')
+
     - A Series of dtype ``timedelta64[ns]`` can now be divided by another
       ``timedelta64[ns]`` object to yield a ``float64`` dtyped Series. This
       is frequency conversion. See :ref:`here<timeseries.timedeltas_convert>` for the docs.
diff --git a/pandas/tseries/api.py b/pandas/tseries/api.py
index ead5a17c4fab1..c2cc3723802fc 100644
--- a/pandas/tseries/api.py
+++ b/pandas/tseries/api.py
@@ -7,5 +7,6 @@
 from pandas.tseries.frequencies import infer_freq
 from pandas.tseries.period import Period, PeriodIndex, period_range, pnow
 from pandas.tseries.resample import TimeGrouper
+from pandas.tseries.timedeltas import to_timedelta
 from pandas.lib import NaT
 import pandas.tseries.offsets as offsets
diff --git a/pandas/tseries/tests/test_timedeltas.py b/pandas/tseries/tests/test_timedeltas.py
index 8b88da1d7ef28..551507039112b 100644
--- a/pandas/tseries/tests/test_timedeltas.py
+++ b/pandas/tseries/tests/test_timedeltas.py
@@ -11,7 +11,7 @@
                     bdate_range, date_range, _np_version_under1p7)
 import pandas.core.common as com
 from pandas.compat import StringIO, lrange, range, zip, u, OrderedDict, long
-from pandas import compat
+from pandas import compat, to_timedelta, tslib
 from pandas.tseries.timedeltas import _coerce_scalar_to_timedelta_type as ct
 from pandas.util.testing import (assert_series_equal,
                                  assert_frame_equal,
@@ -33,10 +33,9 @@ def setUp(self):
     def test_numeric_conversions(self):
         _skip_if_numpy_not_friendly()
 
-        # ns not converted properly
         self.assert_(ct(0) == np.timedelta64(0,'ns'))
-        self.assert_(ct(10) == np.timedelta64(0,'ns'))
-        self.assert_(ct(10,unit='ns') == np.timedelta64(0,'ns').astype('m8[ns]'))
+        self.assert_(ct(10) == np.timedelta64(10,'ns'))
+        self.assert_(ct(10,unit='ns') == np.timedelta64(10,'ns').astype('m8[ns]'))
 
         self.assert_(ct(10,unit='us') == np.timedelta64(10,'us').astype('m8[ns]'))
         self.assert_(ct(10,unit='ms') == np.timedelta64(10,'ms').astype('m8[ns]'))
@@ -56,11 +55,10 @@ def test_short_format_converters(self):
         def conv(v):
             return v.astype('m8[ns]')
 
-        # ns not converted properly
-        self.assert_(ct('10') == np.timedelta64(0,'ns'))
-        self.assert_(ct('10ns') == np.timedelta64(0,'ns'))
-        self.assert_(ct('100') == np.timedelta64(0,'ns'))
-        self.assert_(ct('100ns') == np.timedelta64(0,'ns'))
+        self.assert_(ct('10') == np.timedelta64(10,'ns'))
+        self.assert_(ct('10ns') == np.timedelta64(10,'ns'))
+        self.assert_(ct('100') == np.timedelta64(100,'ns'))
+        self.assert_(ct('100ns') == np.timedelta64(100,'ns'))
 
         self.assert_(ct('1000') == np.timedelta64(1000,'ns'))
         self.assert_(ct('1000ns') == np.timedelta64(1000,'ns'))
@@ -122,6 +120,48 @@ def conv(v):
         # invalid
         self.assertRaises(ValueError, ct, '- 1days, 00')
 
+    def test_nat_converters(self):
+        _skip_if_numpy_not_friendly()
+
+        self.assert_(to_timedelta('nat') == tslib.iNaT)
+        self.assert_(to_timedelta('nan') == tslib.iNaT)
+
+    def test_to_timedelta(self):
+        _skip_if_numpy_not_friendly()
+
+        def conv(v):
+            return v.astype('m8[ns]')
+        d1 = np.timedelta64(1,'D')
+
+        self.assert_(to_timedelta('1 days 06:05:01.00003') == conv(d1+np.timedelta64(6*3600+5*60+1,'s')+np.timedelta64(30,'us')))
+        self.assert_(to_timedelta('15.5us') == conv(np.timedelta64(15500,'ns')))
+
+        # empty string
+        result = to_timedelta('')
+        self.assert_(result == tslib.iNaT)
+
+        result = to_timedelta(['', ''])
+        self.assert_(isnull(result).all())
+
+        # pass thru
+        result = to_timedelta(np.array([np.timedelta64(1,'s')]))
+        expected = np.array([np.timedelta64(1,'s')])
+        tm.assert_almost_equal(result,expected)
+
+        # ints
+        result = np.timedelta64(0,'ns')
+        expected = to_timedelta(0)
+        self.assert_(result == expected)
+
+        # Series
+        expected = Series([timedelta(days=1), timedelta(days=1, seconds=1)])
+        result = to_timedelta(Series(['1d','1days 00:00:01']))
+        tm.assert_series_equal(result, expected)
+
+        # with units
+        result = Series([ np.timedelta64(0,'ns'), np.timedelta64(10,'s').astype('m8[ns]') ],dtype='m8[ns]')
+        expected = to_timedelta([0,10],unit='s')
+        tm.assert_series_equal(result, expected)
 
 if __name__ == '__main__':
     nose.runmodule(argv=[__file__, '-vvs', '-x', '--pdb', '--pdb-failure'],
diff --git a/pandas/tseries/timedeltas.py b/pandas/tseries/timedeltas.py
index d12791755859e..4d8633546e017 100644
--- a/pandas/tseries/timedeltas.py
+++ b/pandas/tseries/timedeltas.py
@@ -8,29 +8,86 @@
 import numpy as np
 import pandas.tslib as tslib
 from pandas import compat, _np_version_under1p7
-from pandas.core.common import (ABCSeries, is_integer,
+from pandas.core.common import (ABCSeries, is_integer, is_timedelta64_dtype,
                                 _values_from_object, is_list_like)
 
 repr_timedelta = tslib.repr_timedelta64
 repr_timedelta64 = tslib.repr_timedelta64
 
+def to_timedelta(arg, box=True, unit='ns'):
+    """
+    Convert argument to timedelta
+
+    Parameters
+    ----------
+    arg : string, timedelta, array of strings (with possible NAs)
+    box : boolean, default True
+        If True returns a Series of the results, if False returns ndarray of values
+    unit : unit of the arg (D,s,ms,us,ns) denote the unit, which is an integer/float number
+
+    Returns
+    -------
+    ret : timedelta64/arrays of timedelta64 if parsing succeeded
+    """
+    if _np_version_under1p7:
+        raise ValueError("to_timedelta is not support for numpy < 1.7")
+
+    def _convert_listlike(arg, box):
+
+        if isinstance(arg, (list,tuple)):
+            arg = np.array(arg, dtype='O')
+
+        if is_timedelta64_dtype(arg):
+            if box:
+                from pandas import Series
+                return Series(arg,dtype='m8[ns]')
+            return arg
+
+        value = np.array([ _coerce_scalar_to_timedelta_type(r, unit=unit) for r in arg ])
+        if box:
+            from pandas import Series
+            value = Series(value,dtype='m8[ns]')
+        return value
+
+    if arg is None:
+        return arg
+    elif isinstance(arg, ABCSeries):
+        from pandas import Series
+        values = _convert_listlike(arg.values, box=False)
+        return Series(values, index=arg.index, name=arg.name, dtype='m8[ns]')
+    elif is_list_like(arg):
+        return _convert_listlike(arg, box=box)
+
+    return _convert_listlike([ arg ], box=False)[0]
+
 _short_search = re.compile(
     "^\s*(?P<neg>-?)\s*(?P<value>\d*\.?\d*)\s*(?P<unit>d|s|ms|us|ns)?\s*$",re.IGNORECASE)
 _full_search = re.compile(
     "^\s*(?P<neg>-?)\s*(?P<days>\d+)?\s*(days|d)?,?\s*(?P<time>\d{2}:\d{2}:\d{2})?(?P<frac>\.\d+)?\s*$",re.IGNORECASE)
+_nat_search = re.compile(
+    "^\s*(nat|nan)\s*$",re.IGNORECASE)
+_whitespace = re.compile('^\s*$')
 
 def _coerce_scalar_to_timedelta_type(r, unit='ns'):
     # kludgy here until we have a timedelta scalar
     # handle the numpy < 1.7 case
 
+    def conv(v):
+        if _np_version_under1p7:
+            return timedelta(microseconds=v/1000.0)
+        return np.timedelta64(v)
+
     if isinstance(r, compat.string_types):
         converter = _get_string_converter(r, unit=unit)
         r = converter()
-        r = timedelta(microseconds=r/1000.0)
-
-    if is_integer(r):
+        r = conv(r)
+    elif r == tslib.iNaT:
+        return r
+    elif isinstance(r, np.timedelta64):
+        r = r.astype("m8[{0}]".format(unit.lower()))
+    elif is_integer(r):
         r = tslib.cast_from_unit(r, unit)
-        r = timedelta(microseconds=r/1000.0)
+        r = conv(r)
 
     if _np_version_under1p7:
         if not isinstance(r, timedelta):
@@ -50,6 +107,12 @@ def _coerce_scalar_to_timedelta_type(r, unit='ns'):
 def _get_string_converter(r, unit='ns'):
     """ return a string converter for r to process the timedelta format """
 
+    # treat as a nan
+    if _whitespace.search(r):
+        def convert(r=None, unit=None):
+            return tslib.iNaT
+        return convert
+
     m = _short_search.search(r)
     if m:
         def convert(r=None, unit=unit, m=m):
@@ -92,6 +155,12 @@ def convert(r=None, unit=None, m=m):
             return tslib.cast_from_unit(value, 's')
         return convert
 
+    m = _nat_search.search(r)
+    if m:
+        def convert(r=None, unit=None, m=m):
+            return tslib.iNaT
+        return convert
+
     # no converter
     raise ValueError("cannot create timedelta string converter")
 
diff --git a/pandas/tslib.pyx b/pandas/tslib.pyx
index 8117ff7d3539b..fd97512b0528b 100644
--- a/pandas/tslib.pyx
+++ b/pandas/tslib.pyx
@@ -1303,7 +1303,9 @@ cpdef inline int64_t cast_from_unit(object ts, object unit) except -1:
     # to avoid precision issues from float -> int
     base = <int64_t> ts
     frac = ts-base
-    return <int64_t> (base*m) + <int64_t> (round(frac,p)*m)
+    if p:
+       frac = round(frac,p)
+    return <int64_t> (base*m) + <int64_t> (frac*m)
 
 def cast_to_nanoseconds(ndarray arr):
     cdef: