If I have this vector:
x = [1 1 1 1 1 2 2 2 3 4 4 6 6 6 6]
I would like to get the position of each unique number according to itself.
y = [1 2 3 4 5 1 2 3 1 1 2 1 2 3 4]
At the moment I'm using:
y = sum(triu(x==x.')) % MATLAB 2016b and above
It's compact but obviously not memory efficient.
For the pure beauty of MATLAB programming I would avoid using a loop. Do you have a better simple implementation ?
Context:
My final goal is to sort the vector x but with the constraint that a number that appear N times has the priority over another number that has appeared more than N times:
[~,ind] = sort(y);
x_relative_sort = x(ind);
% x_relative_sort = 1 2 3 4 6 1 2 4 6 1 2 6 1 6 1
Assuming x is sorted, here's one vectorized alternative using unique, diff, and cumsum:
[~, index] = unique(x);
y = ones(size(x));
y(index(2:end)) = y(index(2:end))-diff(index).';
y = cumsum(y);
And now you can apply your final sorting:
>> [~, ind] = sort(y);
>> x_relative_sort = x(ind)
x_relative_sort =
1 2 3 4 6 1 2 4 6 1 2 6 1 6 1
If you have positive integers you can use sparse matrix:
[y ,~] = find(sort(sparse(1:numel(x), x, true), 1, 'descend'));
Likewise x_relative_sort can directly be computed:
[x_relative_sort ,~] = find(sort(sparse(x ,1:numel(x),true), 2, 'descend'));
Just for variety, here's a solution based on accumarray. It works for x sorted and containing positive integers, as in the question:
y = cell2mat(accumarray(x(:), x(:), [], #(t){1:numel(t)}).');
You can be more memory efficient by only comparing to unique(x), so you don't have a large N*N matrix but rather N*M, where N=numel(x), M=numel(unique(x)).
I've used an anonymous function syntax to avoid declaring an intermediate matrix variable, needed as it's used twice - this can probably be improved.
f = #(X) sum(cumsum(X,2).*X); y = f(unique(x).'==x);
Here's my solution that doesn't require sorting:
x = [1 1 1 1 1 2 2 2 3 4 4 6 6 6 6 1 1 1];
y = cell2mat( splitapply(#(v){cumsum(v)},x,cumsum(logical([1 diff(x)]))) ) ./ x;
Explanation:
% Turn each group new into a unique number:
t1 = cumsum(logical([1 diff(x)]));
% x = [1 1 1 1 1 2 2 2 3 4 4 6 6 6 6 1 1 1];
% t1 = [1 1 1 1 1 2 2 2 3 4 4 5 5 5 5 6 6 6];
% Apply cumsum separately to each group:
t2 = cell2mat( splitapply(#(v){cumsum(v)},x,t1) );
% t1 = [1 1 1 1 1 2 2 2 3 4 4 5 5 5 5 6 6 6];
% t2 = [1 2 3 4 5 2 4 6 3 4 8 6 12 18 24 1 2 3];
% Finally, divide by x to get the increasing values:
y = t2 ./ x;
% x = [1 1 1 1 1 2 2 2 3 4 4 6 6 6 6 1 1 1];
% t2 = [1 2 3 4 5 2 4 6 3 4 8 6 12 18 24 1 2 3];
a = [1 1 1 1 2 2 2 2 3 3 3 3; 1 2 3 4 5 6 7 8 9 10 11 12]';
What is the quickest way to sum the values in column 2 that correspond to the first and last occurrence of each number in column 1?
The desired output:
1 5
2 13
3 21
EDIT: The result should be the same if the numbers in column 1 are ordered differently.
a = [2 2 2 2 1 1 1 1 3 3 3 3; 1 2 3 4 5 6 7 8 9 10 11 12]';
2 5
1 13
3 21
You can use accumarray as follows. Not sure how fast it is, especially because it uses a custom anonymous function:
[u, ~, v] = unique(a(:,1), 'stable');
s = accumarray(v, a(:,2), [], #(x) x(1)+x(end));
result = [u s];
If the values in the first column of a are always in contiguous groups, the following approach can be used as well:
ind_diff = find(diff(a(:,1))~=0);
ind_first = [1; ind_diff+1];
ind_last = [ind_diff; size(a,1)];
s = a(ind_first,2) + a(ind_last,2);
result = [unique(a(:,1), 'stable') s];
As the title says, I want to find all rows in a Matlab matrix that in certain columns the values in the row are equal with the values in the previous row, or in general, equal in some row in the matrix. For example I have a matrix
1 2 3 4
1 2 8 10
4 5 7 9
2 3 6 4
1 2 4 7
and I want to find the following rows:
1 2 3 4
1 2 3 10
1 2 4 7
How do I do something like that and how do I do it generally for all the possible pairs in columns 1 and 2, and have equal values in previous rows, that exist in the matrix?
Here's a start to see if we're headed in the right direction:
>> M = [1 2 3 4;
1 2 8 10;
4 5 7 9;
2 3 6 4;
1 2 4 7];
>> N = M; %// copy M into a new matrix so we can modify it
>> idx = ismember(N(:,1:2), N(1,1:2), 'rows')
idx =
1
1
0
0
1
>> N(idx, :)
ans =
1 2 3 4
1 2 8 10
1 2 4 7
Then you can remove those rows from the original matrix and repeat.
>> N = N(~idx,:)
N =
4 5 7 9
2 3 6 4
this will give you the results
data1 =[1 2 3 4
1 2 8 10
4 5 7 9
2 3 6 4
1 2 4 7];
data2 = [1 2 3 4
1 2 3 10
1 2 4 7];
[exists,position] = ismember(data1,data2, 'rows')
where the exists vector tells you wheter the row is on the other matrix and position gives you the position...
a less elegant and simpler version would be
array_data1 = reshape (data1',[],1);
array_data2 = reshape (data2',[],1);
matchmatrix = zeros(size(data2,1),size(data1,1));
for irow1 = 1: size(data2,1)
for irow2 = 1: size(data1,1)
matchmatrix(irow1,irow2) = min(data2(irow1,:) == data1(irow2,:))~= 0;
end
end
the matchmatrix is to read as a connectivity matrix where value of 1 indicates which row of data1 matches with which row of data2
A = [1,4,2,5,10
2,4,5,6,2
2,1,5,6,10
2,3,5,4,2]
And I want split it into two matrix by the last column
A ->B and C
B = [1,4,2,5,10
2,1,5,6,10]
C = [2,4,5,6,2
2,3,5,4,2]
Also, this method could be applied to a big matrix, like matrix 100*22 according to the last column value into 9 groups by matlab.
Use logical indexing
B=A(A(:,end)==10,:);
C=A(A(:,end)==2,:);
returns
>> B
B =
1 4 2 5 10
2 1 5 6 10
>> C
C =
2 4 5 6 2
2 3 5 4 2
EDIT: In reply to Dan's comment here is the extension for general case
e = unique(A(:,end));
B = cell(size(e));
for k = 1:numel(e)
B{k} = A(A(:,end)==e(k),:);
end
or more compact way
B=arrayfun(#(x) A(A(:,end)==x,:), unique(A(:,end)), 'UniformOutput', false);
so for
A =
1 4 2 5 10
2 4 5 6 2
2 1 5 6 10
2 3 5 4 2
0 3 1 4 9
1 3 4 5 1
1 0 4 5 9
1 2 4 3 1
you get the matrices in elements of cell array B
>> B{1}
ans =
1 3 4 5 1
1 2 4 3 1
>> B{2}
ans =
2 4 5 6 2
2 3 5 4 2
>> B{3}
ans =
0 3 1 4 9
1 0 4 5 9
>> B{4}
ans =
1 4 2 5 10
2 1 5 6 10
Here is a general approach which will work on any number of numbers in the last column on any sized matrix:
A = [1,4,2,5,10
2,4,5,6,2
1,1,1,1,1
2,1,5,6,10
2,3,5,4,2
0,0,0,0,2];
First sort by the last column (many ways to do this, don't know if this is the best or not)
[~, order] = sort(A(:,end));
As = A(order,:);
Then create a vector of how many rows of the same number appear in that last col (i.e. how many rows per group)
rowDist = diff(find([1; diff(As(:, end)); 1]));
Note that for my example data rowDist will equal [1 3 2] as there is 1 1, 3 2s and 2 10s.
Now use mat2cell to split by these row groupings:
Ac = mat2cell(As, rowDist);
If you really want to you can now split it into separate matrices (but I doubt you would)
Ac{:}
results in
ans =
1 1 1 1 1
ans =
0 0 0 0 2
2 3 5 4 2
2 4 5 6 2
ans =
1 4 2 5 10
2 1 5 6 10
But I think you would find Ac itself more useful
EDIT:
Many solutions so might as well do a time comparison:
A = [...
1 4 2 5 10
2 4 5 6 2
2 1 5 6 10
2 3 5 4 2
0 3 1 4 9
1 3 4 5 3
1 0 4 5 9
1 2 4 3 1];
A = repmat(A, 1000, 1);
tic
for l = 1:100
[~, y] = sort(A(:,end));
As = A(y,:);
rowDist = diff(find([1; diff(As(:, end)); 1]));
Ac = mat2cell(As, rowDist);
end
toc
tic
for l = 1:100
D=arrayfun(#(x) A(A(:,end)==x,:), unique(A(:,end)), 'UniformOutput', false);
end
toc
tic
for l = 1:100
for k = 1:numel(e)
B{k} = A(A(:,end)==e(k),:);
end
end
toc
tic
for l = 1:100
Bb = sort(A(:,end));
[~,b] = histc(A(:,end), Bb([diff(Bb)>0;true]));
C = accumarray(b, (1:size(A,1))', [], #(r) {A(r,:)} );
end
toc
resulted in
Elapsed time is 0.053452 seconds.
Elapsed time is 0.17017 seconds.
Elapsed time is 0.004081 seconds.
Elapsed time is 0.22069 seconds.
So for even for a large matrix the loop method is still the fastest.
Use accumarray in combination with histc:
% Example data (from Mohsen Nosratinia)
A = [...
1 4 2 5 10
2 4 5 6 2
2 1 5 6 10
2 3 5 4 2
0 3 1 4 9
1 3 4 5 1
1 0 4 5 9
1 2 4 3 1];
% Get the proper indices to the specific rows
B = sort(A(:,end));
[~,b] = histc(A(:,end), B([diff(B)>0;true]));
% Collect all specific rows in their specific groups
C = accumarray(b, (1:size(A,1))', [], #(r) {A(r,:)} );
Results:
>> C{:}
ans =
1 3 4 5 1
1 2 4 3 1
ans =
2 3 5 4 2
2 4 5 6 2
ans =
0 3 1 4 9
1 0 4 5 9
ans =
2 1 5 6 10
1 4 2 5 10
Note that
B = sort(A(:,end));
[~,b] = histc(A(:,end), B([diff(B)>0;true]));
can also be written as
[~,b] = histc(A(:,end), unique(A(:,end)));
but unique is not built-in and is therefore likely to be slower, especially when this is all used in a loop.
Note also that the order of the rows has changed w.r.t. the order they had in the original matrix. If the order matters, you'll have to throw in another sort:
C = accumarray(b, (1:size(A,1))', [], #(r) {A(sort(r),:)} );
I have a large matrix (time x frequency), which I want to reduce partially. I want to sum every 1000 rows (time-samples) together keepinq the frequency information, it is kind of a segmentation.
Is there any way to do it without any cycle in MATLAB?
A smaller example:
M=[1 2 3; 2 3 4; 5 8 7; 5 6 7; 1 2 3; 1 2 4];
and I want to sum every 2 rows together so, that I get:
[3 5 7; 10 14 14; 2 4 7]
Suppose you have a matrix with N rows and M columns and you want to sum every R rows together (where N is divisible by R),
>> mat = [1 2 3; 2 3 4; 5 8 7; 5 6 7; 1 2 3; 1 2 4]
mat =
1 2 3
2 3 4
5 8 7
5 6 7
1 2 3
1 2 4
>> [N, M] = size(mat); %=> [6, 3]
>> R = 2;
The following will allow you to sum groups of R rows:
>> res = reshape(mat, R, [])
res =
1 5 1 2 8 2 3 7 3
2 5 1 3 6 2 4 7 4
>> res = sum(res)
res =
3 10 2 5 14 4 7 14 7
>> res = reshape(res, [], M)
res =
3 5 7
10 14 14
2 4 7
You can also do everything in one line:
>> reshape(sum(reshape(mat, R, [])), [], M)
ans =
3 5 7
10 14 14
2 4 7