Is there a efficient way to do the computation of a multivariate gaussian (as below) that returns matrix p , that is, making use of some sort of vectorization? I am aware that matrix p is symmetric, but still for a matrix of size 40000x3, for example, this will take quite a long time.
Matlab code example:
DataMatrix = [3 1 4; 1 2 3; 1 5 7; 3 4 7; 5 5 1; 2 3 1; 4 4 4];
[rows, cols ] = size(DataMatrix);
I = eye(cols);
p = zeros(rows);
for k = 1:rows
p(k,:) = mvnpdf(DataMatrix(:,:),DataMatrix(k,:),I);
end
Stage 1: Hack into source code
Iteratively we are performing mvnpdf(DataMatrix(:,:),DataMatrix(k,:),I)
The syntax is : mvnpdf(X,Mu,Sigma).
Thus, the correspondence with our input becomes :
X = DataMatrix(:,:);
Mu = DataMatrix(k,:);
Sigma = I
For the sizes relevant to our situation, the source code mvnpdf.m reduces to -
%// Store size parameters of X
[n,d] = size(X);
%// Get vector mean, and use it to center data
X0 = bsxfun(#minus,X,Mu);
%// Make sure Sigma is a valid covariance matrix
[R,err] = cholcov(Sigma,0);
%// Create array of standardized data, and compute log(sqrt(det(Sigma)))
xRinv = X0 / R;
logSqrtDetSigma = sum(log(diag(R)));
%// Finally get the quadratic form and thus, the final output
quadform = sum(xRinv.^2, 2);
p_out = exp(-0.5*quadform - logSqrtDetSigma - d*log(2*pi)/2)
Now, if the Sigma is always an identity matrix, we would have R as an identity matrix too. Therefore, X0 / R would be same as X0, which is saved as xRinv. So, essentially quadform = sum(X0.^2, 2);
Thus, the original code -
for k = 1:rows
p(k,:) = mvnpdf(DataMatrix(:,:),DataMatrix(k,:),I);
end
reduces to -
[n,d] = size(DataMatrix);
[R,err] = cholcov(I,0);
p_out = zeros(rows);
K = sum(log(diag(R))) + d*log(2*pi)/2;
for k = 1:rows
X0 = bsxfun(#minus,DataMatrix,DataMatrix(k,:));
quadform = sum(X0.^2, 2);
p_out(k,:) = exp(-0.5*quadform - K);
end
Now, if the input matrix is of size 40000x3, you might want to stop here. But with system resources permitting, you can vectorize everything as discussed next.
Stage 2: Vectorize everything
Now that we see what's actually going on and that the computations look parallelizable, it's time to step-up to use bsxfun in 3D with his good friend permute for a vectorized solution, like so -
%// Get size params and R
[n,d] = size(DataMatrix);
[R,err] = cholcov(I,0);
%// Calculate constants : "logSqrtDetSigma" and "d*log(2*pi)/2`"
K1 = sum(log(diag(R)));
K2 = d*log(2*pi)/2;
%// Major thing happening here as we calclate "X0" for all iterations
%// in one go with permute and bsxfun
diffs = bsxfun(#minus,DataMatrix,permute(DataMatrix,[3 2 1]));
%// "Sigma" is an identity matrix, so it plays no in "/R" at "xRinv = X0 / R".
%// Perform elementwise squaring and summing rows to get vectorized "quadform"
quadform1 = squeeze(sum(diffs.^2,2))
%// Finally use "quadform1" and get vectorized output as a 2D array
p_out = exp(-0.5*quadform1 - K1 - K2)
Related
Given two vectors X and Y of length n, representing points on the plane, and a neighborhood radius rad, is there a vectorized way to compute the neighborhood matrix of the points?
In other words, can the following (painfully slow for large n) loop be vectorized:
neighborhood_mat = zeros(n, n);
for i = 1 : n
for j = 1 : i - 1
dist = norm([X(j) - X(i), Y(j) - Y(i)]);
if (dist < radius)
neighborhood_mat(i, j) = 1;
neighborhood_mat(j, i) = 1;
end
end
end
Approach #1
bsxfun based approach -
out = bsxfun(#minus,X,X').^2 + bsxfun(#minus,Y,Y').^2 < radius^2
out(1:n+1:end)= 0
Approach #2
Distance matrix calculation using matrix-multiplication based approach (possibly faster) -
A = [X(:) Y(:)]
A_t = A.'; %//'
out = [-2*A A.^2 ones(n,3)]*[A_t ; ones(3,n) ; A_t.^2] < radius^2
out(1:n+1:end)= 0
Approach #3
With pdist and squareform -
A = [X(:) Y(:)]
out = squareform(pdist(A))<radius
out(1:n+1:end)= 0
Approach #4
You can use pdist as with the previous approach, but avoid squareform with some logical indexing to get the final output of neighbourhood matrix as shown below -
A = [X(:) Y(:)]
dists = pdist(A)< radius
mask_lower = bsxfun(#gt,[1:n]',1:n) %//'
%// OR tril(true(n),-1)
mask_upper = bsxfun(#lt,[1:n]',1:n) %//'
%// OR mask_upper = triu(true(n),1)
%// OR mask_upper = ~mask_lower; mask_upper(1:n+1:end) = false;
out = zeros(n)
out(mask_lower) = dists
out_t = out' %//'
out(mask_upper) = out_t(mask_upper)
Note: As one can see, for the all above mentioned approaches, we are using pre-allocation for the output. A fast way to pre-allocate would be with out(n,n) = 0 and is based upon this wonderful blog on undocumented MATLAB. This should really speed up those approaches!
The following approach is great if the number of points in your neighborhoods is small or you run low on memory using the brute-force approach:
If you have the statistics toolbox installed, you can have a look at the rangesearch method.
(Free alternatives include the k-d tree implementations of a range search on the File Exchange.)
The usage of rangesearch is straightforward:
P = [X,Y];
[idx,D] = rangesearch(P, P, rad);
It returns a cell-array idx of the indices of nodes within reach and their distances D.
Depending on the size of your data, this could be beneficial in terms of speed and memory.
Instead of computing all pairwise distances and then filtering out those that are large, this algorithm builds a data structure called a k-d tree to more efficiently search close points.
You can then use this to build a sparse matrix:
I = cell2mat(idx.').';
J = runLengthDecode(cellfun(#numel,idx));
n = size(P,1);
S = sparse(I,J,1,n,n)-speye(n);
(This uses the runLengthDecode function from this answer.)
You can also have a look at the KDTreeSearcher class if your data points don't change and you want to query your data lots of times.
In Matlab I am looking for a way to most efficiently calculate a frequency averaged periodogram on a GPU.
I understand that the most important thing is to minimise for loops and use the already built in GPU functions. However my code still feels relatively unoptimised and I was wondering what changes I can make to it to gain a better speed up.
r = 5; % Dimension
n = 100; % Time points
m = 20; % Bandwidth of smoothing
% Generate some random rxn data
X = rand(r, n);
% Generate normalised weights according to a cos window
w = cos(pi * (-m/2:m/2)/m);
w = w/sum(w);
% Generate non-smoothed Periodogram
FT = (n)^(-0.5)*(ctranspose(fft(ctranspose(X))));
Pdgm = zeros(r, r, n/2 + 1);
for j = 1:n/2 + 1
Pdgm(:,:,j) = FT(:,j)*FT(:,j)';
end
% Finally smooth with our weights
SmPdgm = zeros(r, r, n/2 + 1);
% Take advantage of the GPU filter function
% Create new Periodogram WrapPdgm with m/2 values wrapped around in front and
% behind it (it seems like there is redundancy here)
WrapPdgm = zeros(r,r,n/2 + 1 + m);
WrapPdgm(:,:,m/2+1:n/2+m/2+1) = Pdgm;
WrapPdgm(:,:,1:m/2) = flip(Pdgm(:,:,2:m/2+1),3);
WrapPdgm(:,:,n/2+m/2+2:end) = flip(Pdgm(:,:,n/2-m/2+1:end-1),3);
% Perform filtering
for i = 1:r
for j = 1:r
temp = filter(w, [1], WrapPdgm(i,j,:));
SmPdgm(i,j,:) = temp(:,:,m+1:end);
end
end
In particular, I couldn't see a way to optimise out the for loop when calculating the initial Pdgm from the Fourier transformed data and I feel the trick I play with the WrapPdgm in order to take advantage of filter() on the GPU feels unnecessary if there were a smooth function instead.
Solution Code
This seems to be pretty efficient as benchmark runtimes in the next section might convince us -
%// Select the portion of FT to be processed and
%// send copy to GPU for calculating everything
gFT = gpuArray(FT(:,1:n/2 + 1));
%// Perform non-smoothed Periodogram, thus removing the first loop
Pdgm1 = bsxfun(#times,permute(gFT,[1 3 2]),permute(conj(gFT),[3 1 2]));
%// Generate WrapPdgm right on GPU
WrapPdgm1 = zeros(r,r,n/2 + 1 + m,'gpuArray');
WrapPdgm1(:,:,m/2+1:n/2+m/2+1) = Pdgm1;
WrapPdgm1(:,:,1:m/2) = Pdgm1(:,:,m/2+1:-1:2);
WrapPdgm1(:,:,n/2+m/2+2:end) = Pdgm1(:,:,end-1:-1:n/2-m/2+1);
%// Perform filtering on GPU and get the final output, SmPdgm1
filt_data = filter(w,1,reshape(WrapPdgm1,r*r,[]),[],2);
SmPdgm1 = gather(reshape(filt_data(:,m+1:end),r,r,[]));
Benchmarking
Benchmarking Code
%// Input parameters
r = 50; % Dimension
n = 1000; % Time points
m = 200; % Bandwidth of smoothing
% Generate some random rxn data
X = rand(r, n);
% Generate normalised weights according to a cos window
w = cos(pi * (-m/2:m/2)/m);
w = w/sum(w);
% Generate non-smoothed Periodogram
FT = (n)^(-0.5)*(ctranspose(fft(ctranspose(X))));
tic, %// ... Code from original approach, toc
tic %// ... Code from proposed approach, toc
Runtime results thus obtained on GPU, GTX 750 Ti against CPU, I-7 4790K -
------------------------------ With Original Approach on CPU
Elapsed time is 0.279816 seconds.
------------------------------ With Proposed Approach on GPU
Elapsed time is 0.169969 seconds.
To get rid of the first loop you can do the following:
Pdgm_cell = cellfun(#(x) x * x', mat2cell(FT(:, 1 : 51), [5], ones(51, 1)), 'UniformOutput', false);
Pdgm = reshape(cell2mat(Pdgm_cell),5,5,[]);
Then in your filter you can do the following:
temp = filter(w, 1, WrapPdgm, [], 3);
SmPdgm = temp(:, :, m + 1 : end);
The 3 lets the filter know to operate along the 3rd dimension of your data.
You can use pagefun on the GPU for the first loop. (Note that the implementation of cellfun is basically a hidden loop, whereas pagefun runs natively on the GPU using a batched GEMM operation). Here's how:
n = 16;
r = 8;
X = gpuArray.rand(r, n);
R = gpuArray.zeros(r, r, n/2 + 1);
for jj = 1:(n/2+1)
R(:,:,jj) = X(:,jj) * X(:,jj)';
end
X2 = X(:,1:(n/2+1));
R2 = pagefun(#mtimes, reshape(X2, r, 1, []), reshape(X2, 1, r, []));
R - R2
for an input matrix
in = [1 1;
1 2;
1 3;
1 4;
2 5;
2 6;
2 7;
3 8;
3 9;
3 10;
3 11];
i want to get the output matrix
out = [1 5 8;
2 6 9;
3 7 10;
4 0 11];
meaning i want to reshape the second input column into an output matrix, where all values corresponding to one value in the first input column are written into one column of the output matrix.
As there can be different numbers of entries for each value in the first input column (here 4 values for "1" and "3", but only 3 for "2"), the normal reshape function is not applicable. I need to pad all columns to the maximum number of rows.
Do you have an idea how to do this matlab-ish?
The second input column can only contain positive numbers, so the padding values can be 0, -x, NaN, ...
The best i could come up with is this (loop-based):
maxNumElem = 0;
for i=in(1,1):in(end,1)
maxNumElem = max(maxNumElem,numel(find(in(:,1)==i)));
end
out = zeros(maxNumElem,in(end,1)-in(1,1));
for i=in(1,1):in(end,1)
tmp = in(in(:,1)==i,2);
out(1:length(tmp),i) = tmp;
end
Either of the following approaches assumes that column 1 of in is sorted, as in the example. If that's not the case, apply this initially to sort in according to that criterion:
in = sortrows(in,1);
Approach 1 (using accumarray)
Compute the required number of rows, using mode;
Use accumarray to gather the values corresponding to each column, filled with zeros at the end. The result is a cell;
Concatenate horizontally the contents of all cells.
Code:
[~, n] = mode(in(:,1)); %//step 1
out = accumarray(in(:,1), in(:,2), [], #(x){[x; zeros(n-numel(x),1)]}); %//step 2
out = [out{:}]; %//step 3
Alternatively, step 1 could be done with histc
n = max(histc(in(:,1), unique(in(:,1)))); %//step 1
or with accumarray:
n = max(accumarray(in(:,1), in(:,2), [], #(x) numel(x))); %//step 1
Approach 2 (using sparse)
Generate a row-index vector using this answer by #Dan, and then build your matrix with sparse:
a = arrayfun(#(x)(1:x), diff(find([1,diff(in(:,1).'),1])), 'uni', 0); %//'
out = full(sparse([a{:}], in(:,1), in(:,2)));
Introduction to proposed solution and Code
Proposed here is a bsxfun based masking approach that uses the binary operators available as builtins for use with bsxfun and as such I would consider this very appropriate for problems like this. Of course, you must also be aware that bsxfun is a memory hungry tool. So, it could pose a threat if you are dealing with maybe billions of elements depending also on the memory available for MATLAB's usage.
Getting into the details of the proposed approach, we get the counts of each ID from column-1 of the input with histc. Then, the magic happens with bsxfun + #le to create a mask of positions in the output array (initialized by zeros) that are to be filled by the column-2 elements from input. That's all you need to tackle the problem with this approach.
Solution Code
counts = histc(in(:,1),1:max(in(:,1)))'; %//' counts of each ID from column1
max_counts = max(counts); %// Maximum counts for each ID
mask = bsxfun(#le,[1:max_counts]',counts); %//'# mask of locations where
%// column2 elements are to be placed
out = zeros(max_counts,numel(counts)); %// Initialize the output array
out(mask) = in(:,2); %// place the column2 elements in the output array
Benchmarking (for performance)
The benchmarking presented here compares the proposed solution in this post against the various methods presented in Luis's solution. This skips the original loopy approach presented in the problem as it appeared to be very slow for the input generated in the benchmarking code.
Benchmarking Code
num_ids = 5000;
counts_each_id = randi([10 100],num_ids,1);
num_runs = 20; %// number of iterations each approach is run for
%// Generate random input array
in = [];
for k = 1:num_ids
in = [in ; [repmat(k,counts_each_id(k),1) rand(counts_each_id(k),1)]];
end
%// Warm up tic/toc.
for k = 1:50000
tic(); elapsed = toc();
end
disp('------------- With HISTC + BSXFUN Masking approach')
tic
for iter = 1:num_runs
counts = histc(in(:,1),1:max(in(:,1)))';
max_counts = max(counts);
out = zeros(max_counts,numel(counts));
out(bsxfun(#le,[1:max_counts]',counts)) = in(:,2);
end
toc
clear counts max_counts out
disp('------------- With MODE + ACCUMARRAY approach')
tic
for iter = 1:num_runs
[~, n] = mode(in(:,1)); %//step 1
out = accumarray(in(:,1), in(:,2), [], #(x){[x; zeros(n-numel(x),1)]}); %//step 2
out = [out{:}];
end
toc
clear n out
disp('------------- With HISTC + ACCUMARRAY approach')
tic
for iter = 1:num_runs
n = max(histc(in(:,1), unique(in(:,1))));
out = accumarray(in(:,1), in(:,2), [], #(x){[x; zeros(n-numel(x),1)]}); %//step 2
out = [out{:}];
end
toc
clear n out
disp('------------- With ARRAYFUN + Sparse approach')
tic
for iter = 1:num_runs
a = arrayfun(#(x)(1:x), diff(find([1,diff(in(:,1).'),1])), 'uni', 0); %//'
out = full(sparse([a{:}], in(:,1), in(:,2)));
end
toc
clear a out
Results
------------- With HISTC + BSXFUN Masking approach
Elapsed time is 0.598359 seconds.
------------- With MODE + ACCUMARRAY approach
Elapsed time is 2.452778 seconds.
------------- With HISTC + ACCUMARRAY approach
Elapsed time is 2.579482 seconds.
------------- With ARRAYFUN + Sparse approach
Elapsed time is 1.455362 seconds.
slightly better, but still uses a loop :(
out=zeros(4,3);%set to zero matrix
for i = 1:max(in(:,1)); %find max in column 1, and loop for that number
ind = find(in(:,1)==i); %
out(1: size(in(ind,2),1),i)= in(ind,2);
end
don't know if you can avoid the loop...
Given two sets of d-dimensional points. How can I most efficiently compute the pairwise squared euclidean distance matrix in Matlab?
Notation:
Set one is given by a (numA,d)-matrix A and set two is given by a (numB,d)-matrix B. The resulting distance matrix shall be of the format (numA,numB).
Example points:
d = 4; % dimension
numA = 100; % number of set 1 points
numB = 200; % number of set 2 points
A = rand(numA,d); % set 1 given as matrix A
B = rand(numB,d); % set 2 given as matrix B
The usually given answer here is based on bsxfun (cf. e.g. [1]). My proposed approach is based on matrix multiplication and turns out to be much faster than any comparable algorithm I could find:
helpA = zeros(numA,3*d);
helpB = zeros(numB,3*d);
for idx = 1:d
helpA(:,3*idx-2:3*idx) = [ones(numA,1), -2*A(:,idx), A(:,idx).^2 ];
helpB(:,3*idx-2:3*idx) = [B(:,idx).^2 , B(:,idx), ones(numB,1)];
end
distMat = helpA * helpB';
Please note:
For constant d one can replace the for-loop by hardcoded implementations, e.g.
helpA(:,3*idx-2:3*idx) = [ones(numA,1), -2*A(:,1), A(:,1).^2, ... % d == 2
ones(numA,1), -2*A(:,2), A(:,2).^2 ]; % etc.
Evaluation:
%% create some points
d = 2; % dimension
numA = 20000;
numB = 20000;
A = rand(numA,d);
B = rand(numB,d);
%% pairwise distance matrix
% proposed method:
tic;
helpA = zeros(numA,3*d);
helpB = zeros(numB,3*d);
for idx = 1:d
helpA(:,3*idx-2:3*idx) = [ones(numA,1), -2*A(:,idx), A(:,idx).^2 ];
helpB(:,3*idx-2:3*idx) = [B(:,idx).^2 , B(:,idx), ones(numB,1)];
end
distMat = helpA * helpB';
toc;
% compare to pdist2:
tic;
pdist2(A,B).^2;
toc;
% compare to [1]:
tic;
bsxfun(#plus,dot(A,A,2),dot(B,B,2)')-2*(A*B');
toc;
% Another method: added 07/2014
% compare to ndgrid method (cf. Dan's comment)
tic;
[idxA,idxB] = ndgrid(1:numA,1:numB);
distMat = zeros(numA,numB);
distMat(:) = sum((A(idxA,:) - B(idxB,:)).^2,2);
toc;
Result:
Elapsed time is 1.796201 seconds.
Elapsed time is 5.653246 seconds.
Elapsed time is 3.551636 seconds.
Elapsed time is 22.461185 seconds.
For a more detailed evaluation w.r.t. dimension and number of data points follow the discussion below (#comments). It turns out that different algos should be preferred in different settings. In non time critical situations just use the pdist2 version.
Further development:
One can think of replacing the squared euclidean by any other metric based on the same principle:
help = zeros(numA,numB,d);
for idx = 1:d
help(:,:,idx) = [ones(numA,1), A(:,idx) ] * ...
[B(:,idx)' ; -ones(1,numB)];
end
distMat = sum(ANYFUNCTION(help),3);
Nevertheless, this is quite time consuming. It could be useful to replace for smaller d the 3-dimensional matrix help by d 2-dimensional matrices. Especially for d = 1 it provides a method to compute the pairwise difference by a simple matrix multiplication:
pairDiffs = [ones(numA,1), A ] * [B'; -ones(1,numB)];
Do you have any further ideas?
For squared Euclidean distance one can also use the following formula
||a-b||^2 = ||a||^2 + ||b||^2 - 2<a,b>
Where <a,b> is the dot product between a and b
nA = sum( A.^2, 2 ); %// norm of A's elements
nB = sum( B.^2, 2 ); %// norm of B's elements
distMat = bsxfun( #plus, nA, nB' ) - 2 * A * B' ;
Recently, I've been told that as of R2016b this method for computing square Euclidean distance is faster than accepted method.
I have 2 matrices: V which is square MxM, and K which is MxN. Calling the dimension across rows x and the dimension across columns t, I need to evaluate the integral (i.e sum) over both dimensions of K times a t-shifted version of V, the answer being a function of the shift (almost like a convolution, see below). The sum is defined by the following expression, where _{} denotes the summation indices, and a zero-padding of out-of-limits elements is assumed:
S(t) = sum_{x,tau}[V(x,t+tau) * K(x,tau)]
I manage to do it with a single loop, over the t dimension (vectorizing the x dimension):
% some toy matrices
V = rand(50,50);
K = rand(50,10);
[M N] = size(K);
S = zeros(1, M);
for t = 1 : N
S(1,1:end-t+1) = S(1,1:end-t+1) + sum(bsxfun(#times, V(:,t:end), K(:,t)),1);
end
I have similar expressions which I managed to evaluate without a for loop, using a combination of conv2 and\or mirroring (flipping) of a single dimension. However I can't see how to avoid a for loop in this case (despite the appeared similarity to convolution).
Steps to vectorization
1] Perform sum(bsxfun(#times, V(:,t:end), K(:,t)),1) for all columns in V against all columns in K with matrix-multiplication -
sum_mults = V.'*K
This would give us a 2D array with each column representing sum(bsxfun(#times,.. operation at each iteration.
2] Step1 gave us all possible summations and also the values to be summed are not aligned in the same row across iterations, so we need to do a bit more work before summing along rows. The rest of the work is about getting a shifted up version. For the same, you can use boolean indexing with a upper and lower triangular boolean mask. Finally, we sum along each row for the final output. So, this part of the code would look like so -
valid_mask = tril(true(size(sum_mults)));
sum_mults_shifted = zeros(size(sum_mults));
sum_mults_shifted(flipud(valid_mask)) = sum_mults(valid_mask);
out = sum(sum_mults_shifted,2);
Runtime tests -
%// Inputs
V = rand(1000,1000);
K = rand(1000,200);
disp('--------------------- With original loopy approach')
tic
[M N] = size(K);
S = zeros(1, M);
for t = 1 : N
S(1,1:end-t+1) = S(1,1:end-t+1) + sum(bsxfun(#times, V(:,t:end), K(:,t)),1);
end
toc
disp('--------------------- With proposed vectorized approach')
tic
sum_mults = V.'*K; %//'
valid_mask = tril(true(size(sum_mults)));
sum_mults_shifted = zeros(size(sum_mults));
sum_mults_shifted(flipud(valid_mask)) = sum_mults(valid_mask);
out = sum(sum_mults_shifted,2);
toc
Output -
--------------------- With original loopy approach
Elapsed time is 2.696773 seconds.
--------------------- With proposed vectorized approach
Elapsed time is 0.044144 seconds.
This might be cheating (using arrayfun instead of a for loop) but I believe this expression gives you what you want:
S = arrayfun(#(t) sum(sum( V(:,(t+1):(t+N)) .* K )), 1:(M-N), 'UniformOutput', true)