Array math on numpy structured arrays

Question

import numpy as np
arr = np.array([(1,2), (3,4)], dtype=[('c1', float), ('c2', float)])
arr += 3

results in an invalid type promotion error. Is there a way I can have nice labeled columns like a structured array, but still be able to do operations like it's a simple dtype=float array?

Alternatively, is there an easy way to cast a dtype=float array into a structured array? i.e.

arr = np.array([(1,2), (3,4)], dtype=float)
arr_struc = arr.astype([('c1', float), ('c2', float)])

only where it doesn't broadcast and matches columns to names. Seems like I shouldn't have to do this loop:

arr_struc = np.zeros(2, dtype=[('c1', float), ('c2', float)])
for i,key in enumerate(arr_struc.dtype.names):  arr_struc[key] = arr[i,:]

On the structured array docs page, they describe the numpy.lib.recfunctions module. In has a structured_to_unstructured and unstructured_to_structured functions. A view if it works is faster, but these are more general purpose. — hpaulj
– hpaulj, Commented Jun 24, 2020 at 3:07
A key thing to keep in mind when doing your own conversion is that data for a structured array is in the form of a list of tuples, while unstructured displays as list of lists. But using structured arrays just to have pretty 'column' names is, I think, more work than its worth. Structured arrays are best when fields have distinct dtypes, esp. some numeric, and others string. — hpaulj
– hpaulj, Commented Jun 24, 2020 at 3:09

juanpa.arrivillaga · Accepted Answer · 2020-06-24 03:21:49Z

2

Hmmm. One option, use a view for this:

>>> import numpy as np
>>> arr = np.array([(1,2), (3,4)], dtype=[('c1', float), ('c2', float)])
>>> view = arr.view(float)
>>> view += 3
>>> arr
array([(4., 5.), (6., 7.)], dtype=[('c1', '<f8'), ('c2', '<f8')])
>>> view
array([4., 5., 6., 7.])

Not the cleanest. But it's a solution.

EDIT:

Yes, don't use astype use a view again:

>>> arr = np.array([(1,2), (3,4)], dtype=float)
>>> arr
array([[1., 2.],
       [3., 4.]])
>>> struct = arr.view(dtype=[('c1', float), ('c2', float)])
>>> struct
array([[(1., 2.)],
       [(3., 4.)]], dtype=[('c1', '<f8'), ('c2', '<f8')])
>>> struct.shape
(2, 1)

You may have to reshape it to your liking:

>>> struct.squeeze()
array([(1., 2.), (3., 4.)], dtype=[('c1', '<f8'), ('c2', '<f8')])

edited Jun 24, 2020 at 3:21

answered Jun 23, 2020 at 22:14

juanpa.arrivillaga

97.6k14 gold badges141 silver badges190 bronze badges

Sign up to request clarification or add additional context in comments.

2 Comments

I.P. Freeley Over a year ago

That's a good way. I guess for my purposes I could also do most of the math on the float array, then convert it at the end. Is there an elegant way to do that (i.e., without a loop)? arr.astype([('c1', '<f8'), ('c2', '<f8')]) doesn't behave like I'd want.

juanpa.arrivillaga Over a year ago

@I.P.Freeley I don't follow, can you edit your question and elaborate?

Collectives™ on Stack Overflow

Array math on numpy structured arrays

1 Answer 1

2 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

2 Comments

Your Answer

Sign up or log in

Post as a guest

Related