np.MaskedArray.transpose() produces partial views

[eric-wieser]

>>> src = np.ma.masked_where([0, 1, 0], [1, 1337, 1])
>>> dst = np.ma.masked_where([1, 0, 0, 0, 0], np.zeros(5))
>>> dst_view = dst.transpose()
# check the views look ok
>>> dst
masked_array(data = [-- 0 0 0 0],
             mask = [ True False False False False],
       fill_value = 999999)
>>> dst_view
masked_array(data = [-- 0 0 0 0],
             mask = [ True False False False False],
       fill_value = 999999)

# now write a slice of the view
>>> dst_view[2:5] = src
>>> dst_view
masked_array(data = [-- 0 1 -- 1],
             mask = [ True False False  True False],
       fill_value = 999999)

# but looking back at the original, the data was updated but the mask was not!
>>> dst
masked_array(data = [-- 0 1 1337 1],
             mask = [ True False False False False],
       fill_value = 999999)

[eric-wieser]

Repeating this with [:] instead of .tranpose() has the same effect, but also prints out

Warning (from warnings module):
  File "__main__", line 1
MaskedArrayFutureWarning: setting an item on a masked array which has a shared mask will not copy the mask and also change the original mask array in the future.
Check the NumPy 1.11 release notes for more information.

[mhvk]

This behaviour is as intended originally (see masked array docs), but it is not obvious it is a good idea (see #5558), and hence plans are afoot to change it: #5580. The latter PR is also the reason you are getting FutureWarning. One thing that might be useful for #5580 is to check that your examples would work with that PR (in particular for the transpose, for which you do not get the FutureWarning).

[seberg]

Its pretty nonsense __setmask__ has a copy argument, but it seems like it probably does not matter. I doubt the mask is ever used as a view, which it should for the methods at least....

Seems like all those changes were not enough yet and there are more things where such a warning/change would be needed. Though, at least it may be ok to do one after another.

[eric-wieser]

This behaviour is as intended originally

Yep, I think the best option here is to duplicate the warning inside transpose

[mhvk]

We tried quite hard not to give warnings that were superfluous -- this only affects those who would like to write to transposed data, which surely is a minority. So, I don't think we can warn inside transpose...

[eric-wieser]

Perhaps there should be a flag on "partial view"s, that causes a warning to be emitted the first time a write happens

[seberg]

I am not sure, but it might be possible to set _sharedmask to true. However, it may not actually warn in all cases that will change, but only for assignments.