I have a question.
Suppose I have matrix
A =
1 2 3
4 5 6
7 8 9
10 11 12
I need to select n rolling rows from A and transpose elements in new matrix C in rows.
The loop that I use is:
n = 3; %for instance every 3 rows of A
B = [];
for i = 1:n
Btemp = transpose(A(i:i+size(A,1)-n,:));
B = [B;Btemp];
end
C=B';
and that produces matrix C which is:
C =
1 2 3 4 5 6 7 8 9
4 5 6 7 8 9 10 11 12
This is what i want too do, but can I do the same job without the loop?
It takes 4 minutes to calculate for an A matrix of 3280x35 size.
I think you can make it work very fast if you make initialization. And one other trick is to take the transpose first, since MATLAB uses columns as first index instead of rows.
tic
A = reshape(1:3280*35,[3280 35])'; %# Generate an example A
[nRows, nCols] = size(A);
n = 3; %for instance every 3 rows of A
B = zeros(nRows-n+1,nCols*n);
At = A';
for i = 1:size(B,1)
B(i,:) = reshape(At(:,i:i+n-1), [1 nCols*n]);
end
toc
The elapsed time is
Elapsed time is 0.004059 seconds.
I would not use reshape in the loop, but transform A first to one single row (actually a column will also work, doesn't matter)
Ar = reshape(A',1,[]); % the ' is important here!
then the selecting of elements out of Ar is really simple:
[nrows, ncols] = size(A);
new_ncols = ncols*n;
B = zeros(nrows-(n-1),new_ncols);
for ii = 1:nrows-(n-1)
B(ii,:) = Ar(n*(ii-1)+(1:new_ncols));
end
Still, the preallocation of B, gives you the largest improvement: more info at http://www.mathworks.nl/help/techdoc/matlab_prog/f8-784135.html
I don't have Matlab on me right now but I think you can do this without loops like this:
reshape(permute(cat(A(1:end-1,:),A(2:end,:),3),[3,2,1]), [2, size(A,2)*(size(A,1) - 1)]);
and in fact won't this do what you want?:
A1 = A(1:end-1,:);
A2 = A(2:end,:);
answer = [A1(:) ; A2(:)]
Related
Say I have a vector Q = [Q1 Q2 .... QN].
I would like to create a matrix A such that the kth "ring" of the matrix is equal to Qk, with the following constraint:
if N is odd, the central patch is composed of one number, which is QN
For Q = [12 3 27] this would be :
A =
12 12 12 12 12
12 3 3 3 12
12 3 27 3 12
12 3 3 3 12
12 12 12 12 12
if N is even, the central patch is a 2x2 patch where QN gets repeated
for Q = [12 3] this would be
A =
12 12 12 12
12 3 3 12
12 3 3 12
12 12 12 12
Two for loops
Two for loops work but it is too slow (~13,3s for 5000x5000 matrices) (Code below) :
%% Two for loops :
% Generate random integer vector Q with unique values
N = 5;
n = 15 * N;
Q = randperm(n,N).';
% Double for loop method
if mod(N,2)==1
mSize = 2*N-1;
else
mSize = 2*N;
end
A = zeros(mSize);
for ii=1:(mSize)
for jj=1:(mSize)
IDX = min([ii,jj,mSize-ii+1,mSize-jj+1]);
A(ii,jj) = Q(IDX);
end
end
Faster approach
I have found a faster approach, which is pretty good (~1.46s for 5000x5000 matrices) but there might still be some room for improvement :
if mod(N,2)==1
mSize = 2*N-1;
I_idx = (1:mSize)-N;
A_fast = Q(end-max(abs(I_idx.'),abs(I_idx)));
else
I_idx = [(N-1):-1:0 0:(N-1)];
A_fast = Q(end-max(I_idx.',I_idx));
end
Any ideas?
The logic of the code is lightly simpler if you follow the advice in Wolfie's comment, and compute only one quadrant that you repeat:
I_idx = 1:N;
B = Q(min(I_idx,I_idx.'));
if mod(N,2)==1
B = [B,B(:,end-1:-1:1)]; % same as [B,fliplr(B(:,1:end-1))]
B = [B;B(end-1:-1:1,:)]; % same as [B;flipud(B(1:end-1,:))]
else
B = [B,fliplr(B)];
B = [B;flipud(B)];
end
This is 2-2.5 times as fast depending on whether Q is even or odd-sized.
Steve's comment suggest building a triangle first, but I don't see that being any faster, due to the complexity of indexing a matrix's upper or lower triangle.
Testing code:
N = 5000;
n = 15 * N;
Q = randperm(n,N).';
tic
if mod(N,2)==1
mSize = 2*N-1;
I_idx = (1:mSize)-N;
A = Q(end-max(abs(I_idx.'),abs(I_idx)));
else
I_idx = [(N-1):-1:0 0:(N-1)];
A = Q(end-max(I_idx.',I_idx));
end
toc
tic
I_idx = 1:N;
B = Q(min(I_idx,I_idx.'));
if mod(N,2)==1
B = [B,B(:,end-1:-1:1)];
B = [B;B(end-1:-1:1,:)];
else
B = [B,fliplr(B)];
B = [B;flipud(B)];
end
toc
isequal(A,B)
I came up with a solution using repmat, then flipping along a diagonal to get a quarter of the solution, finally flipping and reversing twice to get the full output matrix.
function A = flip_it_and_reverse_it(Q)
N = length(Q);
QQ = repmat(Q(:), 1, N);
quarter_A = triu(QQ) + triu(QQ, 1).';
half_A = [quarter_A, quarter_A(:, end-1:-1:1)];
A = [half_A; half_A(end-1:-1:1, :)];
end
There may be improvements that can be made to get faster flips/reverses with some clever transposes.
For the even case in your updated question, the indices in the lines starting half_A and A should be end:-1:1 instead of end-1:-1:1.
Running some quick timings, it looks like my solution is comprable (sometimes slightly slower) to your faster approach:
N = 5000;
n = 15 * N;
Q = randperm(n,N).';
disp('double loop')
tic
double_loop(Q);
disp(toc)
disp('faster approach')
tic
faster_approach(Q);
disp(toc)
disp('flip_it_and_reverse_it')
tic
flip_it_and_reverse_it(Q);
disp(toc)
Results:
double loop
14.4767
faster approach
1.8137
flip_it_and_reverse_it
1.6556
Note: sometimes faster_approach wins, sometimes flip - I've got some other jobs running on my laptop.
I have an NxM matrix in MATLAB that I would like to reorder in similar fashion to the way JPEG reorders its subblock pixels:
(image from Wikipedia)
I would like the algorithm to be generic such that I can pass in a 2D matrix with any dimensions. I am a C++ programmer by trade and am very tempted to write an old school loop to accomplish this, but I suspect there is a better way to do it in MATLAB.
I'd be rather want an algorithm that worked on an NxN matrix and go from there.
Example:
1 2 3
4 5 6 --> 1 2 4 7 5 3 6 8 9
7 8 9
Consider the code:
M = randi(100, [3 4]); %# input matrix
ind = reshape(1:numel(M), size(M)); %# indices of elements
ind = fliplr( spdiags( fliplr(ind) ) ); %# get the anti-diagonals
ind(:,1:2:end) = flipud( ind(:,1:2:end) ); %# reverse order of odd columns
ind(ind==0) = []; %# keep non-zero indices
M(ind) %# get elements in zigzag order
An example with a 4x4 matrix:
» M
M =
17 35 26 96
12 59 51 55
50 23 70 14
96 76 90 15
» M(ind)
ans =
17 35 12 50 59 26 96 51 23 96 76 70 55 14 90 15
and an example with a non-square matrix:
M =
69 9 16 100
75 23 83 8
46 92 54 45
ans =
69 9 75 46 23 16 100 83 92 54 8 45
This approach is pretty fast:
X = randn(500,2000); %// example input matrix
[r, c] = size(X);
M = bsxfun(#plus, (1:r).', 0:c-1);
M = M + bsxfun(#times, (1:r).'/(r+c), (-1).^M);
[~, ind] = sort(M(:));
y = X(ind).'; %'// output row vector
Benchmarking
The following code compares running time with that of Amro's excellent answer, using timeit. It tests different combinations of matrix size (number of entries) and matrix shape (number of rows to number of columns ratio).
%// Amro's approach
function y = zigzag_Amro(M)
ind = reshape(1:numel(M), size(M));
ind = fliplr( spdiags( fliplr(ind) ) );
ind(:,1:2:end) = flipud( ind(:,1:2:end) );
ind(ind==0) = [];
y = M(ind);
%// Luis' approach
function y = zigzag_Luis(X)
[r, c] = size(X);
M = bsxfun(#plus, (1:r).', 0:c-1);
M = M + bsxfun(#times, (1:r).'/(r+c), (-1).^M);
[~, ind] = sort(M(:));
y = X(ind).';
%// Benchmarking code:
S = [10 30 100 300 1000 3000]; %// reference to generate matrix size
f = [1 1]; %// number of cols is S*f(1); number of rows is S*f(2)
%// f = [0.5 2]; %// plotted with '--'
%// f = [2 0.5]; %// plotted with ':'
t_Amro = NaN(size(S));
t_Luis = NaN(size(S));
for n = 1:numel(S)
X = rand(f(1)*S(n), f(2)*S(n));
f_Amro = #() zigzag_Amro(X);
f_Luis = #() zigzag_Luis(X);
t_Amro(n) = timeit(f_Amro);
t_Luis(n) = timeit(f_Luis);
end
loglog(S.^2*prod(f), t_Amro, '.b-');
hold on
loglog(S.^2*prod(f), t_Luis, '.r-');
xlabel('number of matrix entries')
ylabel('time')
The figure below has been obtained with Matlab R2014b on Windows 7 64 bits. Results in R2010b are very similar. It is seen that the new approach reduces running time by a factor between 2.5 (for small matrices) and 1.4 (for large matrices). Results are seen to be almost insensitive to matrix shape, given a total number of entries.
Here's a non-loop solution zig_zag.m. It looks ugly but it works!:
function [M,index] = zig_zag(M)
[r,c] = size(M);
checker = rem(hankel(1:r,r-1+(1:c)),2);
[rEven,cEven] = find(checker);
[cOdd,rOdd] = find(~checker.'); %'#
rTotal = [rEven; rOdd];
cTotal = [cEven; cOdd];
[junk,sortIndex] = sort(rTotal+cTotal);
rSort = rTotal(sortIndex);
cSort = cTotal(sortIndex);
index = sub2ind([r c],rSort,cSort);
M = M(index);
end
And a test matrix:
>> M = [magic(4) zeros(4,1)];
M =
16 2 3 13 0
5 11 10 8 0
9 7 6 12 0
4 14 15 1 0
>> newM = zig_zag(M) %# Zig-zag sampled elements
newM =
16
2
5
9
11
3
13
10
7
4
14
6
8
0
0
12
15
1
0
0
Here's a way how to do this. Basically, your array is a hankel matrix plus vectors of 1:m, where m is the number of elements in each diagonal. Maybe someone else has a neat idea on how to create the diagonal arrays that have to be added to the flipped hankel array without a loop.
I think this should be generalizeable to a non-square array.
% for a 3x3 array
n=3;
numElementsPerDiagonal = [1:n,n-1:-1:1];
hadaRC = cumsum([0,numElementsPerDiagonal(1:end-1)]);
array2add = fliplr(hankel(hadaRC(1:n),hadaRC(end-n+1:n)));
% loop through the hankel array and add numbers counting either up or down
% if they are even or odd
for d = 1:(2*n-1)
if floor(d/2)==d/2
% even, count down
array2add = array2add + diag(1:numElementsPerDiagonal(d),d-n);
else
% odd, count up
array2add = array2add + diag(numElementsPerDiagonal(d):-1:1,d-n);
end
end
% now flip to get the result
indexMatrix = fliplr(array2add)
result =
1 2 6
3 5 7
4 8 9
Afterward, you just call reshape(image(indexMatrix),[],1) to get the vector of reordered elements.
EDIT
Ok, from your comment it looks like you need to use sort like Marc suggested.
indexMatrixT = indexMatrix'; % ' SO formatting
[dummy,sortedIdx] = sort(indexMatrixT(:));
sortedIdx =
1 2 4 7 5 3 6 8 9
Note that you'd need to transpose your input matrix first before you index, because Matlab counts first down, then right.
Assuming X to be the input 2D matrix and that is square or landscape-shaped, this seems to be pretty efficient -
[m,n] = size(X);
nlim = m*n;
n = n+mod(n-m,2);
mask = bsxfun(#le,[1:m]',[n:-1:1]);
start_vec = m:m-1:m*(m-1)+1;
a = bsxfun(#plus,start_vec',[0:n-1]*m);
offset_startcol = 2- mod(m+1,2);
[~,idx] = min(mask,[],1);
idx = idx - 1;
idx(idx==0) = m;
end_ind = a([0:n-1]*m + idx);
offsets = a(1,offset_startcol:2:end) + end_ind(offset_startcol:2:end);
a(:,offset_startcol:2:end) = bsxfun(#minus,offsets,a(:,offset_startcol:2:end));
out = a(mask);
out2 = m*n+1 - out(end:-1:1+m*(n-m+1));
result = X([out2 ; out(out<=nlim)]);
Quick runtime tests against Luis's approach -
Datasize: 500 x 2000
------------------------------------- With Proposed Approach
Elapsed time is 0.037145 seconds.
------------------------------------- With Luis Approach
Elapsed time is 0.045900 seconds.
Datasize: 5000 x 20000
------------------------------------- With Proposed Approach
Elapsed time is 3.947325 seconds.
------------------------------------- With Luis Approach
Elapsed time is 6.370463 seconds.
Let's assume for a moment that you have a 2-D matrix that's the same size as your image specifying the correct index. Call this array idx; then the matlab commands to reorder your image would be
[~,I] = sort (idx(:)); %sort the 1D indices of the image into ascending order according to idx
reorderedim = im(I);
I don't see an obvious solution to generate idx without using for loops or recursion, but I'll think some more.
I have a problem with sorting some finance data based on firmnumbers. So given is a matrix that looks like:
[1 3 4 7;
1 2 7 8;
2 3 7 8;]
On Matlab i would like the matrix to be sorted as follows:
[1 0 3 4 7 0;
1 2 0 0 7 8;
0 2 3 0 7 8;]
So basically every column needs to consist of 1 type of number.
I have tried many things but i cant get the matrix sorted properly.
A = [1 3 4 7;
1 2 7 8;
2 3 7 8;]
%// Get a unique list of numbers in the order that you want them to appear as the new columns
U = unique(A(:))'
%'//For each column (of your output, same as columns of U), find which rows have that number. Do this by making A 3D so that bsxfun compares each element with each element
temp1 = bsxfun(#eq,permute(A,[1,3,2]),U)
%// Consolidate this into a boolean matrix with the right dimensions and 1 where you'll have a number in your final answer
temp2 = any(temp1,3)
%// Finally multiply each line with U
bsxfun(#times, temp2, U)
So you can do that all in one line but I broke it up to make it easier to understand. I suggest you run each line and look at the output to see how it works. It might seem complicated but it's worthwhile getting to understand bsxfun as it's a really useful function. The first use which also uses permute is a bit more tricky so I suggest you first make sure you understand that last line and then work backwards.
What you are asking can also be seen as an histogram
A = [1 3 4 7;
1 2 7 8;
2 3 7 8;]
uniquevalues = unique(A(:))
N = histc(A,uniquevalues' ,2) %//'
B = bsxfun(#times,N,uniquevalues') %//'
%// bsxfun can replace the following instructions:
%//(the instructions are equivalent only when each value appears only once per row )
%// B = repmat(uniquevalues', size(A,1),1)
%// B(N==0) = 0
Answer without assumptions - Simplified
I did not feel comfortable with my old answer that makes the assumption of everything being an integer and removed the possibility of duplicates, so I came up with a different solution based on #lib's suggestion of using a histogram and counting method.
The only case I can see this not working for is if a 0 is entered. you will end up with a column of all zeros, which one might interpret as all rows initially containing a zero, but that would be incorrect. you could uses nan instead of zeros in that case, but not sure what this data is being put into, and if it that processing would freak out.
EDITED
Includes sorting of secondary matrix, B, along with A.
A = [-1 3 4 7 9; 0 2 2 7 8.2; 2 3 5 9 8];
B = [5 4 3 2 1; 1 2 3 4 5; 10 9 8 7 6];
keys = unique(A);
[counts,bin] = histc(A,transpose(unique(A)),2);
A_sorted = cell(size(A,1),1);
for ii = 1:size(A,1)
for jj = 1:numel(keys)
temp = zeros(1,max(counts(:,jj)));
temp(1:counts(ii,jj)) = keys(jj);
A_sorted{ii} = [A_sorted{ii},temp];
end
end
A_sorted = cell2mat(A_sorted);
B_sorted = nan(size(A_sorted));
for ii = 1:size(bin,1)
for jj = 1:size(bin,2)
idx = bin(ii,jj);
while ~isnan(B_sorted(ii,idx))
idx = idx+1;
end
B_sorted(ii,idx) = B(ii,jj);
end
end
B_sorted(isnan(B_sorted)) = 0
You can create at the beginning a matrix with 9 columns , and treat the values in your original matrix as column indexes.
A = [1 3 4 7;
1 2 7 8;
2 3 7 8;]
B = zeros(3,max(A(:)))
for i = 1:size(A,1)
B(i,A(i,:)) = A(i,:)
end
B(:,~any(B,1)) = []
Suppose I have an array X of size n by p by q. I would like to reshape it as a matrix with p rows, and in each row put the concatenation of the n rows of size q, resulting in a matrix of size p by nq.
I managed to do it with a loop but it takes a while say if n=1000, p=300, q=300.
F0=[];
for k=1:size(F,1)
F0=[F0,squeeze(X(k,:,:))];
end
Is there a faster way?
I think this is what you want:
Y = reshape(permute(X, [2 1 3]), size(X,2), []);
Example with n=2, p=3, q=4:
>> X
X(:,:,1) =
0 6 9
8 3 0
X(:,:,2) =
4 7 1
3 7 4
X(:,:,3) =
4 7 2
6 7 6
X(:,:,4) =
6 1 9
1 4 3
>> Y = reshape(permute(X, [2 1 3]), size(X,2), [])
Y =
0 8 4 3 4 6 6 1
6 3 7 7 7 7 1 4
9 0 1 4 2 6 9 3
Try this -
reshape(permute(X,[2 3 1]),p,[])
Thus, for code verification, one can look into a sample case run -
n = 2;
p = 3;
q = 4;
X = rand(n,p,q)
F0=[];
for k=1:n
F0=[F0,squeeze(X(k,:,:))];
end
F0
F0_noloop = reshape(permute(X,[2 3 1]),p,[])
Output is -
F0 =
0.4134 0.6938 0.3782 0.4775 0.2177 0.0098 0.7043 0.6237
0.1257 0.8432 0.7295 0.2364 0.3089 0.9223 0.2243 0.1771
0.7261 0.7710 0.2691 0.8296 0.7829 0.0427 0.6730 0.7669
F0_noloop =
0.4134 0.6938 0.3782 0.4775 0.2177 0.0098 0.7043 0.6237
0.1257 0.8432 0.7295 0.2364 0.3089 0.9223 0.2243 0.1771
0.7261 0.7710 0.2691 0.8296 0.7829 0.0427 0.6730 0.7669
Rather than using vectorization to solve the problem, you could look at the code to try and figure out what may improve performance. In this case, since you know the size of your output matrix F0 should be px(n*q), you could pre-allocate memory to F0 and avoid the constant resizing of the matrix at each iteration of the for loop
n=1000;
p=300;
q=300;
F0=zeros(p,n*q);
for k=1:size(F,1)
F0(:,(k-1)*q+1:k*q) = squeeze(F(k,:,:));
end
While probably not as efficient as the other two solutions, it is an alternative. Try the above and see what happens!
Assume matrix M:
1 2 3
3 5 6
6 8 9
How do I store I extract the following row vector a from it?
1
5
9
You just need to use diag:
octave-3.4.0:1> A = [ 1 2 3; 3 5 6; 6 8 9 ]
A =
1 2 3
3 5 6
6 8 9
octave-3.4.0:2> D = diag(A)
D =
1
5
9
Note that you can also extract other diagonals by passing a second parameter to diag, e.g.
octave-3.4.0:3> D = diag(A, 1)
D =
2
6
octave-3.4.0:4> D = diag(A, -1)
D =
3
8
If you know the dimensions of your matrix (square or otherwise), you can extract any diagonal you like, or even modified diagonals (such as numbers in (1,1), (2,3), (3,5), etc), somewhat faster than using diag, by simply using an index call like this:
a=M(1:4:9)
(note: this produces a row vector; for a column vector, just transpose) For an NxN matrix, simply start at the desired value (1 for the top-left corner, 2 for next one down vertically, and so on), then increment by N+1 until you reach the appropriate value.
octave:35> tic; for i=1:10000 diag(rand(3)); end; toc;
Elapsed time is 0.13973 seconds.
octave:36> tic; for i=1:10000 rand(3)(1:4:9); end; toc;
Elapsed time is 0.10966 seconds.
For reference:
octave:49> tic; for i=1:10000 rand(3); end; toc;
Elapsed time is 0.082429 seconds.
octave:107> version
ans = 3.6.3
So the overhead for the for loop and the rand function, subtracted off, shows that using indices is about twice as fast as using diag. I suspect that this is purely due to the overhead of calling diag, as the operation itself is very straightforward and fast, and is almost certainly how diag itself works.