numpy numbers converted wrong

[Robert Kern]

robert wrote:
> in Gnuplot (Gnuplot.utils) the input array will be converted to a Numeric float array as shown below. When I insert a numpy array into Gnuplot like that below, numbers 7.44 are cast to 7.0 
> Why is this and what should I do ? Is this bug in numpy or in Numeric?
> 
> 
> [Dbg]>>> m         #numpy array
> array([[  9.78109200e+08,   7.44000000e+00],
>        [  9.78454800e+08,   7.44000000e+00],
>        [  9.78541200e+08,   8.19000000e+00],
>        ..., 
>        [  1.16162280e+09,   8.14600000e+01],
>        [  1.16170920e+09,   8.10500000e+01],
>        [  1.16179560e+09,   8.16800000e+01]])
> [Dbg]>>> Numeric.asarray(m, Numeric.Float32)[:10]
> array([[ 9.78109184e+008,  7.00000000e+000],
>        [ 9.78454784e+008,  7.00000000e+000],
>        [ 9.78541184e+008,  8.00000000e+000],
>        [ 9.78627584e+008,  8.00000000e+000],
>        [ 9.78713984e+008,  8.00000000e+000],
>        [ 9.78973184e+008,  8.00000000e+000],
>        [ 9.79059584e+008,  8.00000000e+000],
>        [ 9.79145984e+008,  8.00000000e+000],
>        [ 9.79232384e+008,  9.00000000e+000],
>        [ 9.79318784e+008,  8.00000000e+000]],'f')
> [Dbg]>>> Numeric.asarray(m, Numeric.Float)[:10]
> array([[ 9.78109200e+008,  7.00000000e+000],
>        [ 9.78454800e+008,  7.00000000e+000],

The problem is with the version of Numeric you are using. I can replicate this 
problem with Numeric 24.0 but not with 24.2.

> and why and what is:
> 
> [Dbg]>>> m[0,1]
> 7.44
> [Dbg]>>> type(_)
> <type 'numpy.float64'>
> [Dbg]>>> 

It is a scalar object. numpy supports more number types than Python does so the 
scalar results of indexing operations need representations beyond the standard 
int, float, complex types. These scalar objects also support the array 
interface, so it's easier to write generic code that may operate on arrays or 
scalars. Their existence also resolves the long-standing problem of maintaining 
the precision of arrays even when performing operations with scalars. In 
Numeric, adding the scalar 2.0 to a single precision array would return a 
double-precision array. Worse, if a and b are single precision arrays, (a+b[0]) 
would give a double-precision result because b[0] would have to be represented 
as a standard Python float.

The _Guide to NumPy_ has a discussion of these in Chapter 2, part of the sample 
chapters:

   http://numpy.scipy.org/numpybooksample.pdf

> does this also slow down python math computations? 

If you do a whole lot of computations with scalar values coming out of arrays, 
then yes, somewhat. You can forestall that by casting to Python floats or ints 
if that is causing problems for you.

> should one better stay away from numpy in current stage of numpy development?
 > I remember, with numarray there were no such problems.

Not really, no.

-- 
Robert Kern

"I have come to believe that the whole world is an enigma, a harmless enigma
  that is made terrible by our own mad attempt to interpret it as though it had
  an underlying truth."
   -- Umberto Eco