No worries.
It helps a lot to keep track of what dimension you think each thing
has, in a little comment like in my examples below. After a while,
you'll begin thinking naturally that way...
Most common operations have zero active dimensions, so every dimension
is a thread dimension. (Multiplication is like that). The collapse
and reduce operators all have one active dimension (e.g. "maximum" and
"sumover").
Cheers,
Craig
On Apr 16, 2008, at 6:10 AM, Ashish Mahabal wrote:
>
> Hello Craig,
>
> Thanks for the quick tutorial. That recipe worked and it has made
> the program hugely faster. Thanks.
>
> Indeed I still have not come to terms with the dimension
> manipulations and threading. Getting there, slowly.
>
> Cheers,
> ashish
>
> On Tue, 15 Apr 2008, Craig DeForest wrote:
>
>> You are looping because you are not trusting the threading engine.
>>
>> The strategy is to push your unused dimensions above the active
>> dimensions of your operator, and then let them ride along. Matrix
>> multiplication has two active dimensions, so you have to add dummy
>> dimensions to make your data into a collection of (1,n,x,y) column
>> vectors and (n,1,x,y) row vectors. Then you can just carry out
>> your multiplication. You probably want to get out of the habit of
>> using 'slice' and into the habit of using "PDL::NiceSlice", which
>> yields a much cleaner syntax.
>>
>> $colvecs = $allims->(*1); # colvecs dims: (1,n,x,y)
>> $rowvecs = $allims->(:,*1); # rowvecs dims: (n,1,x,y)
>> $result = $rowvecs x $matrix x $colvecs; # result dims: (1,1,x,y)
>> $out = $result->clump(3); # out dims: (x,y)
>>
>> A little more concise:
>>
>> $vecs = $allims->(*1);
>> $out = ($vecs->transpose x $matrix x $vecs)->clump(3);
>>
>> You can even do it in one expression if you prefer:
>>
>> $out = ($allims->(:,*1) x $matrix x $allims->(*1))->clump(3);
>>
>>
>> BTW, your '->cat' call is doing nothing, since it sticks together a
>> bunch of PDLs and you're just giving it one PDL.
>>
>> Cheers,
>> Craig
>>
>>
>> On Apr 15, 2008, at 12:49 PM, Ashish Mahabal wrote:
>>> Hi,
>>> I know there has to be a better, more succinct way of doing this,
>>> but I
>>> just don't seem to get it, not having done any explicit threading:
>>> $allims is a stack of n images of size (xsize,ysize) from which I
>>> need to
>>> pull out vectors that have the same pixel for each image in the
>>> stack and
>>> do some matrix operations on it. The loops clearly slow it down.
>>> How do I
>>> get rid of the loops? ($invcovari is a nxn matrix).
>>> my $out= ones($xsize, $ysize);
>>> for my $j (0..$xsize-1){
>>> for my $k (0..$ysize-1){
>>> my $vecf = (slice $allims,"*,$j,$k")->flat->cat;
>>> my $vecft = $vecf->transpose;
>>> (slice $out,"$j,$k") .= $vecf x ($invcovari x $vecft);
>>> }
>>> }
>>> Thanks.
>>> -ashish
>>> Ashish Mahabal, Caltech Astronomy, Pasadena, CA 91125
>>> http://www.astro.caltech.edu/~aam aam at astro.caltech.edu
>>> Why is "abbreviation" such a long word?
>>> _______________________________________________
>>> Perldl mailing list
>>> [email protected]
>>> http://mailman.jach.hawaii.edu/mailman/listinfo/perldl
>>
>
> -ashish
>
> Ashish Mahabal, Caltech Astronomy, Pasadena, CA 91125
> http://www.astro.caltech.edu/~aam aam at astro.caltech.edu
>
> My computer isn't that nervous...it's just a bit ANSI.
>
_______________________________________________
Perldl mailing list
[email protected]
http://mailman.jach.hawaii.edu/mailman/listinfo/perldl
