[Numpy-discussion] Less dimensions than expected with record array

[josef.pktd at gmail.com]

On Fri, Apr 29, 2011 at 10:56 PM, Alan Gibson <dyssident at gmail.com> wrote:
> Hello all,
>
> This question may seem elementary (mostly because it is), but I can't
> find documentation anywhere as to why the following are true:
>
>>>> import numpy as np
>>>> data = [(1,2,3),(4,5,6),(7,8,9)]
>>>> dt = [('a',int),('b',int),('c',int)]
>>>> normal_array = np.array(data)
>>>> record_array = np.array(data, dtype=dt)
>>>> print "ndarray has shape %s but record array has shape %s" % \
> ...     (normal_array.shape, record_array.shape)
> ndarray has shape (3, 3) but record array has shape (3,)
>>>> print "ndarray has %s dimensions but record array has %s dimensions" % \
> ...     (normal_array.ndim, record_array.ndim)
> ndarray has 2 dimensions but record array has 1 dimensions
>
> This makes seemingly reasonable things, like using apply_along_axis()
> over a table of data with named columns, impossible:
>
>>>> np.apply_along_axis(record_array, 1, lambda x: x)
> Traceback (most recent call last):
>  File "<stdin>", line 1, in <module>
>  File "/usr/local/lib/python2.6/dist-packages/numpy/lib/shape_base.py",
> line 72, in apply_along_axis
>   % (axis,nd))
> ValueError: axis must be less than arr.ndim; axis=1, rank=0.
>
> What's the reason for this behavior? Is there a way to make such
> operations work with record arrays?

each row (record) is treated as one array element, so the structured
array is only 1d.

If you have rows/records with content that is not homogenous, then
working along axis=1 (across elements of a record) doesn't make sense.
for example I just struggle with 2 datetime columns and the rest are
integers.

If you want an array with homogenous elements (all floats or all ints)
with operations along axis, then larry (la) is, I think, still the
best bet. I don't know what the status with the dataarray for numpy
is.

Josef

>
> Thanks,
>
> Alan
> _______________________________________________
> NumPy-Discussion mailing list
> NumPy-Discussion at scipy.org
> http://mail.scipy.org/mailman/listinfo/numpy-discussion
>