How can I speed up importing large .d files?

Hi,

soon I'll have large files, about 50000x6 (more than one of them) and I'll have to work on them several times to plot different stuff. I've realised that "importdata" command is very slow and I've been recommended to use GNUPlot, but I would prefer to stick to MATLAB if possible. So, is there any command to load data to workspace in a faster way? Maybe using .xls instead of .d?

Thanks.

[Cedric]

Could you provide a sample file?

[Felix Lauwaert]

I had to upload it as .txt because of website requirements. I tic-tocked the importdata and it was arround 12s, but I'll have to deal with way larger files soon.

Answer 1

[Cedric]

I would do something along the following line:

 buffer = fileread( 'test.txt' ) ; 
 data   = sscanf( buffer(76:end), '%f' ) ;
 data   = reshape( data, 6, [] )' ;

I built a file with more than 50k rows to test, and it takes half the time of IMPORTDATA. We can easily improve it so it doesn't rely on a fixed header size. You could also do something like:

 fName  = 'test.txt' ;
 fId    = fopen( fName, 'r' ) ;
 header = strsplit( strtrim( fgetl( fId ) ), ' ' ) ;
 data   = fscanf( fId, '%f' ) ;
 data   = reshape( data, 6, [] )' ;
 fclose( fId ) ;

but this is slower.

PS: I don't understand why your IMPORTDATA is that slow. On my test file with >50k lines, here is the timing:

 IMPORTDATA: Elapsed time is 0.312664 seconds.
 FSCANF    : Elapsed time is 0.457961 seconds.
 FILEREAD  : Elapsed time is 0.171987 seconds.

[Felix Lauwaert]

It was perfect, 0.09s, I'll have a look at it. Thanks

[per isakson]

"understand why your IMPORTDATA is that slow" &nbsp importdata (and txt2mat) are slow with long rows. It's because the time needed to parse the header and the first data row. Wasn't the original question about a file with long rows?

I'm surprised fileread in combination with sscanf is three times faster than fscanf.

[Cedric]

That was a surprise for me too actually!

About IMPORTDATA, I thought that the 12s were with the demo file that the OP attached to the question, but a version with 50k lines. I don't understand because I added lines for reaching ~50k, and the timings that I provide are for this version, with IMPORTDATA being timed at 0.3s.

[per isakson]

I created a test file with

cssm0(50000)

where

function    cssm0( N )
    h = sprintf( 'H%06d ', 1:N );
    d = sprintf( '%f '   , 1:N );
    fid = fopen('test_long_rows.txt','w');
    fprintf( fid, '%s\n', h,d,d,d,d,d,d );
    fclose(fid);
end

and used profile. Maybe, I misread the original question. Anyhow, whether the rows or the columns are long and short, respectively, makes a huge difference with importdata

[Cedric]

It's interesting, 6x5e4 -> 2s and 5e4x6 -> 0.3s.

In any case, I've always been avoiding IMPORTDATA like plague (especially after looking at its source code), because its behavior is size-dependent and difficult to predict. This leads to situations like the one reported recently on the forum, where it works with a files that contains thousands of rows of data, but fails when there are only 30 lines (for the same data structure).

Answer 2

[per isakson]

Does the file consist of header rows followed by data rows, which contains only numerical data? (No string data such as date time data.)

Try txt2mat, by Andres "txt2mat basically is a wrapper for sscanf, it quickly converts ascii files containing m-by-n numeric data, allowing for header lines"

[Felix Lauwaert]

It has a header line and then just double precision numbers. I uploaded an example file if you want to have a look.

[per isakson]

Yes, please upload the six row file.

[Felix Lauwaert]

I tried txt2mat out and it's great, timing 0.157s! Awesome, I hope one day I'll be ready to answer questions and make such functions :)