Is it possible to extract numbers from formated strings without a for cycle?

I have {'abc12', 'abc23', 'abc24', 'abc99'} and I need the vector [12,23,24,99]. How to do this without a for cycle?

Answer 1

[Stephen23]

Much faster than cellfun, str2double, or strrep, and with no explicit loop:

>> C = {'abc12', 'abc23', 'abc24', 'abc99'};
>> V = sscanf([C{:}],'%*3c%d')
V =
   12
   23
   24
   99
>>

[Jan]

+1. Very efficient.

If C is large (e.g. 5000 elements), the concatenation needs a lot of time. It seems like Matlab's horzcat has a problem with the pre-allocation. Using FEX: Cell2Vec and with the format string 'abc%d' the code is even two times faster. But if the OP needs to do this for short arrays strings, the time for compiling the fast C-Mex function might be wasted.

See timings in my answer.

Remark: What a pity, that Matlab's sprintfc is not documented and that there is no corresponding sscanfc.

[Stephen23]

@Jan Simon: thank you for your in-depth timing and investigation.

"with the format string 'abc%d' the code is even two times faster"

I thought this might be the case, but suspected (based on this question) that the user would not always want the same 'abc' characters being matched.

Answer 2

[Matt J]

No, it is not really possible to do this without a for-loop. The suggestions Star Strider and I have given you use str2double and/or cellfun, which have for-loops inside them.

If for-loops hidden in functions don't count for you, then okay, but you could just as easily write your own function to hide the loop.

[Jan]

@Matt J: strrep and str2double contain the same kind of for loops inside(!) also. I think "without for loop" is used as synonym for "fast". This might come from the times before the JIT acceleration was introduced (Matlab 6.5, July 2002).

[Matt J]

Hmmm. In R2017b, str2double is Mcoded but strrep is not, so maybe both have loops, but they can't be the "same kind".

[Jan]

@Matt J: Exactly, this is what I actually wanted to express. Mentioning strrep and str2double was thought to supplement your answer. They contain loops in M or C code level. As you wrote: "it is not really possible to do this without a for-loop". +1

Remark: Old versions of Matlab contained many C codes of the builtin functions, like cellfun.c or histc.c. Even for some P coded files the corresponding M files have been shipped. These source codes have been very good examples. What a pity, that they are not available in modern versions anymore.

Answer 3

[Jan]

With FEX: Cell2Vec:

V = sscanf(Cell2Vec(C), 'abc%d');

Some timings:

C = repmat({'abc12', 'abc23', 'abc24', 'abc99'}, 1, 1000);

tic;
for k = 1:100
  V = sscanf([C{:}],'%*3c%d');
end;
toc

tic;
for k = 1:100
  V = sscanf(Cell2Vec(C), '%*3c%d');
end;
toc

% While '%*3c%d' requires some work, 'abc%d' is cheaper:
tic;
for k = 1:100
  V = sscanf(Cell2Vec(C), 'abc%d');
end;
toc

tic;
for k = 1:100
  V = str2double(strrep(C, 'abc', ''));
end
toc

tic;
for k = 1:100
  V = cellfun(@str2double, regexp(C, '\d*', 'match'));
end
toc

Results:

 Elapsed time is 0.556899 seconds.        % sccanf([C{:}], '%*3c%d')
 Elapsed time is 0.311912 seconds.        % sccanf(Cell2Vec(C), '%*3c%d')
 Elapsed time is 0.254616 seconds.        % sccanf(Cell2Vec(C), 'abc%d')
 Elapsed time is 12.940544 seconds.       % str2double(strrep)
 Elapsed time is 22.966292 seconds.       % cellfun(@str2double, regexp)

Answer 4

[Matt J]

str2double(  strrep( {'abc12', 'abc23', 'abc24', 'abc99'},'abc','') )