[SciPy-user] NaN's in numpy (and Scipy)

[A. M. Archibald]

On 06/01/07, Vincent Nijs <v-nijs at kellogg.northwestern.edu> wrote:
> It may be relevant to note that 'isnan' only seems to work with floats. If
> you change Pierre's example a bit, mask creation doesn't work as I might
> expect. If you create the array as follows
>
> >>> x = N.array([1,2,N.nan,4],dtype='int16')

[...]

> It might be convenient if an array would be automatically up-cast to float
> is an nan is present.

This is in fact the normal behaviour of numpy: if there's a
floating-point number in the array when you create it, the array gets
upcast - unless you specify the data type, forcing it to a particular
type. Which you did.

The unavailablity of NaNs for integer types is sometimes unfortunate,
but there's really nothing numpy can do about it.

A. M. Archibald