NumPy dtype forcing structured array expansion over type

Question

I use special np.dtypes for a program that have a structure like so:

POINT = np.dtype([('vertices', '<f4', 2), ('center', '<f4', 2), ('bbox', '<f4', 4)])

I need to specify another np.dtype that only uses the last field above, like so:

MBR = np.dtype([('bbox', '<f4', 4)])

This is so I can later access that field of both arrays like so:

def intersection(s, t):

    sxmin, sxmax, symin, sxmax = s['bbox']
    txmin, txmax, tymin, tymax = t['bbox']

    # do stuff

However, when I create the following array, it is being expanded and I'm unsure why:

box = np.array([1, 2, 3, 4], dtype=MBR)
# expected output...
array([1., 2., 3., 4.], dtype=[('bbox', '<f4', 4)])
# actual output...
array([([1., 1., 1., 1.],), ..., ([4., 4., 4., 4.],)], dtype=[('bbox', '<f4', 4)])

A quick test returns what I had expected...

np.empty([], dtype=MBR)
array(([nan, nan, inf, nan],), dtype=[('bbox', '<f4', 4)])

Edit:

Doing the following returns the expected result:

box = np.array(([1, 2, 3, 4],), dtype=MBR)

So now the question: Why do I have to wrap it in a tuple to conform to the dtype?

hpaulj · Accepted Answer · 2020-01-28 17:39:27Z

The short answer is that the input format, with nested lists and tuples, has to match the display format:

In [59]: MBR = np.dtype([('bbox', '<f4', 4)])                                                    
In [60]: arr = np.zeros(3, dtype=MBR)                                                            
In [61]: arr                                                                                     
Out[61]: 
array([([0., 0., 0., 0.],), ([0., 0., 0., 0.],), ([0., 0., 0., 0.],)],
      dtype=[('bbox', '<f4', (4,))])
In [62]: arr[0]                                                                                  
Out[62]: ([0., 0., 0., 0.],)
In [63]: arr[0]=[1,2,3,4]                                                                        
In [64]: arr[1]=[10,11,12,13]                                                                    
In [65]: arr                                                                                     
Out[65]: 
array([([ 1.,  2.,  3.,  4.],), ([10., 11., 12., 13.],),
       ([ 0.,  0.,  0.,  0.],)], dtype=[('bbox', '<f4', (4,))])
In [66]: np.array([([1,2,3,4],)],MBR)                                                            
Out[66]: array([([1., 2., 3., 4.],)], dtype=[('bbox', '<f4', (4,))])

So with a typical compound dtype, we say the inputs should be a list of tuples, one tuple per 'record' of the array. Within the tuple, one element per field.

You have the added complication of the size (4,) dimension within the field.

Note that the shape of the field, extracted from the array, combines the outer array shape with the inner field shape:

In [67]: arr['bbox']                                                                             
Out[67]: 
array([[ 1.,  2.,  3.,  4.],
       [10., 11., 12., 13.],
       [ 0.,  0.,  0.,  0.]], dtype=float32)

Often it is easier to assign values to a structured array by field, rather than by record:

In [68]: arr['bbox']=np.arange(12).reshape(3,4)                                                  
In [69]: arr                                                                                     
Out[69]: 
array([([ 0.,  1.,  2.,  3.],), ([ 4.,  5.,  6.,  7.],),
       ([ 8.,  9., 10., 11.],)], dtype=[('bbox', '<f4', (4,))])

Collectives™ on Stack Overflow

NumPy dtype forcing structured array expansion over type

1 Answer 1

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Related