BUG: structured array indexing dtype and dtype.descr changed in 1.14.0 #10387
Description
Running this code on 1.13.x and 1.14.0 yields different outputs:
import numpy as np
dt = [('a', int), ('b', float)]
data = np.zeros(1, dtype=dt)
data['a'] = 0
data['b'] = 1
data = data[['a']]
print(data.dtype)
print(data.dtype.descr)
np.save('test.npy', data)
data_read = np.load('test.npy')
print(data_read.dtype)
print(data_read.dtype.descr)
On 1.13.x the output is consistent -- for all four print statements it gives:
[('a', '<i8')]
On 1.14.x it changes to these, the first of which is now dict-like, the second (descr) has a new field ('', '|V8'), the third (round-trip IO) changes dtype back to a list, and the fourth (descr) changes the blank field name to '':
{'names':['a'], 'formats':['<i8'], 'offsets':[0], 'itemsize':16}
[('a', '<i8'), ('', '|V8')]
[('a', '<i8'), ('f1', 'V8')]
[('a', '<i8'), ('f1', '|V8')]
The real reason I found this bug is because the addition of the blank field name breaks scipy.io.savemat:
>>> from scipy import io # on latest dev
>>> io.savemat('test.mat', dict(data=data))
File "untitled0.py", line 16, in <module>
io.savemat('test.mat', dict(data=data))
File "scipy/io/matlab/mio.py", line 219, in savemat
MW.put_variables(mdict)
File "scipy/io/matlab/mio5.py", line 846, in put_variables
self._matrix_writer.write_top(var, asbytes(name), is_global)
File "scipy/io/matlab/mio5.py", line 587, in write_top
self.write(arr)
File "scipy/io/matlab/mio5.py", line 616, in write
self.write_struct(narr)
File "scipy/io/matlab/mio5.py", line 735, in write_struct
self._write_items(arr)
File "scipy/io/matlab/mio5.py", line 752, in _write_items
self.write(el[f])
ValueError: no field of name
If these structured dtype changes are intentional and not bugs, I can open a SciPy issue instead (which would either be to avoid writing void data, or write out unknown fields).