[SciPy-user] sparse+slice? (and hello!)

[Ed Schofield]

Bram de Jong - MTG wrote:
> Hi,
>
> Ed Schofield wrote:
>   
>> On 17/05/2006, at 9:19 AM, Bram de Jong - MTG wrote:
>>     
>>> * A simple question to start of with, is it correct to say that the
>>> sparse matrices do not support slicing?
>>>       
>> I've added some support for slicing in more recent SVN versions.   
>> It's now possible to use csc[:, j], csr[i, :], or lil[i, :].  These  
>> should be efficient operations.  But slicing along the 'non-native'  
>> axes, with csc[i, :], csr[:, j], or lil[:, j] would be slower (and  
>> more complex to write), and these raise an IndexError.
>>     
>
> Oh, that's great news... Is this Python-only, or C code? Just to know if 
> I will have to recompile. (Are there daily builds somewhere?)
>   

It's Python-only, but I suggest you re-compile the package, since some
bugs were also fixed in the sparsetools Fortran code.  You'll need a
fairly recent NumPy version too (I think at least 0.9.6).

> Perhaps slicing along the non-native axes could be supported as well, 
> even if it's really slow... A simple docstring with a message about the 
> speed would solve the problem, while supplying the user with a "full 
> set". For small matrices or really sparse ones it won't be much of a 
> problem I guess (?).
>   

If you don't mind it being slow, an easy way to get slices along
non-native axes is to convert the matrix format first.  For example,

>>> csc.tocsr()[i, :]

does the trick :)

>>> * a not-so-simple question: is there any documentation, tutorial,
>>> introduction, ... to the sparse matrices of scipy? It seems neither  
>>> the
>>> official (feebased) docs mention them, nor are the docstrings *that*
>>> extensive.
>>>       
>> There are also more docstrings in SVN HEAD.  No tutorial yet,  
>> although we've had several requests recently.  I'll see what I can do  
>> after I've submitted my thesis in a few weeks ...
>>     
>
> Ok, that's great as well...
>
> thanks for the rapid reply!
>   

You're welcome!

-- Ed