BUG: np.ma.mean and var should return scalar if no mask #8142
Conversation
ahaldane
force-pushed
the
ma_mean_scalar
branch
from
f40633e
to
2f792b1
Compare
|
Hmm this led me to notice some other (old) problems with |
| cnt = self.count(axis=axis, **kwargs) - ddof | ||
| danom = self - self.mean(axis, dtype, keepdims=True) | ||
| if iscomplexobj(self): | ||
| danom = umath.absolute(danom) ** 2 |
There was a problem hiding this comment.
Good to check for a complex object, but might as well do (there was a PR to make a ufunc out of this...):
danom = danom.real**2 + danom.imag**2
There was a problem hiding this comment.
I'm going to leave this alone for now, as outside the scope of this PR. (this code is already present).
|
@ahaldane Still planning more for this PR? |
|
Yeah, I need a day or two more to decide what to do in I have an old branch lying around with a "proper" overhaul of |
ahaldane
force-pushed
the
ma_mean_scalar
branch
from
2f792b1
to
d8d7c25
Compare
|
I finished my changes here, it should be ready to go. |
| @@ -5057,7 +5057,7 @@ def mean(self, axis=None, dtype=None, out=None, keepdims=np._NoValue): | |||
|
|
|||
| if self._mask is nomask: | |||
| result = super(MaskedArray, self).mean(axis=axis, | |||
| dtype=dtype, **kwargs) | |||
| dtype=dtype, **kwargs)[()] | |||
| else: | |||
There was a problem hiding this comment.
I assume the super mean returns a 0-d array scalar for relevant cases. I note the following odd behavior of ordinary arrays that is new to me.
>>> np.float64(1)
1.0
>>> np.float64(1)[()]
array(1.0)
>>> np.float64(1)[()][()]
1.0
That is, the [()] construct flips the result back and forth between arrays and scalars.
There was a problem hiding this comment.
For anyone coming here in future, the above is no longer (1.11.1) the case. [()] does not promote scalars to 0d-arrays
|
I note some odd behavior of |
|
What numpy version? I don't get that on 1.11, and I think there is a lot of numpy code depends on |
|
Hmm, but actually maybe you are right. I just remembered #7267. And actually grepping shows the empty tumple indexing is only used in a few places in python, and I know of only one place in C it is used. |
|
I'm running on my travel machine that has numpy 1.8. I can't upgrade until I figure out how to get Cython installed. Yeah, my mac environment is a mess, I don't use it often enough... |
|
I'm not particularly worried as long as the tests cover the relevant possibilities. |
|
If you can't replicate what I'm seeing on 1.8 let's give it a shot. Thanks Allan. |
|
Thanks Chuck! |
As discussed in my comments for issue numpy#8145, this patch adds the equal_nan argument to assert_array_compare(), and assert_allclose() passes the value it receives for the same argument through to assert_array_compare(). Closes numpy#8142.
Fixes #5769
This is a followup to #7350 (first added in 1.12), and makes sure
np.ma.meanandnp.ma.varreturn a scalar if appropriate in thenomaskbranch of the code.