timedelta error with numpy.longdouble dtype

Question

I have times with dtype numpy.longdouble and when I'm trying to use that values with timedelta function I've got errors. But when I convert it to numpy.float64 everything is fine. Could somebody explain that behaviour?

import numpy as np
from datetime import timedelta
t1 = np.array([1000], dtype=np.longdouble)
t2 = np.array([1000], dtype=np.float64)

In [166]: timedelta(seconds=t1[0])
TypeError: unsupported type for timedelta seconds component: numpy.float64

In [167]: timedelta(seconds=t2[0])
Out[167]: datetime.timedelta(0, 1000)

In [168]: timedelta(seconds=t1[0].astype(np.float64))
Out[168]: datetime.timedelta(0, 1000)

When I'm trying to see dtypes of variables they are look the similar but not the same:

In [172]: t1[0].dtype
Out[172]: dtype('float64')

In [173]: t2[0].dtype
Out[173]: dtype('float64')

In [174]: np.longdouble == np.float64
Out[174]: False

In [175]: t1[0].dtype == t2[0].dtype
Out[175]: True

Edit

And it's strange that it's not working for np.int32 and np.int64 either:

t3 = np.array([1000], dtype=np.int32)
t4 = np.array([1000], dtype=np.int64)

In [29]: timedelta(t3[0])
TypeError: unsupported type for timedelta days component: numpy.int32

In [69]: timedelta(t4[0])
TypeError: unsupported type for timedelta days component: numpy.int64

Answer 1

I ran into a similar issue where

timedelta(days = value.astype(int)) did not work but

timedelta(days = int(value)) worked.

I hope it is useful for future Googlers.

Answer 2

So maybe timedelta for dtype np.longdouble isn't implemented?

In short, yes.

From the documentation:

class datetime.timedelta([days[, seconds[, microseconds[, milliseconds[, minutes[, hours[, weeks]]]]]]])

All arguments are optional and default to 0. Arguments may be ints, longs, or floats, and may be positive or negative.

Here "long" refers to a Python long integer, not a longdouble float.

---

Update

I think I've figured the seemingly inconsistent behavior of np.float64. What seems to be relevant is whether or not the numpy dtype subclasses one of the native Python scalar types that timedelta accepts.

On my 64 bit machine, running Python 2.7.9, numpy v1.10.1:

In [1]: timedelta(np.float64(1))
Out[1]: datetime.timedelta(1)

In [2]: timedelta(np.float32(1))
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-2-4a7874ba393b> in <module>()
----> 1 timedelta(np.float32(1))

TypeError: unsupported type for timedelta days component: numpy.float32

In [3]: timedelta(np.int64(1))
Out[3]: datetime.timedelta(1)

In [4]: timedelta(np.int32(1))
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-4-0475c6c8f1aa> in <module>()
----> 1 timedelta(np.int32(1))

TypeError: unsupported type for timedelta days component: numpy.int32

In [5]: issubclass(np.float64, float)
Out[5]: True

In [6]: issubclass(np.float32, float)
Out[6]: False

In [7]: issubclass(np.int64, int)
Out[7]: True

In [8]: issubclass(np.int32, int)
Out[8]: False

However the OP reported in the comments that timedelta(np.int64(1)) did not work using Python 3.4.3. I think that is due to the fact that when numpy is built on Python 3x, np.int64 no longer subclasses int.

Here's what happens in Python 3.4.3:

In [1]: timedelta(np.int64(1))
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-1-9902ea26a52d> in <module>()
----> 1 timedelta(np.int64(1))

TypeError: unsupported type for timedelta days component: numpy.int64

In [2]: issubclass(np.int64, int)
Out[2]: False