|
|
Blitz Support : |
From: Julian Cummings (cummings_at_[hidden])
Date: 2003-10-14 13:04:51
Hi Roger,
I was thinking about the all-but-one indexing problem you described. I
wonder how the indexing code you show below is actually implemented
internally. It seems to me that it would be terribly inefficient to use
indirect addressing to do this, which is what normally happens when you
index an array with an array of arbitrary indices. It is probably much
better to handle the all-but-one indexing using two ordinary index
ranges: one for all the indices below the one being excluded and another
for all the indices above. That way, you take advantage of the fact
that almost all of your indexing is contiguous (unit stride). The minor
disadvantage is that you would have to repeat any expressions involving
this all-but-one indexing twice, once for the "lower" indices and once
for the "upper" indices.
But this would still be trivial to code up in Blitz using Range objects.
Regards, Julian C.
roger.grant wrote:
>Todd,
>I have enclosd an example, sorry for the formatting problem :)
>Regards,
>Roger
>
>//
>class rgGRNN : public doubleObjFunction
>{
>private:
>// passed arrays
>DoubleVec actualOUTCOME;
>DoubleArray actualINPUTS;
>DoubleVec predictedOUTCOME;
>// some good stuff to only build once
>int numROWS, numCOLS;
>
>public:
>
>virtual double objFunc(const DoubleVec &) const;
>virtual DoubleVec rgGRNN::objGrad(const DoubleVec &);
>virtual DoubleSymMatrix objHess(const DoubleVec &);
>rgGRNN(const DoubleArray &, const DoubleVec &, DoubleVec &);
>DoubleVec rgGRNN::getOUTCOME();
>
>};
>
>rgGRNN::rgGRNN(const DoubleArray &x, const DoubleVec &y, DoubleVec &z)
>: actualINPUTS(x),actualOUTCOME(y),predictedOUTCOME(z)
>{
> numROWS = rows(actualINPUTS);
> numCOLS = cols(actualINPUTS);
>}
>
>DoubleVec rgGRNN::getOUTCOME()
>{
> return predictedOUTCOME;
>}
>
>double rgGRNN::objFunc(const DoubleVec & inSIGMA) const
>{
>// DoubleArray rowSIGMA = inSIGMA;
>// rowSIGMA = rowSIGMA.Transpose();
> int i,j,k;
> double holdROWS;
> double sumDENOM;
> double sumNUMER;
> double partialERROR;
> double retERROR = 0;
> Index oneELEMENT = 0;
> Index oneCOLUMN = 0;
> Index allCOLS(0,numCOLS,1);
> int * includeINDEX = new int [numROWS-1];
> double *dptr;
> double temp;
> for (i=0; i<numROWS; i++)
> {
> partialERROR = 0;
> sumDENOM = 0;
> sumNUMER = 0;
> k=0;
>
>// This is the section where it builds the index using numROWS -1
>
> for (j=0; j<numROWS; j++)
> {
> if (j != i)
> {
> includeINDEX[k] = j;
> k++;
> }
> }
> Index testROWS(includeINDEX, numROWS-1);
>// As you can see the Index is built from an array
>
> oneCOLUMN = 0;
> dptr = &inSIGMA(0);
> for (j=0; j< numCOLS; j++)
> {
> temp = *dptr;
> if (temp == 0) {}
> else {
> DoubleArray tempX = actualINPUTS(testROWS, oneCOLUMN) -
>actualINPUTS(oneELEMENT, oneCOLUMN);
> tempX /= inSIGMA(j);
> tempX *= tempX;
> tempX += 1;
> tempX = 1 / tempX;
> sumDENOM += sum(tempX);
> tempX *= actualOUTCOME(testROWS);
> sumNUMER += sum(tempX);
> }
> oneCOLUMN++;
> dptr++;
> }
> if (sumDENOM == 0)
> {
> sumDENOM = 1e-40;
> }
> partialERROR = sumNUMER / sumDENOM;
> predictedOUTCOME(oneELEMENT) = partialERROR;
> DoubleVec trialOUTCOME = actualOUTCOME(oneELEMENT);
> partialERROR -= trialOUTCOME(0);
> partialERROR *= partialERROR;
>//
> holdROWS = (i / (numROWS - 1));
> holdROWS *= 4;
> holdROWS++;
> holdROWS /= 5;
> partialERROR *= holdROWS;
>//
> retERROR += partialERROR;
> oneELEMENT++;
> }
>
> delete [] includeINDEX;
> retERROR /= numROWS;
> return retERROR;
>}
>
>
>----- Original Message -----
>From: "Todd Veldhuizen" <tveldhui_at_[hidden]>
>To: "roger.grant" <roger.grant_at_[hidden]>
>Cc: <blitz-support_at_[hidden]>
>Sent: Sunday, October 12, 2003 10:49 AM
>Subject: Re: [Blitz-support] question
>
>
>
>
>>Hi Roger,
>>
>>Could you post an example of the kind of iteration you'd like to do,
>>(i.e. an excerpt of M++ code)?
>>
>>Regarding BLAS, does your code spend a substantial amount of time in an
>>array operation that corresponds to a BLAS routine? Blitz++ is designed to
>>get decent performance for any array expression, whereas BLAS provides a
>>selected set of fast implementations of common expressions in linear
>>algebra. The idea of invoking BLAS from blitz when possible has been
>>kicked around, but so far it seems like nobody has had their code run
>>unacceptably slow because BLAS isn't used (or, they've been willing
>>to grab pointers and invoke blas themselves.) I guess it just hasn't
>>seemed worth the effort to anyone to write the BLAS hooks.
>>
>>Cheers,
>>Todd
>>
>>
>>On Fri, 10 Oct 2003, roger.grant wrote:
>>
>>
>>>I am currently using the M++ package on Windows. I am transferring my
>>>
>>>
>work to FreeBSD and the M++ is not easily changeable to go to a UNIX
>environment and is also of very old technology as compared to Bltz++.
>Anyway, M++ defines an Index as a class and I am able to define a special
>index which I will call a Leave-One-Out index. I can then transverse my
>array iteratively leaving one out.
>
>
>>>Example:
>>>Define a Vector of 100 items.
>>>define a leave-one-out index which will be of size 99.
>>>pass thru the Vector 100 times; each time leaving out a different point
>>>
>>>
>( from 0 to 99)
>
>
>>>This is VERY useful for working to develope pattern matching algorithms
>>>
>>>
>as you never use the point that you are trying to predict within the
>algorithm. My software may make 1000's of passes on the Vector/Array working
>to optimize the algorithm, to minimize the "error". The Vector/Array is only
>created once and only the "view" of the data is changed.
>
>
>>>How can I accomplish this with Blitz++???
>>>Thank You!
>>>Roger Grant
>>>PS: I am surprised that you do not use the blas libraries. Wouldn't this
>>>
>>>
>lead to a severe performance hit as most architectures have optimized
>assembler coded blas?
>
>
>>--
>>Todd Veldhuizen / tveldhui_at_[hidden] / Indiana University Computer
>>
>>
>Science
>
>
>
>_______________________________________________
>Blitz-support mailing list
>Blitz-support_at_[hidden]
>http://www.oonumerics.org/mailman/listinfo.cgi/blitz-support
>
>
-- Dr. Julian C. Cummings E-mail: cummings_at_[hidden] California Institute of Technology Phone: 626-395-2543 1200 E. California Blvd., Mail Code 158-79 Fax: 626-584-5917 Pasadena, CA 91125