问题
Matlab/Octave algorithm example:
input vector: [ 1 0 2 0 7 7 7 0 5 0 0 0 9 ]
output vector: [ 1 1 2 2 7 7 7 7 5 5 5 5 9 ]
The algorithm is very simple: it goes through the vector and replaces all zeros with the last non-zero value. It seems trivial, and is so when done with a slow for (i=1:length) loop and being able to refer to the previous element (i-1), but looks impossible to be formulated in the fast vectorized form. I tried the merge() and shift() but it only works for the first occurrence of zero, not an arbitrary number of them.
Can it be done in a vectorized form in Octave/Matlab or must C be used for this to have sufficient performance on big amount of data?
I have another similar slow for-loop algorithm to speed up and it seems generally impossible to refer to previous values in a vectorized form, like an SQL lag() or group by or loop (i-1) would easily do. But Octave/Matlab loops are terribly slow.
Has anyone found a solution to this general problem or is this futile for fundamental Octave/Matlab design reasons?
Performance benchmark:
SOLUTION 1 (slow loop)
in = repmat([ 1 0 2 0 7 7 7 0 5 0 0 0 9 ] ,1 ,100000);
out = in;
tic
for i=2:length(out)
if (out(i)==0)
out(i)=out(i-1);
end
end
toc
[in(1:20); out(1:20)] % test to show side by side if ok
Elapsed time is 15.047 seconds.
SOLUTION 2 by Dan (~80 times faster)
in = V = repmat([ 1 0 2 0 7 7 7 0 5 0 0 0 9 ] ,1 ,100000);
tic;
d = double(diff([0,V])>0);
d(find(d(2:end))+1) = find(diff([0,~V])==-1) - find(diff([0,~V])==1);
out = V(cumsum(~~V+d)-1);
toc;
[in(1:20); out(1:20)] % shows it works ok
Elapsed time is 0.188167 seconds.
15.047 / 0.188167 = 79.97 times improvement
SOLUTION 3 by GameOfThrows (~115 times faster)
in = repmat([ 1 0 2 0 7 7 7 0 5 0 0 0 9 ] ,1 ,100000);
a = in;
tic;
pada = [a,888];
b = pada(pada >0);
bb = b(:,1:end-1);
c = find (pada==0);
d = find(pada>0);
len = d(2:end) - (d(1:end-1));
t = accumarray(cumsum([1,len])',1);
out = bb(cumsum(t(1:end-1)));
toc;
Elapsed time is 0.130558 seconds.
15.047 / 0.130558 = 115.25 times improvement
Magical SOLUTION 4 by Luis Mendo (~250 times faster)
in = repmat([ 1 0 2 0 7 7 7 0 5 0 0 0 9 ] , 1, 100000);
tic;
u = nonzeros(in);
out = u(cumsum(in~=0)).';
toc;
Elapsed time is 0.0597501 seconds.
15.047 / 0.0597501 = 251.83 times improvement
(Update 2019/03/13) Timings with MATLAB R2017a:
Slow loop: 0.010862 seconds.
Dan: 0.072561 seconds.
GameOfThrows: 0.066282 seconds.
Luis Mendo: 0.032257 seconds.
fillmissing: 0.053366 seconds.
So we draw yet again the same conclusion: loops in MATLAB are no longer slow!
See also: Trivial/impossible algorithm challenge in Octave/Matlab Part II: iterations memory
回答1:
The following simple approach does what you want, and is probably very fast:
in = [1 0 2 0 7 7 7 0 5 0 0 0 9];
t = cumsum(in~=0);
u = nonzeros(in);
out = u(t).';
回答2:
I think it is possible, let's start with the basics, you want to capture where number is greater than 0:
a = [ 1 0 2 0 7 7 7 0 5 0 0 0 9 ] %//Load in Vector
pada = [a,888]; %//Pad A with a random number at the end to help in case the vector ends with a 0
b = pada(find(pada >0)); %//Find where number if bigger than 0
bb = b(:,1:end-1); %//numbers that are bigger than 0
c = find (pada==0); %//Index where numbers are 0
d = find(pada>0); %//Index where numbers are greater than 0
length = d(2:end) - (d(1:end-1)); %//calculate number of repeats needed for each 0 trailing gap.
%//R = [cell2mat(arrayfun(@(x,nx) repmat(x,1,nx), bb, length,'uniformoutput',0))]; %//Repeat the value
----------EDIT---------
%// Accumarray and cumsum method, although not as nice as Dan's 1 liner
t = accumarray(cumsum([1,length])',1);
R = bb(cumsum(t(1:end-1)));
NOTE: I used arrayfun, but you can use accumarray as well.I think this demonstrates that it is possible to do this in parallel?
R =
Columns 1 through 10
1 1 2 2 7 7 7 7 5 5
Columns 11 through 13
5 5 9
TESTs:
a = [ 1 0 2 0 7 7 7 0 5 0 0 0 9 0 0 0 ]
R =
Columns 1 through 10
1 1 2 2 7 7 7 7 5 5
Columns 11 through 16
5 5 9 9 9 9
PERFORMANCE:
a = repmat([ 1 0 2 0 7 7 7 0 5 0 0 0 9 ] ,1,10000); %//Double of 130,000
Arrayfun Method : Elapsed time is 6.840973 seconds.
AccumArray Method : Elapsed time is 2.097432 seconds.
回答3:
I think is a vectorized solution. Works on your example:
V = [1 0 2 0 7 7 7 0 5 0 0 0 9]
%// This is where the numbers you will repeat lie. You have to cast to a double otherwise later when you try assign numbers to it it caps them at logical 1s
d = double(diff([0,V])>0)
%// find(diff([0,~V])==-1) - find(diff([0,~V])==1) is the length of each zero cluster
d(find(d(2:end))+1) = find(diff([0,~V])==-1) - find(diff([0,~V])==1)
%// ~~V is the same as V ~= 0
V(cumsum(~~V+d)-1)
回答4:
Here is another solution, using linear interpolation with previous neighbor lookup.
I assume it to be quite fast as well, as there are just look-ups and indexing and no calculations:
in = [1 0 2 0 7 7 7 0 5 0 0 0 9]
mask = logical(in);
idx = 1:numel(in);
in(~mask) = interp1(idx(mask),in(mask),idx(~mask),'previous');
%// out = in
Explanation
You need to create an index vector:
idx = 1:numel(in) $// = 1 2 3 4 5 ...
And a logical mask, masking all your non-zero values:
mask = logical(in);
This way you get the grid points idx(mask) and grid data in(mask) for the interpolation. The query points idx(~mask) are indices of the zero data. The query data in(~mask) is then "calculated" by next previous neighbor interpolation, so it basically looks in the grid what is the value for the previous grid point. Exactly what you want. Unfortunately the involved functions have a huge overhead for all thinkable cases, thats why it is still slower than Luis Mendo's Answer, though there are no arithmetic calculations involved.
Furthermore one could reduce the overhead of interp1 a little:
F = griddedInterpolant(idx(mask),in(mask),'previous');
in(~mask) = F(idx(~mask));
But there is not too much effect.
in = %// = out
1 1 2 2 7 7 7 7 5 5 5 5 9
Benchmark
0.699347403200000 %// thewaywewalk
1.329058123200000 %// GameOfThrows
0.408333643200000 %// LuisMendo
1.585014923200000 %// Dan
Code
function [t] = bench()
in = repmat([ 1 0 2 0 7 7 7 0 5 0 0 0 9 ] ,1 ,100000);
% functions to compare
fcns = {
@() thewaywewalk(in);
@() GameOfThrows(in);
@() LuisMendo(in);
@() Dan(in);
};
% timeit
t = zeros(4,1);
for ii = 1:10;
t = t + cellfun(@timeit, fcns);
end
format long
end
function in = thewaywewalk(in)
mask = logical(in);
idx = 1:numel(in);
in(~mask) = interp1(idx(mask),in(mask),idx(~mask),'previous');
end
function out = GameOfThrows(a)
pada = [a,888];
b = pada(find(pada >0));
bb = b(:,1:end-1);
c = find (pada==0);
d = find(pada>0);
length = d(2:end) - (d(1:end-1));
t = accumarray(cumsum([1,length])',1);
out = bb(cumsum(t(1:end-1)));
end
function out = LuisMendo(in)
t = cumsum(in~=0);
u = nonzeros(in);
out = u(t).';
end
function out = Dan(V)
d = double(diff([0,V])>0);
d(find(d(2:end))+1) = find(diff([0,~V])==-1) - find(diff([0,~V])==1);
out = V(cumsum(~~V+d)-1);
end
回答5:
New in MATLAB R2016b: fillmissing, it does exactly as described in the question:
in = [ 1 0 2 0 7 7 7 0 5 0 0 0 9 ];
in(in==0) = NaN;
out = fillmissing(in,'previous');
[This new functionality discovered in this duplicate question].
回答6:
Vector operations generally assume independence of the individual items. If you have a dependence on an earlier item, then looping is the best way to do it.
Some extra background on matlab: In matlab the operations are typically faster not because of vector operations specifically, but because a vector operation simply does the loop in native C++ code instead of through the interpreter
来源：https://stackoverflow.com/questions/34041614/replace-all-zeros-in-vector-by-previous-non-zero-value