Efficiently convert cell to double

Hi, I have a cell array of strings ("position") and need to convert the first cell column into a vector of doubles ("pointX"). The following loop works as intended, but takes a very long time. Is there a more efficient way to accomplish the same goal? Thank you.

    parfor j = 1:length(position)
       pointX(j,1) = str2num(cell2mat(position{j,1}(1)));
    end

[Jan]

If "position" is really a cell string, "cell2mat(position{j,1}(1))" fails, because "position{j,1}(1)" is the first character of the first string, such that CELL2MAT does not get a cell as input.

[John]

The variable "position" is actually a cell array, having dimension 6000 x 1 cell. Each cell in that column is a 3x1 consisting of 3 strings that give x, y, z coordinate strings. What I need to do is step into position{1,1} which is then a 3x1 cell array and convert the entire string in (1,1) to a double. Then repeat down the position column until position{6000,1}(1). The desired result is a 6000x1 vector of doubles that represents all the X points.

[Oleg Komarov]

Shouldn't you get then a 6000 by 3 matrix where the columns are respectively the x,y,z?

[John]

What I would really like to do is:

pointX = str2double(position{:,1}(1));

but it gives me a "bad cell reference" error.

[Oleg Komarov]

You cannot do that.

You can try to change the problem at the core, e.g. by importing a file in a different way or by changing the routine that generates the cell, or you have to manipulate the cell array.

[Jan]

@John, reading the textual description is still complictaed. It would reduce the chance of misunderstanding, if you post the Matlab code to create the test data and the wanted result, e.g.

position = {{'21', '34.5', '-17'}, {'324', '1e-19', '15'}}

result = [21, 324]

3 contributors tried to guess as good as they can, but there is a large chance, that we have wasted half an hour, because we did not understand the problem completely.

Answer 1

[Jan]

[EDITED: THIS DOES NOT MEET YOUR DEMANDS] Sorry, due to a missing example, I've overseen the fact, that you need the first column only. Before I guess to much around, could you please add some example values by editing the original question?

str2double is slow. sscanf is fast:

c = cell(1, 10000);  % Create test data
for i = 1:numel(c)
  c{i} = sprintf('%d', floor(rand*1000));
end

tic;
v1 = str2double(c);
toc

tic;
S = sprintf('%s*', c{:});
v2 = sscanf(S, '%d*');
toc

 Elapsed time is 0.411949 seconds.      (Matlab 2009a/64 Win7.)
 Elapsed time is 0.013777 seconds.

But sprintf cannot pre-allocate sufficiently, such that a dedicated MEX-function for concatenating cell elements beat this, see FEX: CStr2String:

tic;
S = CStr2String(c, '*');
v2 = sscanf(S, '%d*');
toc

 Elapsed time is 0.005835 seconds.

71 times faster. To my deep surprise calling SSCANF inside the MEX directly is not faster, although it avoids the expensive allocation of the large string, see e.g. FEX: String to double. Therefore I'd rely either on the pure Matlab sscanf(sprint()) or use CStr2String in addition.

[MUKESH KUMAR]

how it can be solved........

[dzid_]

This only works with integers

[madhan ravi]

@dzid: read about sscanf() in the documentation

[MacMilan ZHANG]

Amazing speed up sscanf

Answer 2

[Sean de Wolski]

str2double() will be faster than str2num()

Also, could probably skip the parfor loop altogether and just call str2double() on the cell array of strings:

str2double({'3';'45'})

Answer 3

[Oleg Komarov]

To address the OP comments:

Suppose you have this fake input (6000 by 1 cell array where each cell is a 1 by 3 cellstring):

position = repmat({{'23','43','120'}},6000,1);

tic
% Manipulate into 6000 by 3
position = cat(1,position{:,1});
% Use str2double
position = str2double(position);
toc

Elapsed time is 0.754681 seconds.

If you want to achieve results faster, consider Jan's suggestions.

[Jan]

I guess, that this is wanted:

positionM = cat(1, position{:, 1});

Btw, the JIT profits from variable not changing their type. Therefore I avoid reusing of names in general, but this can increase the memory footprint.

[Sean de Wolski]

@Jan:

function jitxtype

t1 = 0;

t2 = 0;

for ii = 1:1000;

A = rand(1000);

tic

B = uint8(A);

t1 = t1+toc;

tic

A = uint8(A);

t2 = t2+toc;

end

[t1 t2]

That is signifanct.

[Jan]

@Sean: Please add the timings you get and the Matlab version. I'm not at my Matlab machine currently.

[Oleg Komarov]

win64 R2012a

ans =

0.0054 5.7938

I also added pause(0.01) before each tic.

[Jan]

I'd dare to call it "significant" also.

[Daniel Shub]

With Oleg's pause on 64-bit Linux with R2011a I get:

0.6394 6.1356