[Numpy-discussion] Dot as a method

Bart Vandereycken bart.vandereycken at cs.kuleuven.be
Thu Jul 6 06:49:57 CDT 2006


Tim Hochberg wrote:
> Bart Vandereycken wrote:
>> Hi all,
>>
>> reading the thread "Ransom proposals" I was wondering why there isn't a 
>> ndarray.dot() method?  There is already a scipy.sparse.dot() so this 
>> would fit nicely in the whole idea of polymorphism.
>>   
> Are you sure about that?
> 
> The problem with a dot method (aside from a over proliferation of 
> methods in general) is that to be done correctly you want to choose a 
> particular implementation of dot based *both* of its arguments. A method 
> does a good job dispatching on a single argument, but falls down when 
> dispatching on two types. Let's look at this specific case. Imagine that 
> in addition ndarray.dot and sparse.dot, we also stick a dot method on 
> ma.masked, etc. Now in order to fully exploit polymorphism and get 
> maximum efficiency, we want asparsearray.dot(amaskedarray) to correctly 
> treat the masked values (ma.dot treats them as zeros) and to not 
> instantiate a dense version of the sparsearray. But in order to do that 
> all three classes need to know about the other two. That's possible, if 
> messy since all three of these are known in advance, but this approach 
> becomes untenable if you classes outside core numpy or scipy to 
> participate as full citizens.

You're right that you need to put the class specific dot code somewhere. 
   However, it would be better if this can be done in a clean way. By 
the way I don't fixate on methods or functions, I just want something 
useful ;-)

Suppose  I have written my own matrix class (custmatrix) with it's own 
fast matrix-vector product. How can I extend numpy code so that is uses 
this fast multiplication when I want a_custmatrix*a_ndarray or 
a_ndarray*a_custmatrix?


> What does appear fit well here is generic functions / multimethods / 
> protocols as discussed in some detail on pydev-3000 a couple of months 
> ago. This would allow classes defined outside of core numpy and scipy to 
> work correctly and efficiently with dot as long as they register 
> appropriate versions of dot. If I wasn't swamped right at the moment I'd 
> prototype this up and see how it works in practice.

Right, I was thinking about multimethods too after reading 
http://www.artima.com/weblogs/viewpost.jsp?thread=101605

The other solution that simulates double dispatching is the Visitor 
pattern, but it has a lot of disadvantages (difficult to add custom 
matrices, slow?, all the matrices have to be children of an abstract class).


Bart





More information about the Numpy-discussion mailing list