[Numpy-discussion] [Cdat-discussion] Arrays containing NaNs

[Pierre GM]

Oh, I guess this one's for me...

On Thursday 01 January 1970 04:21:03 Charles Doutriaux wrote:

> Basically it was suggested to automarically mask NaN (and Inf ?) when
> creating ma.
> I'm sure you already thought of this on this list and was curious to
> know why you decided not to do it.

Because it's always best to let the user decide what to do with his/her data 
and not impose anything ?

Masking a point doesn't necessarily mean that the point is invalid (in the 
sense of NaNs/Infs), just that it doesn't satisfy some particular condition. 
In that sense, masks act as selecting tools.

By forcing invalid data to be masked at the creation of an array, you run the 
risk to tamper with the (potential) physical meaning of the mask you have 
given as input, and/or miss the fact that some data are actually invalid when 
you don't expect it to be.

Let's take an example: 
I want to analyze sea surface temperatures at the world scale. The data comes 
as a regular 2D ndarray, with NaNs for missing or invalid data. In a first 
step, I create a masked array of this data, filtering out the land masses by 
a predefined geographical mask. The remaining NaNs in the masked array 
indicate areas where the sensor failed... It's an important information I 
would probably have missed by masking all the NaNs at first...


As Eric F. suggested, you can use numpy.ma.masked_invalid to create a masked 
array with NaNs/Infs filtered out:

>>>import numpy as np,. numpy.ma as ma
>>>x = np.array([1,2,None,4], dtype=float)
>>>x
array([  1.,   2.,  NaN,   4.])
>>>mx = ma.masked_invalid(x)
>>>mx
masked_array(data = [1.0 2.0 -- 4.0],
      mask = [False False  True False],
      fill_value=1e+20)

Note that the underlying data still has NaNs/Infs:
>>>mx._data
array([  1.,   2.,  NaN,   4.])

You can also use the ma.fix_invalid function: it creates a mask where the data 
is not finite (NaNs/Infs), and set the corresponding points to fill_value.
>>>mx = ma.fix_invalid(x, fill_value=999)
>>>mx
masked_array(data = [1.0 2.0 -- 4.0],
      mask = [False False  True False],
      fill_value=1e+20)
>>>mx._data
array([   1.,    2.,  999.,    4.])


The advantage of the second approach is that you no longer have NaNs/Infs in 
the underlying data, which speeds things up during computation. The obvious 
disadvantage is that you no longer know where the data was invalid...