[Genabel-commits] r1658 - pkg/ProbABEL/src

[Yurii Aulchenko]

Sum of logs = log of prods

The latter is in principle dangerous practice when we multiply
probabilities - the product is getting close to zero very fast
potentially leading to numerical problems (of which machine zero is
the smallest because easily detected)

Yurii

----------------------
Yurii Aulchenko
(sent from mobile device)

> On Mar 21, 2014, at 11:57 PM, "noreply at r-forge.r-project.org" <noreply at r-forge.r-project.org> wrote:
>
> Author: lckarssen
> Date: 2014-03-21 23:57:20 +0100 (Fri, 21 Mar 2014)
> New Revision: 1658
>
> Modified:
>   pkg/ProbABEL/src/reg1.cpp
> Log:
> Speed-ups in ProbABEL's logistic regression.
> Profiling showed lots of (expensive) calls to exp() and log(). I got rid of ~ 1/3 of the exp() calls by saving the result in a variable and reusing it in the calculation of exp(mu) / ( 1+exp(mu) ).
> The number of log() calls was reduced even more by using the fact that sum_i( log(x_i) ) = log( prod_i(x_i) )
>
> In total this gives roughly a speed up of 30% -- 40% when reading txt dosage files and roughly 20% -- 30% when using filevector files (measured for dosage data).
>
>
> Modified: pkg/ProbABEL/src/reg1.cpp
> ===================================================================
> --- pkg/ProbABEL/src/reg1.cpp    2014-03-21 21:05:44 UTC (rev 1657)
> +++ pkg/ProbABEL/src/reg1.cpp    2014-03-21 22:57:20 UTC (rev 1658)
> @@ -719,7 +719,8 @@
>             double emu = eMu.get(i, 0);
>             double value = emu;
>             double zval;
> -            value = exp(value) / (1. + exp(value));
> +            double expval = exp(value);
> +            value = expval / (1. + expval);
>             residuals[i] = (rdata.Y).get(i, 0) - value;
>             eMu.put(value, i, 0);
>             W.put(value * (1. - value), i, 0);
> @@ -778,11 +779,24 @@
>             beta.print();
>         }
>         // std::cout << "beta:\n"; beta.print();
> -        // compute likelihood
> +
> +        // Compute the likelihood. The following commented code gives
> +        // the 'easy to understand' algorithm. The code that's
> +        // actually used is mathematically equivalent (remember:
> +        // log(a*b) = log(a)+log(b)), but faster because log() is
> +        // relatively expensive.
> +        //    for (int i = 0; i < eMu.nrow; i++) {
> +        //          loglik += rdata.Y[i] * eMu_us[i] - log(1. +
> +        //                    exp(eMu_us[i]));
> +        //    }
>         prevlik = loglik;
>         loglik = 0.;
> -        for (int i = 0; i < eMu.nrow; i++)
> -            loglik += rdata.Y[i] * eMu_us[i] - log(1. + exp(eMu_us[i]));
> +        double logterm = 1;
> +        for (int i = 0; i < eMu.nrow; i++) {
> +            loglik  += rdata.Y[i] * eMu_us[i];
> +            logterm *= 1. + exp(eMu_us[i]);
> +        }
> +        loglik += - log(logterm);
>
>         delta = fabs(1. - (prevlik / loglik));
>         niter++;
>
> _______________________________________________
> Genabel-commits mailing list
> Genabel-commits at lists.r-forge.r-project.org
> https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/genabel-commits