Given an array like this:

arr = np.array([(1, 2), (3, 4)], dtype=np.dtype([('a','u2'), ('b','u4')], align=True))

The dtype is:

{'names':['a','b'], 'formats':['<u2','<u4'], 'offsets':[0,4], 'itemsize':8, 'aligned':True}

Note the difference with the default align=False dtype:

[('a', '<u2'), ('b', '<u4')]

Now, the problem is that np.save('foo.npy', arr) saves the aligned dtype this way:

[('a', '<u2'), ('', '|V2'), ('b', '<u4')]

Then, np.load('foo.npy') produces:

array([(1, [62, -18], 2), (3, [62, -18], 4)], 
      dtype=[('a', '<u2'), ('f1', 'V2'), ('b', '<u4')])

This is despite the fact that np.load() can parse the original dtype with names, formats, offsets - only np.save() doesn't save it that way. Editing foo.npy by hand lets np.load() produce an identical array to arr (if you try it, the 9th byte must change from 0x66 to 0x96 to allow for the longer descr).

As far as I can tell, fixing this would be as simple as storing str(arr.dtype) instead of arr.dtype.descr in the NPY file. For aligned arrays the two seem to produce the same string. But I'm not sure if there's some reason this wasn't done in the first place.

I'm using NumPy 1.11.1.