I have a data set that consists of two 1800 x 900 x 3 x 3 arrays, which should each be interpreted as a 1800x900 array of 3x3 matrices. As part of the analysis, at one point I need to create another such array by, at each point of the 1800x900 array, multiplying the corresponding 3x3 matrices together.
There seem to be two ways to do this that I can think of. The obvious way is
C = zeros(size(A))
for i = 1:900
for j = 1:1800
C(i,j,:,:) = A(i,j,:,:)*B(i,j,:,:)
end
end
But that's quite a long loop and doesn't really take advantage of MATLAB's vectorization. The other way is
C = zeros(size(A))
for i = 1:3
for j = 1:3
for k = 1:3
C(:,:,i,j) = C(:,:,i,j) + A(:,:,i,k).*B(:,:,k,j)
end
end
end
where the large dimensions are getting vectorized and I'm basically using the for loops to implement the Einstein summation convention. This seems really inelegant, though, which makes me think there should be a better way. Is there?
Perfect job for bsxfun with permute:
C = permute(sum(bsxfun(#times, A, permute(B, [1 2 5 3 4])), 4), [1 2 3 5 4]);
In R2016b onwards you can avoid bsxfun thanks to implicit singleton expansion:
C = permute(sum(A .* permute(B, [1 2 5 3 4]), 4), [1 2 3 5 4]);
Related
Is there a quick way of normalizing each row of a 3 dimensional Matrix without resorting to slow for loops in Matlab?
Say my input data looks like this:
d(:,:,1) =
1 2 3
4 5 6
d(:,:,2) =
7 8 9
10 11 12
I know that I can get the norm of each row by using
norms = sqrt(sum(d.^2,2))
norms(:,:,1) =
3.7417
8.7750
norms(:,:,2) =
13.9284
19.1050
But how to divide now the second dimension with these norm values?
I know that in 2 dims I can use ./ however this seems not to work for 3 dimensional data.
bsxfun is your friend:
out = bsxfun(#rdivide, d, norms);
What this does is that it temporarily creates a 3D matrix that replicates each row of norms for as many columns as there are in d and it divides each element in an element-wise manner with d and norms.
We get:
>> d = cat(3, [1 2 3; 4 5 6], [7 8 9; 10 11 12]);
>> norms = sqrt(sum(d.^2,2));
>> out = bsxfun(#rdivide, d, norms)
out(:,:,1) =
0.2673 0.5345 0.8018
0.4558 0.5698 0.6838
out(:,:,2) =
0.5026 0.5744 0.6462
0.5234 0.5758 0.6281
We can also verify that each row is L2-normalized by determining the sum of squares along each row independently and ensuring that each result sums to 1:
>> sum(out.^2, 2)
ans(:,:,1) =
1.0000
1.0000
ans(:,:,2) =
1.0000
1.0000
If the approach with bsxfun doesn't quite make sense, an alternative you could use is to create a matrix that respects the same dimensions as d by using repmat... then you can perform the element-wise division you desire:
>> out = d ./ repmat(norms, [1 size(d,2) 1])
out(:,:,1) =
0.2673 0.5345 0.8018
0.4558 0.5698 0.6838
out(:,:,2) =
0.5026 0.5744 0.6462
0.5234 0.5758 0.6281
With repmat you specify how many times you want the matrix to be copied in each dimension. We only want the matrix to be replicated over the columns while the number of rows and slices are the same... hence the vector [1 size(d,2) 1] that specifies how many times you want the matrix copied in each dimension.
Actually, this is what bsxfun does under the hood without you having to deal with the headaches of creating this temporary matrix. This replication is done for you without having you think about it.
Thanks in advance for the help.
Suppose I have a multi D array x such that
x(:,:,1) = [1 2; 3 4];
x(:,:,2) = [5 6; 7 8];
and a matrix y such that
y = [1 2; 5 6];
I would like to add the first row of y to each row of x(:,:,1) and the second row of y to each element of x(:,:,2). This will produce an array z such that
z(:,:,1) = [2 4; 4 6];
z(:,:,2) = [10 12; 12 14];
In reality (not the example I am giving) I would like to do this operation on a very large multi D array x and a very large matrix y. I therefore want to do this as efficiently as possible.
The naive approach would be to use for loops to do this, but this would not be efficient what-so-ever. I believe that an efficient approach would be to incorporate bsxfun to accomplish this, but I haven't been able to figure out an approach. y and x can and be restructured to accomplish this task as long as the same z is produced and, most importantly, the amount of time needed to build z is less than the for loop approach.
I was able to find this which does what I want, but only for multiplication and not summation. I could modify this code to do what I want but I feel as though with summation that there has to be a simpler approach.
You just need bsxfun and a little bit of permute:
z = bsxfun(#plus, x, permute(y, [3 2 1]));
The key is to properly rearrange the dimensions of y so that the singleton expansion performed by bsxfun gives you the desired result.
I have a 5 by 3 matrix, e.g the following:
A=[1 1 1; 2 2 2; 3 3 3; 4 4 4; 5 5 5]
I run a for loop:
for i = 1:5
AA = A(i)'*A(i);
end
My question is how to store each of the 5 (3 by 3) AA matrices?
Thanks.
You could pre-allocate enough memory to the AA matrix to hold all the results:
[r,c] = size(A); % get the rows and columns of A (r and c respectively)
AA = zeros(c,c,r); % pre-allocate memory to AA for all 5 products
% (so we have 5 3x3 arrays)
Now do almost the same loop as above BUT realize that A(i) in the above code only returns one element whereas you want the full row. So you want the data from row i but all columns which can be represented as 1:3 or just the colon :
for i=1:r
AA(:,:,i) = A(i,:)' * A(i,:);
end
In the above, A(i,:) is the ith row of A and we are setting all rows and columns in the third dimension (i) of AA to the result of the product.
Assuming, as in Geoff's answer, that you mean A(i,:)'*A(i,:) (to get 5 matrices of size 3x3 in your example), you can do it in one line with bsxfun and permute:
AA = bsxfun(#times, permute(A, [3 2 1]), permute(A, [2 3 1]));
(I'm also assuming that your matrices only contain real numbers, as in your example. If by ' you really mean conjugate transpose, you need to add a conj in the above).
I have two matrices, A and B. (B is continuous like 1:n)
I need to find all the occurrences of each individual row of B in A, and store those row indices accordingly in cell array C. See below for an example.
A = [3,4,5;1,3,5;1,4,3;4,2,1]
B = [1;2;3;4;5]
Thus,
C = {[2,3,4];[4];[1,2,3];[1,3,4];[1,2]}
Note C does not need to be in a cell array for my application. I only suggest it because the row vectors of C are of unequal length. If you can suggest a work-around, this is fine too.
I've tried using a loop running ismember for each row of B, but this is too slow when the matrices A and B are huge, with around a million entries. Vectorized code is appreciated.
(To give you context, the purpose of this is to identify, in a mesh, those faces that are attached to a single vertex. Note I cannot use the function edgeattachments because my data are not of the form "TR" in triangulation representation. All I have is a list of faces and list of vertices.)
Well, the best answer for this would require knowledge of how A is filled. If A is sparse, that is, if it has few columns values and B is quite large, then I think the best way for memory saving may be using a sparse matrix instead of a cell.
% No fancy stuff, just fast and furious
bMax = numel(B);
nRows = size(A,1);
cLogical = sparse(nRows,bMax);
for curRow = 1:nRows
curIdx = A(curRow,:);
cLogical(curRow,curIdx) = 1;
end
Answer:
cLogical =
(2,1) 1
(3,1) 1
(4,1) 1
(4,2) 1
(1,3) 1
(2,3) 1
(3,3) 1
(1,4) 1
(3,4) 1
(4,4) 1
(1,5) 1
(2,5) 1
How to read the answer. For each column the rows show the indexes that the column index appears in A. That is 1 appears in rows [2 3 4], 2 appear in row [4], 3 rows [1 2 3], 4 row [1 3 4], 5 in row [1 2].
Then you can use cLogical instead of a cell as an indexing matrix in the future for your needs.
Another way would be to allocate C with the expected value for how many times an index should appear in C.
% Fancier solution using some assumed knowledge of A
bMax = numel(B);
nRows = size(A,1);
nColumns = size(A,2);
% Pre-allocating with the expected value, an attempt to reduce re-allocations.
% tic; for rep=1:10000; C = mat2cell(zeros(bMax,nColumns),ones(1,bMax),nColumns); end; toc
% Elapsed time is 1.364558 seconds.
% tic; for rep=1:10000; C = repmat({zeros(1,nColumns)},bMax,1); end; toc
% Elapsed time is 0.606266 seconds.
% So we keep the not fancy repmat solution
C = repmat({zeros(1,nColumns)},bMax,1);
for curRow = 1:nRows
curIdxMsk = A(curRow,:);
for curCol = 1:nColumns
curIdx = curIdxMsk(curCol);
fillIdx = ~C{curIdx};
if any(fillIdx)
fillIdx = find(fillIdx,1);
else
fillIdx = numel(fillIdx)+1;
end
C{curIdx}(fillIdx) = curRow;
end
end
% Squeeze empty indexes:
for curRow = 1:bMax
C{curRow}(~C{curRow}) = [];
end
Answer:
>> C{:}
ans =
2 3 4
ans =
4
ans =
1 2 3
ans =
1 3 4
ans =
1 2
Which solution will performs best? You do a performance test in your code because it depends on how big is A, bMax, the memory size of your computer and so on. Yet, I'm still curious with solutions other people can do for this x). I liked chappjc's solution although it has the cons that he has pointed out.
For the given example (10k times):
Solution 1: Elapsed time is 0.516647 seconds.
Solution 2: Elapsed time is 4.201409 seconds (seems that solution 2 is a bad idea hahaha, but since it was created to the specific issue of A having many rows it has to be tested in those conditions).
chappjc' solution: Elapsed time is 2.405341 seconds.
We can do it without making any assumptions about B. Try this use of bsxfun and mat2cell:
M = squeeze(any(bsxfun(#eq,A,permute(B,[3 2 1])),2)); % 4x3x1 #eq 1x1x5 => 4x3x5
R = sum(M); % 4x5 -> 1x5
[ii,jj] = find(M);
C = mat2cell(ii,R)
The cells in C above will be column vectors rather than rows as in your example. To make the cells contain row vectors, use C = mat2cell(ii',1,R)' instead.
My only concern is that mat2cell could be slow for millions of values of R, but if you want your output in a cell, I'm not sure how much better you can do. EDIT: If you can deal with a sparse matrix like in Werner's first solution with the loop, replace the last line of the above with the following:
>> Cs = sparse(ii,jj,1)
Cs =
(2,1) 1
(3,1) 1
(4,1) 1
(4,2) 1
(1,3) 1
(2,3) 1
(3,3) 1
(1,4) 1
(3,4) 1
(4,4) 1
(1,5) 1
(2,5) 1
Unfortunately, bsxfun will probably run out of memory if both size(A,1) and numel(B) are large! You may have to loop over the elements of A or B if memory becomes an issue. Here's one way to do it by looping over your vertexes in B:
for i=1:numel(B), C{i} = find(any(A==B(i),2)); end
Yup, that easy. Cell array growing is extremely fast in MATLAB as it similar to a sequence container that stores contiguous references to the data, rather than keeping the data itself contiguous. Perhaps ismember was the bottleneck in your test.
I have 2 matrices A (nxm) and B (nxd) and want to multiply element-wise each column of A with a row of B. There are m columns in A and n 1xd vectors in B so the results are m nxd matrices. Then I want to sum(result_i, 1) to get m 1xd vectors, which I want to apply vertcat to get a mxd matrix. I'm doing this operations using for loop and it is slow because n and d are big. How can I vectorize this in matlab to make it faster? Thank you.
EDIT:
You're all right: I was confused by my own question. What I meant by "multiply element-wise each column of A with a row of B" is to multiply n elements of a column in A with the corresponding n rows of B. What I want to do with one column of A is as followed (and I repeat this for m columns of A, then vertcat the C's vector together to get an mxd matrix):
column_of_A =
3
3
1
B =
3 1 3 3
2 2 1 2
1 3 3 3
C = sum(diag(column_of_A)*B, 1)
16 12 15 18
You can vectorize your operation the following way. Note, however, that vectorizing comes at the cost of higher memory usage, so the solution may end up not working for you.
%# multiply nxm A with nx1xd B to create a nxmxd array
tmp = bsxfun(#times,A,permute(B,[1 3 2]));
%# sum and turn into mxd
out = squeeze(sum(tmp,1));
You may want to do everything in one line, which may help the Matlab JIT compiler to save on memory.
EDIT
Here's a way to replace the first line if you don't have bsxfun
[n,m] = size(A);
[n,d] = size(B);
tmp = repmat(A,[1 1 d]) .* repmat(permute(B,[1 3 2]),[1,m,1]);
It's ugly, but as far as I can see, it works. I'm not sure it will be faster than your loop though, plus, it has a large memory overhead. Anyway, here goes:
A_3D = repmat(reshape(A, size(A, 1), 1, size(A, 2)), 1, size(B, 2));
B_3D = repmat(B, [ 1 1 size(A, 2)]);
result_3D = sum(A_3D .* B_3D, 1);
result = reshape(result_3D, size(A, 2), size(B, 2))
What it does is: make A into a 3D matrix of size n x 1 x m, so one column in each index of the 3rd dimension. Then we repeat the matrix so we get an n x d x m matrix. We repeat B in the 3rd dimension as well. We then do a piecewise multiplication of all the elements and sum them. The resulting matrix is a 1 x d x m matrix. We reshape this into a m x d matrix.
I'm pretty sure I switched around the size of the dimensions a few times in my explanation, but I hope you get the general gist.
Multiplying with a diagonal matrix seems at least twice as fast, but I couldn't find a way to use diag, since it wants a vector or 2D matrix as input. I might try again later tonight, I feel there must be a faster way :).
[Edit] Split up the command in parts to at least make it a little bit readable.
This is the way I would do this:
sum(repmat(A,1,4).*B)
If you don't know the number of columns of B:
sum(repmat(A,1,size(B,2)).*B)