Fastest possible way to convert non-numbers to zeros?
in [Matlab]
|
Prev: Help in patch command ...
Next: Fault signal
From: Michael Snyder on 7 Jun 2010 10:11 Hi Y'all, I am working with a large 4D matrix using many iterations to solve a problem. My problem is once in a while my matrix operations produce a non-number, inf, -inf, or even an nan. There is no easy way to know ahead of time when this happens. I am looking for the fastest possible way to convert the non-numbers to zeros between iterations. If I convert the non-numbers to zeros between iterations then it does not effect my final results. I have gotten my main iteration loop down to 1.5 seconds. If I use the following code to convert non-numbers to zeros, it adds 2.7 seconds to my loop. This isn’t good. I have tried converting my 4D matrix to 3D matrixes, but it doesn’t speed up my code in this case maxtrix_out(maxtrix_out==Inf)=0; maxtrix_out(maxtrix_out==-Inf)=0; maxtrix_out(maxtrix_out==NaN)=0; There has to be a faster way. Can anyone point me in the right direction? Thanks, Michael
From: John D'Errico on 7 Jun 2010 10:27 "Michael Snyder" <sirzerp(a)mathworks.com> wrote in message <huiulq$75v$1(a)fred.mathworks.com>... > Hi Y'all, > > I am working with a large 4D matrix using many iterations to solve a problem. > > My problem is once in a while my matrix operations produce a non-number, inf, -inf, or even an nan. There is no easy way to know ahead of time when this happens. > > I am looking for the fastest possible way to convert the non-numbers to zeros between iterations. If I convert the non-numbers to zeros between iterations then it does not effect my final results. > > I have gotten my main iteration loop down to 1.5 seconds. If I use the following code to convert non-numbers to zeros, it adds 2.7 seconds to my loop. This isn’t good. I have tried converting my 4D matrix to 3D matrixes, but it doesn’t speed up my code in this case > > maxtrix_out(maxtrix_out==Inf)=0; > maxtrix_out(maxtrix_out==-Inf)=0; > maxtrix_out(maxtrix_out==NaN)=0; > > There has to be a faster way. Can anyone point me in the right direction? Well, actually, you are wasting some time here. Try this test... >> nan == nan ans = 0 Instead, read these. help isnan help isinf matrix(isnan(matrix) | isinf(matrix)) = 0; John
From: Michael Snyder on 7 Jun 2010 10:41 "John D'Errico" <woodchips(a)rochester.rr.com> wrote in message <huivka$e2u$1(a)fred.mathworks.com>... > "Michael Snyder" <sirzerp(a)mathworks.com> wrote in message <huiulq$75v$1(a)fred.mathworks.com>... > > Hi Y'all, > > > > I am working with a large 4D matrix using many iterations to solve a problem. > > > > My problem is once in a while my matrix operations produce a non-number, inf, -inf, or even an nan. There is no easy way to know ahead of time when this happens. > > > > I am looking for the fastest possible way to convert the non-numbers to zeros between iterations. If I convert the non-numbers to zeros between iterations then it does not effect my final results. > > > > I have gotten my main iteration loop down to 1.5 seconds. If I use the following code to convert non-numbers to zeros, it adds 2.7 seconds to my loop. This isn’t good. I have tried converting my 4D matrix to 3D matrixes, but it doesn’t speed up my code in this case > > > > maxtrix_out(maxtrix_out==Inf)=0; > > maxtrix_out(maxtrix_out==-Inf)=0; > > maxtrix_out(maxtrix_out==NaN)=0; > > > > There has to be a faster way. Can anyone point me in the right direction? > > Well, actually, you are wasting some time here. > Try this test... > > >> nan == nan > ans = > 0 > > Instead, read these. > > help isnan > help isinf > > matrix(isnan(matrix) | isinf(matrix)) = 0; > > John Thanks John Maybe 2.7 seconds wasn't that bad. :) 46.640 1 x 43 maxtrix_out(isnan(maxtrix_out) | isinf(maxtrix_out)) = 0 I try your command with 3d matrixes, but for sure it doesn't like 4d, at 46 seconds. :)
From: Walter Roberson on 7 Jun 2010 10:44 Michael Snyder wrote: > I am working with a large 4D matrix using many iterations to solve a > problem. > My problem is once in a while my matrix operations produce a non-number, > inf, -inf, or even an nan. There is no easy way to know ahead of time > when this happens. > > I am looking for the fastest possible way to convert the non-numbers to > zeros between iterations. If I convert the non-numbers to zeros between > iterations then it does not effect my final results. The most direct way (though I do not promise it is the fastest available) would be matrix_out(~isfinite(matrix_out)) = 0;
From: Walter Roberson on 7 Jun 2010 10:47
Michael Snyder wrote: > Maybe 2.7 seconds wasn't that bad. :) > > 46.640 1 x 43 maxtrix_out(isnan(maxtrix_out) | > isinf(maxtrix_out)) = 0 > I try your command with 3d matrixes, but for sure it doesn't like 4d, at > 46 seconds. :) Those command should be very nearly insensitive to number of dimensions, as this is just straight-forward logical indexing, which internally just copies the matrix dimensions and then sequences through consequative memory locations. |