Repository navigation
Questionable behaviour in MaskedArray.__eq__ on flexible arrays #8589
Description
Activity
In the commit pointed too, in another similar change, there is a comment: "# In case we have nested fields..."
So, at least that gives a hint why this was introduced. Indeed, one of the test cases shows that it is at least one thing the
ValueErroris meant to catch:ndtype = [('A', int), ('B', [('BA', int), ('BB', int)])] a = np.ma.array([(1, (1, 1)), (2, (2, 2))], mask=[(0, (1, 0)), (0, (0, 1))], dtype=ndtype) a.mask # array([(False, ( True, False)), (False, (False, True))], # dtype=[('A', '?'), ('B', [('BA', '?'), ('BB', '?')])]) a.mask.view((np.bool_, len(a.dtype))) # ValueError: new type not compatible with array.In principle, I guess one should really write
try: mask = mask.view((bool_, len(self.dtype))) except ValueError: mask = np.all(....) else: mask = mask.all(axis)But where does the indexing go wrong? I must say having a default of
axis=1seems very odd; surely, this should beaxis=-1?!?!Does this mean there is a bug with masked 2D record arrays?
Does this mean there is a bug with masked 2D record arrays?
Yes. What a code... Anyway, fix almost done.
Yes!
These lines, introduced in a5da87c by @pierregm, appear multiple times in
numpy/ma/core.py, such asMaskedArray.__eq__:But where is this
ValueErrorexpected to come from? The.view, or the.all, due to an invalid axis.It seems like there is a bug in the code that prepares
axishere, and this bug is being silenced by an overzealous try/catch.