[Numpy-discussion] [SciPy-User] vectorized version of 'multinomial' sampling function

[per freem]

On Tue, Oct 13, 2009 at 7:59 PM, David Warde-Farley <dwf at cs.toronto.edu> wrote:
> On 13-Oct-09, at 5:01 PM, per freem wrote:
>
>> hi all,
>>
>> i have a series of probability vector that i'd like to feed into
>> multinomial to get an array of vector outcomes back. for example,
>> given:
>>
>> p = array([[ 0.9 ,  0.05,  0.05],
>>       [ 0.05,  0.05,  0.9 ]])
>>
>> i'd like to call multinomial like this:
>>
>> multinomial(1, p)
>>
>> to get a vector of multinomial samplers, each using the nth list in
>> 'p'. something like:
>>
>> array([[1, 0, 0], [0, 0 1]]) in this case. is this possible? it seems
>> like 'multinomial' takes only a one dimensional array. i could write
>> this as a "for" loop of course but i prefer a vectorized version since
>> speed is crucial for me here.
>>
>> thanks very much.
>
> Your best bet is probably to copy the pyrex/Cython code for
> multinomial in numpy/random/mtrand/mtrand.pyx, and add the
> functionality you want there.  If you do it right (i.e. type your loop
> indices) then it should be fast.
>
> David
> _______________________________________________
> SciPy-User mailing list
> SciPy-User at scipy.org
> http://mail.scipy.org/mailman/listinfo/scipy-user
>

Hi David

thanks for your reply. i am not sure how to do this though -- is the
vectorized version i would write in pyrex/cython simply going to
iterate through this vector of vectors and do the operation? will that
really be efficient?

is there some other library that can do vectorized multinomial like i
described? i really am not sure how to write this cython.