[Numpy-discussion] numpy.ma.compress

[Pierre GM]

On Thursday 24 January 2008 15:58:14 Stefan van der Walt wrote:

> How about masking the output where the condition is masked?
> I.e. keep where condition is True, remove where condition is False and
> mask where condition is masked.

Won't systematically do, check one of the examples I gave in a previous 
emails.
>>>a=masked_array([1,2,3,4,5],mask=[0,0,1,1,0])
>>>a.compress((a<4).filled(True))
masked_array(data = [1 2 -- --],
      mask = [False False  True  True],
      fill_value=999999)


Fundamentally, it's a very bad idea to use a masked array as a condition: what 
should be done when the condition is masked ? Here, we're going against one 
of the basic principles of Paul Dubois, the original author: we are making 
assumptions about the masked values of the condition when we consider the 
condition as a ndarray.
In timeseries, we prevent the use of masked arrays as conditions. That's not 
much of a problem, as we already overloading __getitem__ to a rather ugly 
extent. I'm a bit reluctant to introduce that feature in numpy.ma, as it 
would crash the performance.
That's something we may want to consider when we'll port numpy.ma to C: a 
masked condition is undefined, therefore cannot be used in selecting array 
elements, therefore should be set to False.