AFAICT, neither is currently allowed.

In [6]: np.zeros(2, 'p,p').view('p,p,p')
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-6-02190acaad61> in <module>()
----> 1 np.zeros(2, 'p,p').view('p,p,p')

ValueError: new type not compatible with array.

In [7]: np.zeros(2, 'p,p,p').view('p,p')
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-7-8075bf4648ec> in <module>()
----> 1 np.zeros(2, 'p,p,p').view('p,p')

ValueError: new type not compatible with array.

And the relevant code hasn't changed since 2011 (1.7 release I suspect). I don't see that either should be acceptable.

EDIT: OK, I see what the difference is. But what if the array is not contiguous?

Looks like contiguity is checked somewhere

In [6]: np.zeros(2, 'p,p').view('p,p,p')
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-6-02190acaad61> in <module>()
----> 1 np.zeros(2, 'p,p').view('p,p,p')

ValueError: new type not compatible with array.

In [7]: np.zeros(2, 'p,p,p').view('p,p')
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-7-8075bf4648ec> in <module>()
----> 1 np.zeros(2, 'p,p,p').view('p,p')

ValueError: new type not compatible with array.

I need to think about this a bit. It would seem that if the first is OK then the second should be OK also, but I'd like to check under what circumstances it is allowed. It seems a bit hinky in that it depends not only on array shape and element size, but also the strides.

I think contiguity is accounted for a couple lines above the lines I changed (through a check for PyArray_ISONESEGMENT). I'm reasonably convinced that code is right, though the NPY_C_CONTIGUOUS and NPY_F_CONTIGUOUS flags have confused me before.

I admit this is a pretty odd case which is unlikely to be used. But the first example above (which works currently) came up in #5508 and #5548, and seemed important enough that a bunch of tests use it (see a few lines above the new test I added here). Somehow I had never noticed until now that while an increasing itemsize worked a decreasing one didn't. I happened on it just now while playing with structured arrays for another PR.

I think the 3 versus 2 in OP's original calls to zeros may be crucial?
On Jul 20, 2015 14:01, "Charles Harris" notifications@github.com wrote:

Looks like contiguity is checked somewhere

In [6]: np.zeros(2, 'p,p').view('p,p,p')

ValueError Traceback (most recent call last)
in ()
----> 1 np.zeros(2, 'p,p').view('p,p,p')

ValueError: new type not compatible with array.

In [7]: np.zeros(2, 'p,p,p').view('p,p')

ValueError Traceback (most recent call last)
in ()
----> 1 np.zeros(2, 'p,p,p').view('p,p')

ValueError: new type not compatible with array.

I need to think about this a bit. It would seem that if the first is OK
then the second should be OK also, but I'd like to check under what
circumstances it is allowed. It seems a bit hinky in that it depends not
only on array shape and element size, but also the strides.

—
Reply to this email directly or view it on GitHub
#6097 (comment).

@njsmith Yes, I noted that in the edit.

Hmm after looking over #5508 again, I just realized that this PR is simply a more limited form of that PR - the changes here are almost word-for-word implemeted there, plus more. Maybe I will go over there and look at that code :)

Do you mind if I close this one in favor of that one?

@ahaldane Not at all ;) One less PR to review.