Matlab: slow with converting date strings to datenum?

Question

With 100,000 date strings to convert to datenums, Matlab is really slow... 20 seconds!! Is there a faster way of doing this?

%strDay  
%strTime

dblDate = zeros(1, nLines);
for i = 1 : nLines
    dblDate(j) = datenum([strDay{i}, ' ', strTime{i}], 'yy.mm.dd HH:MM:SS.FFF');
    j = j + 1;
end

Unfortunately, `dblDate = cellfun(@(x,y) datenum([x,' ',y],'yy.mm.dd HH:MM:SS.FFF'),strDay,strTime,'UniformOutput',false)` takes 26 seconds. — Doresoom, Feb 20 '13 at 14:33

score 4 · Answer 1 · answered Feb 20 '13 at 17:36

Internally, datenum does a large amount of fiddling around to process many different date formats. Then it calls the internal function datenummx with different input arguments depending on the date format and the syntax with which you called datenum.

If you know ahead of time which syntax you will be calling datenum with, and which date format you will be using, you can call datenummx directly. Since datenummx is a built-in MEX file, and avoids all the overhead processing, it is typically vastly faster.

Take a look into datenum (type edit datenum) to find out the appropriate way to call datenummx for your application.

(Plus, do all the vectorization that other answers have suggested).

H.Muster · Answer 2 · 2013-02-20T14:53:42.687

You could try to avoid the loop:

tmp = repmat({' '}, length(strDay), 1);
datenum(cell2mat([strDay  (:), tmp(:),  strTime(:)]))

~edit~
A little bit of benchmarking of all the solutions so far:

% creating some bogus data
a = ones(10000,1).*(now.*rand(10000,1));
str = datestr(a,'yy.mm.dd HH:MM:SS.FFF');
strDay = cellstr(str(:,1:8));
strTime = cellstr(str(:,10:end));

%% jdl's original code
disp('jdl''s original code')
tic
nLines = length(strDay);
dblDate = zeros(1, nLines);
j = 1;
for i = 1 : nLines
    dblDate(j) = datenum([strDay{i}, ' ', strTime{i}]);
    j = j + 1;
end
toc

%% hmuster's solution
disp('hmuster''s solution')
tic
tmp = repmat({' '}, length(strDay), 1);% cell(size(str1));
b = datenum(cell2mat([strDay(:), tmp(:),  strTime(:)]));
toc

%% ThijsW's solution
disp('ThijsW''s solution')
tic
strDay = cell2mat(strDay);
strTime = cell2mat(strTime);

dblDate = datenum([strDay, (' ')*ones(size(strTime)), strTime], 'yy.mm.dd HH:MM:SS.FFF');
toc

% jdl's solution
disp('jdl''s solution')
tic
dblDate = datenum(strDay, 'yy.mm.dd') + ...
    datenum(strTime, 'HH:MM:SS.FFF') - ...
    datenum('00:00:00.000', 'HH:MM:SS.FFF');
toc

This results in:

jdl's original code
   Elapsed time is 19.624597 seconds.
hmuster's solution
   Elapsed time is 4.029291 seconds.
ThijsW's solution
   Elapsed time is 0.183376 seconds.
jdl's solution
   Elapsed time is 0.222996 seconds.

issue with cell2mat... the date string coming can vary in width, ie: 3.5.12, 10.25.12 and will yield "Error using cat Dimensions of matrices being concatenated are not consistent. ", is there a work around? thx — jdl, Feb 20 '13 at 14:56
Since your own solution is quite as fast as the one from ThijsW, you maybe should not bother with the `cell2mat` issue and simply use your own solution? — H.Muster, Feb 20 '13 at 15:01
cell2mat is unique, good to use when lengths are the same. Be nice to use for varying lengths as well. — jdl, Feb 20 '13 at 15:11

score 2 · Answer 3 · answered Feb 20 '13 at 14:28

2

Try it without the for loop. datenum should be able to handle a vectors.

strDay = cell2mat(strDay);
strTime = cell2mat(strTime);

dblDate = datenum([strDay, (' ')*ones(size(strTime)), strTime], 'yy.mm.dd HH:MM:SS.FFF');

answered Feb 20 '13 at 14:28

ThijsW

2,599
15
17

issue with cell2mat... the date string coming can vary in width, ie: 3.5.12, 10.25.12 and will yield "Error using cat Dimensions of matrices being concatenated are not consistent. ", is there a work around? thx – jdl Feb 20 '13 at 14:37
You may want to check out this question: http://stackoverflow.com/questions/6210495/how-to-combine-vectors-of-different-length-in-a-cell-array-into-matrix-in-matlab. I'm not sure if padding with `' '` will give problems with datenum. Sorry, but I have no time to try it myself right now. – ThijsW Feb 20 '13 at 14:45

score 2 · Answer 4 · answered Feb 20 '13 at 14:40

2

This is what I found to reduce time from 20 sec to 2 sec:

dblDate = datenum(strDay, 'yy.mm.dd') + ...
          datenum(strTime, 'HH:MM:SS.FFF') - ...
          datenum('00:00:00.000', 'HH:MM:SS.FFF');

answered Feb 20 '13 at 14:40

jdl

6,151
19
83
132

Matlab: slow with converting date strings to datenum?

4 Answers4