[Numpy-svn] [numpy/numpy] 53facf: BUG/TST: Fix for #6724, make numpy.ma.mvoid consis...

[GitHub]

  Branch: refs/heads/master
  Home:   https://github.com/numpy/numpy
  Commit: 53facf327c7b2949a5f31c308fd66b0b1f24b615
      https://github.com/numpy/numpy/commit/53facf327c7b2949a5f31c308fd66b0b1f24b615
  Author: Gerrit Holl <g.holl at reading.ac.uk>
  Date:   2015-12-03 (Thu, 03 Dec 2015)

  Changed paths:
    M numpy/ma/core.py
    M numpy/ma/tests/test_core.py

  Log Message:
  -----------
  BUG/TST: Fix for #6724, make numpy.ma.mvoid consistent with numpy.void

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, 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.


  Commit: 3cc797e5107d5f3fe812b0ccdc9bc96f107f0b20
      https://github.com/numpy/numpy/commit/3cc797e5107d5f3fe812b0ccdc9bc96f107f0b20
  Author: Charles Harris <charlesr.harris at gmail.com>
  Date:   2015-12-04 (Fri, 04 Dec 2015)

  Changed paths:
    M numpy/ma/core.py
    M numpy/ma/tests/test_core.py

  Log Message:
  -----------
  Merge pull request #6763 from gerritholl/structured_multidim_masked_array_mvoid_alt

BUG/TST: Fix for #6724, make numpy.ma.mvoid consistent with numpy.void


Compare: https://github.com/numpy/numpy/compare/c890b10a8ade...3cc797e5107d