I want to a element-by-element binary operation apply to large logical vectors. The content of these vectors is manly false, so for performance considerations it better to work with sparse matrices. If i do so the resulting matrix is not correct.
Examble
A = logical([0;1;0;0]);
B = logical([0 0 1 1]);
C = bsxfun(#and,A,B)
In this case C is
C =
0 0 0 0
0 0 1 1
0 0 0 0
0 0 0 0
If i use sparse matrices C is
C = full(bsxfun(#and,sparse(A),sparse(B)))
C =
0 0 0 0
1 1 1 1
0 0 0 0
0 0 0 0
Which is obviously wrong.
Did i oversee something or is this a Matlab bug.
I can reproduce this so it certainly seems to be a MATLAB bug. Especially considering that:
C = full(bsxfun(#times,sparse(A),sparse(B)))
C =
0 0 0 0
0 0 1 1
0 0 0 0
0 0 0 0
So, I would report it to The Mathworks.
However, in this particular case, I can't help feeling that bsxfun with sparse matrices isn't going to be the most efficient. Consider the following:
A = sparse(logical([0;1;0;0]));
B = sparse(logical([0 0 1 1]));
C_bsxfun = bsxfun(#and,full(A),full(B));
[i j] = ndgrid(find(A), find(B));
C_sparse = sparse(i, j, true, numel(A), numel(B));
isequal(C_bsxfun, full(C_sparse))
Related
I want to construct a matrix A in Matlab of dimension w x (m*w) where
each row is full of zeros except m consecutive ones that shift towards the right hand side as we move down to the rows.
Few examples can clarify
w=3,m=4
A=[1 1 1 1 0 0 0 0 0 0 0 0;
0 0 0 0 1 1 1 1 0 0 0 0;
0 0 0 0 0 0 0 0 1 1 1 1]
or
w=3, m=3
A=[1 1 1 0 0 0 0 0 0;
0 0 0 1 1 1 0 0 0;
0 0 0 0 0 0 1 1 1]
or
w=2, m=3
A=[1 1 1 0 0 0;
0 0 0 1 1 1]
I can't see how to proceed and any hint would be extremely helpful.
Step 1. Simplify the problem
If you write the "modified diagonal matrix" you are asking about as a row vector it will always look like the following
% 1 ... 1 0 ... ... 0 ... ... ... ... ... ... ... ... 1 ... 1
% m ones m*w zeros w-1 times the same as before m ones
Step 2. Think how to solve the simplified problem
The fundamental unit you need is a vector of m ones followed by m*w zeros;
Once you have built such vector, you need it to be repeated w times, MATLAB already knows how to do that;
The only thing you miss are the trailing ones: append them;
Now that the vector you were looking for is completed, you need to turn it into a matrix. MATLAB already knows how to do this too.
Final code
Once you understood the above steps, the final behaviour can be achieved even with a one-liner
>> m = 4; w = 3;
>> vec2mat([repmat([ones(1, m) zeros(1, m*w)], 1, w-1) ones(1, m)], w*m)
ans =
1 1 1 1 0 0 0 0 0 0 0 0
0 0 0 0 1 1 1 1 0 0 0 0
0 0 0 0 0 0 0 0 1 1 1 1
About speed
It's true, for loops aren't so slow anymore. I timed my one-liner solution, the trivial for loop and Luis Mendo's solution with eye() and repelem().
Click on images to zoom
Tested on the same machine, with MATLAB R2018a.
As you can see, as long as m and w are quite small, even if you could point out some differences in speed, them won't be noticeable to humans.
Anyway if you are going to work with bigger matrices, it becomes quite obvious which solution is the best.
Here are some approaches:
Using eye and repelem:
A = repelem(eye(w), 1, m);
Using eye and indexing:
A = eye(w);
A = A(1:w, ceil(1/m:1/m:w));
Using eye and kron:
A = kron(eye(w), ones(1,m));
Using singleton expansion:
A = bsxfun(#eq, (1:m).', ceil(1/m:1/m:w)); % Or A = (1:m).'==ceil(1/m:1/m:w);
I am trying to generate a rectangular matrix with 1s on the diagonal above the main diagonal and -1s on the main diagonal. I used "eye" which does not create the diagonal above the main.
Please find my attempt to this below.
N = 5
M1 = -eye([N-1 N])
M2 = eye([N N-1])'
M = M1+M2
I am unable to resolve this issue on my own. Any help or links to relevant documentation would be greatly appreciated.
I don't know of any prebuild function, but you can easily make such a matrix yourself:
N=5;
M=7;
diag=-eye(N,M);
upper_diag=horzcat(zeros(N,1),eye(N,M-1))
final=diag+upper_diag
using the identity matrix and some concatenation to shift the diagonal around. This example assumes you are looking for a square matrix.
The result looks like:
final =
-1 1 0 0 0 0 0
0 -1 1 0 0 0 0
0 0 -1 1 0 0 0
0 0 0 -1 1 0 0
0 0 0 0 -1 1 0
Just create eye and diag matrices as per normal, add them together, then chop away the rows you do not need:
nCol = 7;
nRow = 5;
M = -eye(nCol) + diag(ones(nCol - 1, 1), 1);
M = M(1:nRow, 1:nCol)
produces
M =
-1 1 0 0 0 0 0
0 -1 1 0 0 0 0
0 0 -1 1 0 0 0
0 0 0 -1 1 0 0
0 0 0 0 -1 1 0
The four-input version of spdiags does just that, producing a sparse matrix. You may need to convert to full then.
M = 5; %// number of rows
N = 7; %// number of columns
d = [0 1]; %// specify main diagonal and the one above
v = [-1 1]; %// values in those diagonals
result = full(spdiags(ones(M,1)*v, d, M, N));
This gives
result =
-1 1 0 0 0 0 0
0 -1 1 0 0 0 0
0 0 -1 1 0 0 0
0 0 0 -1 1 0 0
0 0 0 0 -1 1 0
I want to convert an integer i to a logical vector with an i-th non-zero element. That can de done with 1:10 == 2, which returns
0 1 0 0 0 0 0 0 0 0
Now, I want to vectorize this process for each row. Writing repmat(1:10, 2, 1) == [2 5]' I expect to get
0 1 0 0 0 0 0 0 0 0
0 0 0 0 1 0 0 0 0 0
But instead, this error occurs:
Error using ==
Matrix dimensions must agree.
Can I vectorize this process, or is a for loop the only option?
You can use bsxfun:
>> bsxfun(#eq, 1:10, [2 5].')
ans =
0 1 0 0 0 0 0 0 0 0
0 0 0 0 1 0 0 0 0 0
Note the transpose .' on the second vector; it's important.
Another way is to use eye and create a logical matrix that is n x n long, then use the indices to index into the rows of this matrix:
n = 10;
ind = [2 5];
E = eye(n,n) == 1;
out = E(ind, :);
We get:
>> out
out =
0 1 0 0 0 0 0 0 0 0
0 0 0 0 1 0 0 0 0 0
Just another possibility using indexing:
n = 10;
ind = [2 5];
x=zeros(numel(ind),n);
x(sub2ind([numel(ind),n],1:numel(ind),ind))=1;
For neural networking, I would like to represent a column vector y = [1;2;3] in a matrix like so:
y = [1 0 0;
0 1 0;
0 0 1]
My vector y is very large, and so hardcoding is not an option. Also, I would like to avoid using for-loops.
What I did so far:
y1 =[y; zeros(1,length(y)) ;zeros(1,length(y))] % add two rows with zeros in orde to give y the right format
idx = find(y1(1,:) == 2); % find all the columns containing a 2
y1(:,idx(1):idx(end)) = y1(:,[0;1;0]); % this does not work because now I am comparing a matrix with a vector
I also tried this:
y1( y1 == [2;0;0] )=[0;1;0]; % This of course does not work
Is there a way to specify I want to compare columns in y1 == [2;0;0], or is there another way to solve this?
From the context of your question, you wish to find a matrix where each column is an identity vector. For an identity vector, each column in this matrix is a non-zero vector where 1 is set in the position of the vector denoted by each position of y and 0 otherwise. Therefore, let's say we had the following example:
y = [1 5 4 3]
You would have y_out as the final matrix, which is:
y_out =
1 0 0 0
0 0 0 0
0 0 0 1
0 0 1 0
0 1 0 0
There are several ways to do this. The easiest one would be to declare the identity matrix with eye, then let y pick out those columns that you want from this matrix and place them as columns into your final matrix. If y had all unique values, then we would simply be rearranging the columns of this identity matrix based on y. As such:
y_out = eye(max(y));
y_out = y_out(:,y)
y_out =
1 0 0 0
0 0 0 0
0 0 0 1
0 0 1 0
0 1 0 0
Another way would be to declare a sparse matrix, where each row index is simply those elements from y and each column index is increasing from 1 up to as many elements as we have y:
y_out = sparse(y, 1:numel(y), 1, max(y), numel(y));
y_out = full(y_out)
y_out =
1 0 0 0
0 0 0 0
0 0 0 1
0 0 1 0
0 1 0 0
One more way would be to use sub2ind to find linear indices into your matrix, then access those elements and set them to 1. Therefore:
ind = sub2ind([max(y) numel(y)], y, 1:numel(y));
y_out = zeros(max(y), numel(y));
y_out(ind) = 1
y_out =
1 0 0 0
0 0 0 0
0 0 0 1
0 0 1 0
0 1 0 0
This works even if y has "missing" values:
n = numel(y);
y_matrix = zeros(n, max(y));
y_matrix((1:n) + (y-1)*n) = 1;
Example:
y = [1 5 3 2];
gives
y_matrix =
1 0 0 0 0
0 0 0 0 1
0 0 1 0 0
0 1 0 0 0
You can use bsxfun:
y_out = bsxfun(#eq, (1:max(y)).', y);
Not as efficient as the #rayryeng's answer but this might also help,
Also if there are repeated values in y this code works fine.
a = [1 2 3 2 5 7 6 8];
[X,Y] = meshgrid(a,1 : length(a));
A = X == Y;
A =
1 0 0 0 0 0 0 0
0 1 0 1 0 0 0 0
0 0 1 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 1 0 0 0
0 0 0 0 0 0 1 0
0 0 0 0 0 1 0 0
0 0 0 0 0 0 0 1
Suppose i have an identity matrix .
I=eye(3)
which will produce
I = [1 0 0
0 1 0
0 0 1]
Now i want to insert I into a (5X5) null matrix such that my result will be
N = [0 0 0 0 0
0 0 0 0 0
0 0 1 0 0
0 0 0 1 0
0 0 0 0 1]
How could i achieve this efficiently .Thanks in advace
With the Image processing toolbox, this could be done using padarray like this:
padarray(eye(3), [2 2], 'pre');
padarray pads an array with zeros. The [2 2] part says how many zeros to pad it with, in this case 2 rows and 2 columns. pre means you want it in front of the matrix, not after it (post).
Without it, you need to tweak it a bit more. One option could be to create an identity matrix of the full size, then make the first elements zero:
m = 5; %// size of matrix
n = 3; %// size of identity matrix
a = eye(m);
a(1:m-n,1:m-n) = 0;
a =
0 0 0 0 0
0 0 0 0 0
0 0 1 0 0
0 0 0 1 0
0 0 0 0 1
... or:
a = zeros(m);
a(m-n+1:m,m-n+1:m) = eye(n)
... or using sparse:
full(sparse(m-n+1:m,m-n+1:m,1))
Alternatively,
m = 5;
n = 3;
a = diag( [zeros(1, m-n), ones(1,n)] );