[Numpy-discussion] Really strange result

[Bruce Southey]

Neal Becker wrote:
> Charles R Harris wrote:
>
>   
>> On Fri, May 1, 2009 at 7:39 PM, Charles R Harris
>> <charlesr.harris at gmail.com>wrote:
>>
>>     
>>> On Fri, May 1, 2009 at 7:24 PM, Neal Becker <ndbecker2 at gmail.com> wrote:
>>>
>>>       
>>>> Charles R Harris wrote:
>>>>
>>>>         
>>>>> On Fri, May 1, 2009 at 1:02 PM, Neal Becker <ndbecker2 at gmail.com>
>>>>>           
>>>> wrote:
>>>>         
>>>>>> In [16]: (np.linspace (0, len (x)-1, len(x)).astype
>>>>>>             
>>>> (np.uint64)*2).dtype
>>>>         
>>>>>> Out[16]: dtype('uint64')
>>>>>>
>>>>>> In [17]: (np.linspace (0, len (x)-1, len(x)).astype
>>>>>>             
>>>> (np.uint64)*n).dtype
>>>>         
>>>>>> Out[17]: dtype('float64')
>>>>>>
>>>>>> In [18]: type(n)
>>>>>> Out[18]: <type 'int'>
>>>>>>
>>>>>> Now that's just strange.  What's going on?
>>>>>>
>>>>>>
>>>>>>             
>>>>> The  n is signed, uint64 is unsigned. So a signed type that can hold
>>>>> uint64 is needed. There ain't no such integer, so float64 is used. I
>>>>>           
>>>> think
>>>>         
>>>>> the logic here is a bit goofy myself since float64 doesn't have the
>>>>>           
>>>> needed
>>>>         
>>>>> 64 bit precision and the conversion from int kind to float kind is
>>>>> confusing. I think it would be better to raise a NotAvailable error or
>>>>> some such. Lest you think this is an isolated oddity, sometimes
>>>>> numeric arrays can be converted to object arrays.
>>>>>
>>>>> Chuck
>>>>>           
>>>> I don't think that any type of integer arithmetic should ever be
>>>> automatically promoted to float.
>>>>
>>>> Besides that, what about the first example?  There, I used '2' rather
>>>> than
>>>> 'n'.  Is not '2' also an int?
>>>>         
>>> What version of numpy are you using?
>>>
>>>       
>> And what is the value of n?
>>
>>     
>
>   
>> Chuck 
>>     
>
> np.version.version
> Out[5]: '1.3.0'
> (I think the previous test was on 1.2.0 and did the same thing)
>
> (np.linspace (0, 1023,1024).astype(np.uint64)*2).dtype
> Out[2]: dtype('uint64')
>
> In [3]: n=-7
>
> In [4]: (np.linspace (0, 1023,1024).astype(np.uint64)*n).dtype
> Out[4]: dtype('float64')
>
>
> _______________________________________________
> Numpy-discussion mailing list
> Numpy-discussion at scipy.org
> http://mail.scipy.org/mailman/listinfo/numpy-discussion
>   
Hi,
//I think this behavior has been raised before. IIRC, Numpy is trying to 
do the operation that is requested by converting the dtype into floats 
since this is a generic solution that will avoid overflow with any ints 
not just unsigned ints.

Note that you get a different result if you use subtraction than 
multiplication.
 >>> np.linspace (0, 1023,1024)
array([  0.00000000e+00,   1.00000000e+00,   2.00000000e+00, ...,
         1.02100000e+03,   1.02200000e+03,   1.02300000e+03])
 >>> np.linspace (0, 1023,1024).astype(np.uint64)*-7
array([ -0.00000000e+00,  -7.00000000e+00,  -1.40000000e+01, ...,
        -7.14700000e+03,  -7.15400000e+03,  -7.16100000e+03])
 >>> np.linspace (0, 1023,1024).astype(np.uint64)-7
array([18446744073709551609, 18446744073709551610, 18446744073709551611,
       ...,                 1014,                 1015,
                       1016], dtype=uint64)



Bruce