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);
Related
I have a logical vector in which I would like to iterate over every n-elements. If in any given window at least 50% are 1's, then I change every element to 1, else I keep as is and move to the next window. For example.
n = 4;
input = [0 0 0 1 0 1 1 0 0 0 0 1 0 1 0 1 0 0 0 1];
output = func(input,4);
output = [0 0 0 1 1 1 1 1 0 0 0 1 1 1 1 1 0 0 0 1];
This function is trivial to implement but is it possible to apply a vectorized implementation using logical indexing?. I am trying to build up the intuition of applying this technique.
here's a one liner (that works for your input):
func = #(input,n) input | kron(sum(reshape(input ,n,[]))>=n/2,ones(1,n));
of course, there are cases to solve that this doesnt answer, what if the size of the input is not commensurate in n? etc...
i'm not sure if that's what you meant by vectorization, and I didnt benchmark it vs a for loop...
Here is one way of doing it. Once understood you can compact it in less lines but I'll details the intermediate steps for the sake of clarity.
%% The inputs
n = 4;
input = [0 0 0 1 0 1 1 0 0 0 0 1 0 1 0 1 0 0 0 1];
1) Split your input into blocks of size n (note that your final function will have to check that the number of elements in input is a integer multiple of n)
c = reshape(input,n,[]) ;
Gives you a matrix with your blocks organized in columns:
c =
0 0 0 0 0
0 1 0 1 0
0 1 0 0 0
1 0 1 1 1
2) Perform your test condition on each of the block. For this we'll take advantage that Matlab is working column wise for the sum function:
>> cr = sum(c) >= (n/2)
cr =
0 1 0 1 0
Now you have a logical vector cr containing as many elements as initial blocks. Each value is the result of the test condition over the block. The 0 blocks will be left unchanged, the 1 blocks will be forced to value 1.
3) Force 1 columns/block to value 1:
>> c(:,cr) = 1
c =
0 1 0 1 0
0 1 0 1 0
0 1 0 1 0
1 1 1 1 1
4) Now all is left is to unfold your matrix. You can do it several ways:
res = c(:) ; %% will give you a column vector
OR
>> res = reshape(c,1,[]) %% will give you a line vector
res =
0 0 0 1 1 1 1 1 0 0 0 1 1 1 1 1 0 0 0 1
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 have a binary matrix like this:
0 0 0 0 0 0
0 0 0 1 0 0
0 1 0 0 0 0
0 0 1 0 1 0
0 0 0 1 0 0
0 0 0 0 0 0
and I want to trim this matrix (in other words, remove zeroes at the boundaries) to be like:
0 0 1 0
1 0 0 0
0 1 0 1
0 0 1 0
How to do this the "Matlab" way? that's not to use conventional loops and conditions.
To be clearer, the matrix should be reduced to start from the first column which has at least one 1, and ends at the last column with the same condition, inclusive. Any column out of this range should be removed. Same rules apply for rows.
Thanks.
If you have the data in matrix M...
x = find(any(M,2),1,'first'):find(any(M,2),1,'last');
y = find(any(M),1,'first'):find(any(M),1,'last');
M(x, y)
Or, if you know that there will be a 1 in every row/col except the edges:
M(any(M,2), any(M))
Extension to higher dimensions:
Assuming a 3D matrix to be trimmed, this is more straightforward:
M=rand(3,3,3); % generating a random 3D matrix
M(2,:,:)=0; % just to make a check if it works in extreme case of having zeros in the middle
padded = padarray(M,[2 2 2]); % making some zero boundaries
[r,c,v]=ind2sub(size(padded),find(padded));
recoveredM=padded(min(r):max(r),min(c):max(c),min(v):max(v));
check=M==recoveredM % checking to see if M is successfully recovered
You could use the fact that find can return row and column indices:
[r1, c1] = find(x, 1, 'first')
[r2, c2] = find(x, 1, 'last')
x(r1:r2, c1:c2)
Given a matrix where 1 is the current subset
test =
0 0 0 0 0 0
0 0 0 0 0 0
0 0 1 1 0 0
0 0 1 1 0 0
0 0 0 0 0 0
0 0 0 0 0 0
Is there a function, or quick method to get change the subset to the boundary of the current subset?
Eg. Get this subset from 'test' above
test =
0 0 0 0 0 0
0 1 1 1 1 0
0 1 0 0 1 0
0 1 0 0 1 0
0 1 1 1 1 0
0 0 0 0 0 0
In the end I just want to get the minimum of the cells surrounding a subset of a matrix. Sure I could loop through and get the minimum of the boundary (cell by cell), but there must be a way to do it with the method i've shown above.
Note the subset WILL be connected, but may not be rectangular. This may be the big catch.
This is a possible subset.... (Would pad this with a NaN border)
test =
0 0 0 0 0 0
0 0 0 0 0 0
0 0 1 1 0 0
0 0 1 1 0 0
0 0 1 1 1 1
0 0 1 1 1 1
Ideas?
The basic steps I'd use are:
Perform a dilation on the shape to get a new area which is the shape plus its boundary
Subtract the original shape from the dilated shape to leave just the boundary
Use the boundary to index your data matrix, then take the minimum.
Dilation
What I want to do here is pass a 3x3 window over each cell and take the maximum value in that window:
[m, n] = size(A); % assuming A is your original shape matrix
APadded = zeros(m + 2, n + 2);
APadded(2:end-1, 2:end-1) = A; % pad A with zeroes on each side
ADilated = zeros(m + 2, n + 2); % this will hold the dilated shape.
for i = 1:m
for j = 1:n
mask = zeros(size(APadded));
mask(i:i+2, j:j+2) = 1; % this places a 3x3 square of 1's around (i, j)
ADilated(i + 1, j + 1) = max(APadded(mask));
end
end
Shape subtraction
This is basically a logical AND and a logical NOT to remove the intersection:
ABoundary = ADilated & (~APadded);
At this stage you may want to remove the border we added to do the dilation, since we don't need it any more.
ABoundary = ABoundary(2:end-1, 2:end-1);
Find the minimum data point along the boundary
We can use our logical boundary to index the original data into a vector, then just take the minimum of that vector.
dataMinimum = min(data(ABoundary));
You should look at this as morphology problem, not set theory. This can be solved pretty easily with imdilate() (requires the image package). You basically only need to subtract the image to its dilation with a 3x3 matrix of 1.
octave> test = logical ([0 0 0 0 0 0
0 0 0 0 0 0
0 0 1 1 0 0
0 0 1 1 0 0
0 0 1 1 1 1
0 0 1 1 1 1]);
octave> imdilate (test, true (3)) - test
ans =
0 0 0 0 0 0
0 1 1 1 1 0
0 1 0 0 1 0
0 1 0 0 1 1
0 1 0 0 0 0
0 1 0 0 0 0
It does not, however, pads with NaN. If you really want that, you could pad your original matrix with false, do the operation, and then check if there's any true values in the border.
Note that you don't have to use logical() in which case you'll have to use ones() instead of true(). But that takes more memory and has worse performance.
EDIT: since you are trying to do it without using any matlab toolbox, take a look at the source of imdilate() in Octave. For the case of logical matrices (which is your case) it's a simple usage of filter2() which belongs to matlab core. That said, the following one line should work fine and be much faster
octave> (filter2 (true (3), test) > 0) - test
ans =
0 0 0 0 0 0
0 1 1 1 1 0
0 1 0 0 1 0
0 1 0 0 1 1
0 1 0 0 0 0
0 1 0 0 0 0
One possible solution is to take the subset and add it to the original matrix, but ensure that each time you add it, you offset its position by +1 row, -1 row and +1 column, -1 column. The result will then be expanded by one row and column all around the original subset. You then use the original matrix to mask the original subet to zero.
Like this:
test_new = test + ...
[[test(2:end,2:end);zeros(1,size(test,1)-1)],zeros(size(test,1),1)] + ... %move subset up-left
[[zeros(1,size(test,1)-1);test(1:end-1,2:end)],zeros(size(test,1),1)] + ... %move down-left
[zeros(size(test,1),1),[test(2:end,1:end-1);zeros(1,size(test,1)-1)]] + ... %move subset up-right
[zeros(size(test,1),1),[zeros(1,size(test,1)-1);test(1:end-1,1:end-1)]]; %move subset down-right
test_masked = test_new.*~test; %mask with original matrix
result = test_masked;
result(result>1)=1; % ensure that there is only 1's, not 2, 3, etc.
The result for this on your test matrix is:
result =
0 0 0 0 0 0
0 1 1 1 1 0
0 1 0 0 1 0
0 1 0 0 1 1
0 1 0 0 0 0
0 1 0 0 0 0
Edited - it now grabs the corners as well, by moving the subset up and to the left, up and to the right, down then left and down then right.
I expect this would be a very quick way to achieve this - it doesn't have any loops, nor functions - just matrix operations.
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))