Wrong dtype for unicode field in np.rec.fromarrays()

[taldcroft]

In Python 2 or 3 with Numpy 1.7.1 there seems to be problem with np.rec.fromarrays creating a dtype format that is a factor of 4 too large:

In [5]: a = np.array([u'xyz'])

In [6]: a.dtype
Out[6]: dtype('<U3')

In [7]: a2 = np.rec.fromarrays([a], names=['a'])

In [8]: a2.dtype
Out[8]: dtype([('a', '<U12')])

In [9]: a3 = np.rec.fromarrays([a2['a']], names=['a'])

In [10]: a3.dtype
Out[10]: dtype([('a', '<U48')])

It looks like the problem is here, where here itemsize is 12 for a 3-character unicode string with UCS-4 encoding:

            if issubclass(obj.dtype.type, nt.flexible):
                formats += repr(obj.itemsize)

[charris]

Hah. Numpy unicode is 4 bytes wide, I'll bet that is where the factor of 4 comes from.

[charris]

I'm guessing this is an easy fix. If not, it will be a pretty hard fix.

[taldcroft]

😄

[embray]

Ah, somehow missed this one before, but this was fixed by #5251. This issue can be closed.

[eric-wieser]

Thanks @embray for noting that