ENH: stats.alexandergovern: vectorize calculation for n-D arrays
#21089
Conversation
| if np.any([(sample == sample[0]).all() for sample in samples]): | ||
| msg = "An input array is constant; the statistic is not defined." | ||
| warnings.warn(stats.ConstantInputWarning(msg), stacklevel=2) | ||
| return AlexanderGovernResult(np.nan, np.nan) |
There was a problem hiding this comment.
See " Special case: statistic is NaN when variance is zero" below.
To produce a warning, I was thinking I'd replace means - var_w with _demean(means, var_w) when gh-21068 merges.
scipy/stats/_stats_py.py
Outdated
| warnings.warn(stats.ConstantInputWarning(msg), stacklevel=2) | ||
| return AlexanderGovernResult(np.nan, np.nan) | ||
| samples = _alexandergovern_input_validation(samples, nan_policy, axis) | ||
| samples = [np.moveaxis(sample, axis, -1) for sample in samples] |
There was a problem hiding this comment.
Important for the rest of the calculation:
axis0 corresponds with separate samplesaxis-1 is along a given sample- any other axes are batch dimensions.
| raise ValueError("Input sample size must be greater than one.") | ||
| if np.isinf(sample).any(): | ||
| raise ValueError("Input samples must be finite.") |
There was a problem hiding this comment.
Don't want to error out for infs in a single slice. Instead, the user will get a warning and NaN.
| args_int16 = [np.asarray(arg, dtype=np.int16) for arg in args] | ||
| args_int32 = [np.asarray(arg, dtype=np.int32) for arg in args] | ||
| args_uint8 = [np.asarray(arg, dtype=np.uint8) for arg in args] | ||
| args_float64 = [np.asarray(arg, dtype=np.float64) for arg in args] |
There was a problem hiding this comment.
Arrays happen to be the same size, but conceptually there are several separate samples.
|
@lucascolley re #21097 (comment), here's an example of adding |
|
Thanks @mdhaber . |
Reference issue
Toward gh-20544
What does this implement/fix?
Vectorizes
alexandergovernfor n-D arrays (i.e. adds nativeaxissupport).Additional information
I've also made some tweaks here that will reduce the diff when we add array API support.