[SciPy-dev] [SciPy-user] default dtype

[Ed Schofield]

Francesc Altet wrote:

>A Dilluns 02 Gener 2006 03:38, Ed Schofield va escriure:
>  
>
>>+5 on changing the default to float.  I think we'd look back on this
>>decision in several years as difficult but right.
>>    
>>
>Sorry, but I don't agree. If we want Python to include the container
>for array objects, making the default be double seems to stress the
>fact that this object is meant primarily for scientific use (which is
>true to some extent). However, for the sake of stablishing a *real*
>standard to keep datasets, I'd advocate the default to remain int. In
>addition, there are a lot of uses for integer arrays (indices,
>images...). IMO, making the double the default would discourage the
>use of the object between people not used to write
>scientific/technical apps.
>
>The only issue is the possible confusion in users when they will
>receive Int32 arrays in 32-bit platforms and Int64 arrays in 64-bit
>ones. I don't know, but perhaps this single reason is strong enough to
>change the default to Float64.
>  
>
Interesting point.

Then what do you think about an integer default that is redefinable by 
the user?  For example:

 >>> import scicore / whatever
 >>> scicore.default_dtype = float64

Then zeros(), empty(), and ones() (any others?) would use the new 
default.  I think the flexibility would be nice, and it should be 
feasible ...

Perhaps the most compelling argument against an integer default is the 
current behaviour with unsafe casting:

 >>> a = zeros(10)
 >>> a[0] = 1.159262
 >>> a[0]
1

and this argument would evaporate if unsafe casts were required to be 
more explicit.  I promised to provide a patch and run some timings for 
this, but I haven't done this yet.


-- Ed