BUG/TST: Fix for #6724, make numpy.ma.mvoid consistent with numpy.void - #6763
Conversation
…y.void Make indexing on numpy.ma.mvoid consistent with indexing on numpy.void. Changes behaviour in rare cases (see below). Fixes numpy#6724. Sometimes, indexing ma.mvoid results in a non-scalar mask. For example, dimension increases if indexing with a multi-dimensional field. Previously, this led to a ValueError (truth value ambiguous). With this commit, indexing now returns an ma.masked_array so that there is no loss of information. Note that there is a precedence for returning from void to array. Z = zeros((2,), dtype="(2,)i2,(2,)i2"), then Z[0] is a void, but Z[0][0] and Z[0]["f1"] are array. This commit therefore implements behaviouk such that numpy.ma.mvoid is consistent with numpy.void. Also adds a related test. The behaviour changes in cases where for a masked array `X`, X.dtype["A"] is multidimensional but size 1, such as in the example below. Any case where X.dtype["A"] is multidimensional but with size>1 would previously fail. Old behaviour: In [15]: X = ma.masked_array(data=[([0],)], mask=[([False],)], dtype=[("A", "(1,1)i2", (1,1))]) In [16]: X[0]["A"] Out[16]: array([[[[0]]]], dtype=int16) In [17]: X = ma.masked_array(data=[([0],)], mask=[([True],)], dtype=[("A", "(1,1)i2", (1,1))]) In [18]: X[0]["A"] Out[18]: masked New behaviour: In [1]: X = ma.masked_array(data=[([0],)], mask=[([False],)], dtype=[("A", "(1,1)i2", (1,1))]) In [2]: X[0]["A"] Out[2]: masked_array(data = [[[[0]]]], mask = [[[[False]]]], fill_value = [[[[16959]]]]) In [3]: X = ma.masked_array(data=[([0],)], mask=[([True],)], dtype=[("A", "(1,1)i2", (1,1))]) In [4]: X[0]["A"] Out[4]: masked_array(data = [[[[--]]]], mask = [[[[ True]]]], fill_value = [[[[16959]]]]) The new behaviour is more consistent with indexing the data themselves: In [7]: X.data[0]["A"] Out[7]: array([[[[0]]]], dtype=int16) In theory, this change in behaviour can break code, but I would consider it very unlikely.
|
@gerritholl @ahaldane Do you have a preference? This looks a bit more direct than the other fix. |
|
I prefer this one, as I think it's more correct. Arguably, the old behaviour for the one case where the old behaviour didn't crash is buggy. And from the observation that I discover (and sometimes fix) those bugs in exotic structured masked arrays, there don't seem to be very many people using them. |
|
OK, I'm going to close the other. It can be reopened if needed. |
There was a problem hiding this comment.
I think this line is good now.
But for any future reference, another possibility which you reminded me of in another PR is if m.dtype[indx].subdtype is not None, which checks specifically for subarrays. But I think a check for ndarray is better because there might be other ways of getting an ndarray when indexing - for instance, multifield indices like marr[['f0', 'f1']] could return an ndarray, which we would want to convert to a masked array. (Currently void scalars to not support multifield indices though, but I think that is an oversight which we might want to fix soon, eg in #6053)
|
OK, let's give this a shot. @ahaldane Are you good with this? |
|
Looks good to me, thanks @gerritholl . I'll let @charris do the merge. |
…rray_mvoid_alt BUG/TST: Fix for #6724, make numpy.ma.mvoid consistent with numpy.void
3 participants
This is an alternative fix for #6724. The other fix is in #6755.
Make indexing on numpy.ma.mvoid consistent with indexing on numpy.void.
Changes behaviour in rare cases (see below). Fixes #6724. Sometimes,
indexing ma.mvoid results in a non-scalar mask. For example, dimension
increases if indexing with a multi-dimensional field. Previously, this
led to a ValueError (truth value ambiguous). With this commit, indexing
now returns an ma.masked_array so that there is no loss of information.
Note that there is a precedence for returning from void to array.
Z = zeros((2,), dtype="(2,)i2,(2,)i2"), then Z[0] is a void, butZ[0][0] and Z[0]["f1"]are array. This commit therefore implements behavioursuch that
numpy.ma.mvoidis consistent withnumpy.void.Also adds a related test.
The behaviour changes in cases where for a masked array
X,X.dtype["A"]is multidimensional but size 1, such as in the example below. Any case
where
X.dtype["A"]is multidimensional but with size>1 would previouslyfail.
Old behaviour:
New behaviour:
The new behaviour is more consistent with indexing the data themselves:
In theory, this change in behaviour can break code, but I would consider
it very unlikely.