Related
I need to calculate the maximum of normalized cross correlation of million of particles. The size of the two parameters of normxcorr2 is 56*56. I can't parallelize the calculations. Is there any suggestion to speed up the code especially that I don't need all the results but only the maximum value of each cross correlation (to know the displacement)?
Example of the algorithm
%The choice of 170 particles is because in each time
%the code detects 170 particles, so over 10000 images it's 1 700 000 particles
particle_1=rand(54,54,170);
particle_2=rand(56,56,170);
for i=1:170
C=normxcorr2(particle_1(:,:,i),particle_2(:,:,i));
L(i)=max(C(:));
end
I don't have MATLAB so I ran the following code on this site: https://www.tutorialspoint.com/execute_matlab_online.php which is actually octave. So I implemented "naive" normalized cross correlation and indeed for these small images sizes the naive performs better:
Elapsed time is 2.62645 seconds - for normxcorr2
Elapsed time is 0.199034 seconds - for my naive_normxcorr2
The code is based on the article http://scribblethink.org/Work/nvisionInterface/nip.pdf which describes how to calculate the standard deviation needed for the normalization in an efficient way using integral image, this is the box_corr function.
Also, MATLAB's normxcorr2 returns a padded image so I took the max on the unpadded part.
pkg load image
function [N] = naive_corr(pat,img)
[n,m] = size(img);
[np,mp] = size(pat);
N = zeros(n-np+1,m-mp+1);
for i = 1:n-np+1
for j = 1:m-mp+1
N(i,j) = sum(dot(pat,img(i:i+np-1,j:j+mp-1)));
end
end
end
%w_arr the array of coefficients for the boxes
%box_arr of size [k,4] where k is the number boxes, each box represented by
%4 something ...
function [C] = box_corr2(img,box_arr,w_arr,n_p,m_p)
% construct integral image + zeros pad (for boundary problems)
I = cumsum(cumsum(img,2),1);
I = [zeros(1,size(I,2)+2); [zeros(size(I,1),1) I zeros(size(I,1),1)]; zeros(1,size(I,2)+2)];
% initialize result matrix
[n,m] = size(img);
C = zeros(n-n_p+1,m-m_p+1);
%C = zeros(n,m);
jump_x = 1;
jump_y = 1;
x_start = ceil(n_p/2);
x_end = n-x_start+mod(n_p,2);
x_span = x_start:jump_x:x_end;
y_start = ceil(m_p/2);
y_end = m-y_start+mod(m_p,2);
y_span = y_start:jump_y:y_end;
arr_a = box_arr(:,1) - x_start;
arr_b = box_arr(:,2) - x_start+1;
arr_c = box_arr(:,3) - y_start;
arr_d = box_arr(:,4) - y_start+1;
% cumulate box responses
k = size(box_arr,1); % == numel(w_arr)
for i = 1:k
a = arr_a(i);
b = arr_b(i);
c = arr_c(i);
d = arr_d(i);
C = C ...
+ w_arr(i) * ( I(x_span+b,y_span+d) ...
- I(x_span+b,y_span+c) ...
- I(x_span+a,y_span+d) ...
+ I(x_span+a,y_span+c) );
end
end
function [NCC] = naive_normxcorr2(temp,img)
[n_p,m_p]=size(temp);
M = n_p*m_p;
% compute template mean & std
temp_mean = mean(temp(:));
temp = temp - temp_mean;
temp_std = sqrt(sum(temp(:).^2)/M);
% compute windows' mean & std
wins_mean = box_corr2(img,[1,n_p,1,m_p],1/M, n_p,m_p);
wins_mean2 = box_corr2(img.^2,[1,n_p,1,m_p],1/M,n_p,m_p);
wins_std = real(sqrt(wins_mean2 - wins_mean.^2));
NCC_naive = naive_corr(temp,img);
NCC = NCC_naive ./ (M .* temp_std .* wins_std);
end
n = 170;
particle_1=rand(54,54,n);
particle_2=rand(56,56,n);
[n_p1,m_p1,c_p1]=size(particle_1);
[n_p2,m_p2,c_p2]=size(particle_2);
L1 = zeros(n,1);
L2 = zeros (n,1);
tic
for i=1:n
C1=normxcorr2(particle_1(:,:,i),particle_2(:,:,i));
C1_unpadded = C1(n_p1:n_p2 , m_p1:m_p2);
L1(i)=max(C1_unpadded(:));
end
toc
tic
for i=1:n
C2=naive_normxcorr2(particle_1(:,:,i),particle_2(:,:,i));
L2(i)=max(C2(:));
end
toc
I want to get exp() of a large matrix (A) with values that repeat at different indices. To speed-up the exp() operation I only perform it on the unique values of A and then reassemble the matrix. However the reassembly of the matrix is quite slow. The following code provides a working example:
% defintion of a grid
gridSp = 5:5:35*5;
X = repmat(gridSp,35,1);
Z = repmat(gridSp',1,35);
% calculation of the distances
locMat = [X(:) Z(:)];
dist=sqrt(bsxfun(#minus,locMat(:,1),locMat(:,1)').^2 +...
bsxfun(#minus,locMat(:,2),locMat(:,2)').^2);
sizeDist = size(dist);
uniqueDist = unique(dist,'stable');
[~, Locb] = ismember(dist,uniqueDist);
nn_A = exp(1i*2*pi*rand(sizeDist(1),100));
H_A = zeros(size(nn_A));
freq = linspace(10^-3,10,100);
psdA = 4096*length(freq).*10.*4.*22.6./((1 + 6.*freq*22.6).^(5/3));
for jj=1:100
b = exp(-8.8*uniqueDist*sqrt((freq(jj)/15).^2 + 10^-7));
b = b.*psdA(jj);
A = b(Locb);
droptol = max(A(:))*10^-10;
if min(A(:))<droptol
A = sparse(A);
HH_A = ichol(A,struct('type','ict','shape','lower','droptol',droptol));
else
HH_A = chol(A,'lower');
end
H_A(:,jj) = HH_A*nn_A(:,jj);
end
Especially the reassembly of the matrix
A = b(Locb);
and the conversion of the matrix to sparse
A = sparse(A);
in the last for-loop take up a lot of time. Is there a quicker way to do this? Interestingly:
B = A + A;
is much faster than
A = b(Locb);
I have to perfom these operations far more often than the 100 iterations in the example.
Here a condensed version of the code up on request (below).
% defintion of a grid
gridSp = 5:5:28*5;
X = repmat(gridSp,35,1);
Z = repmat(gridSp',1,35);
% calculation of the distances
locMat = [X(:) Z(:)];
dist=sqrt(bsxfun(#minus,locMat(:,1),locMat(:,1)').^2 +bsxfun(#minus,locMat(:,2),locMat(:,2)').^2);
uniqueDist = unique(dist,'stable');
[~, Locb] = ismember(dist,uniqueDist);
for jj=1:100
b = exp(jj.*uniqueDist);
A = b(Locb);
end
In your example, the dimension of dist is just 980 x 980 in which case you would be better off to just perform a dense matrix operation, i.e.
for jj=1:100
A=exp(jj*dist);
end
which is 2 times faster than
for jj=1:100
b = exp(jj.*uniqueDist);
A = b(Locb);
end
for your given example.
Can anyone help vectorize this Matlab code? The specific problem is the sum and bessel function with vector inputs.
Thank you!
N = 3;
rho_g = linspace(1e-3,1,N);
phi_g = linspace(0,2*pi,N);
n = 1:3;
tau = [1 2.*ones(1,length(n)-1)];
for ii = 1:length(rho_g)
for jj = 1:length(phi_g)
% Coordinates
rho_o = rho_g(ii);
phi_o = phi_g(jj);
% factors
fc = cos(n.*(phi_o-phi_s));
fs = sin(n.*(phi_o-phi_s));
Ez_t(ii,jj) = sum(tau.*besselj(n,k(3)*rho_s).*besselh(n,2,k(3)*rho_o).*fc);
end
end
You could try to vectorize this code, which might be possible with some bsxfun or so, but it would be hard to understand code, and it is the question if it would run any faster, since your code already uses vector math in the inner loop (even though your vectors only have length 3). The resulting code would become very difficult to read, so you or your colleague will have no idea what it does when you have a look at it in 2 years time.
Before wasting time on vectorization, it is much more important that you learn about loop invariant code motion, which is easy to apply to your code. Some observations:
you do not use fs, so remove that.
the term tau.*besselj(n,k(3)*rho_s) does not depend on any of your loop variables ii and jj, so it is constant. Calculate it once before your loop.
you should probably pre-allocate the matrix Ez_t.
the only terms that change during the loop are fc, which depends on jj, and besselh(n,2,k(3)*rho_o), which depends on ii. I guess that the latter costs much more time to calculate, so it better to not calculate this N*N times in the inner loop, but only N times in the outer loop. If the calculation based on jj would take more time, you could swap the for-loops over ii and jj, but that does not seem to be the case here.
The result code would look something like this (untested):
N = 3;
rho_g = linspace(1e-3,1,N);
phi_g = linspace(0,2*pi,N);
n = 1:3;
tau = [1 2.*ones(1,length(n)-1)];
% constant part, does not depend on ii and jj, so calculate only once!
temp1 = tau.*besselj(n,k(3)*rho_s);
Ez_t = nan(length(rho_g), length(phi_g)); % preallocate space
for ii = 1:length(rho_g)
% calculate stuff that depends on ii only
rho_o = rho_g(ii);
temp2 = besselh(n,2,k(3)*rho_o);
for jj = 1:length(phi_g)
phi_o = phi_g(jj);
fc = cos(n.*(phi_o-phi_s));
Ez_t(ii,jj) = sum(temp1.*temp2.*fc);
end
end
Initialization -
N = 3;
rho_g = linspace(1e-3,1,N);
phi_g = linspace(0,2*pi,N);
n = 1:3;
tau = [1 2.*ones(1,length(n)-1)];
Nested loops form (Copy from your code and shown here for comparison only) -
for ii = 1:length(rho_g)
for jj = 1:length(phi_g)
% Coordinates
rho_o = rho_g(ii);
phi_o = phi_g(jj);
% factors
fc = cos(n.*(phi_o-phi_s));
fs = sin(n.*(phi_o-phi_s));
Ez_t(ii,jj) = sum(tau.*besselj(n,k(3)*rho_s).*besselh(n,2,k(3)*rho_o).*fc);
end
end
Vectorized solution -
%%// Term - 1
term1 = repmat(tau.*besselj(n,k(3)*rho_s),[N*N 1]);
%%// Term - 2
[n1,rho_g1] = meshgrid(n,rho_g);
term2_intm = besselh(n1,2,k(3)*rho_g1);
term2 = transpose(reshape(repmat(transpose(term2_intm),[N 1]),N,N*N));
%%// Term -3
angle1 = repmat(bsxfun(#times,bsxfun(#minus,phi_g,phi_s')',n),[N 1]);
fc = cos(angle1);
%%// Output
Ez_t = sum(term1.*term2.*fc,2);
Ez_t = transpose(reshape(Ez_t,N,N));
Points to note about this vectorization or code simplification –
‘fs’ doesn’t change the output of the script, Ez_t, so it could be removed for now.
The output seems to be ‘Ez_t’,which requires three basic terms in the code as –
tau.*besselj(n,k(3)*rho_s), besselh(n,2,k(3)*rho_o) and fc. These are calculated separately for vectorization as terms1,2 and 3 respectively.
All these three terms appear to be of 1xN sizes. Our aim thus becomes to calculate these three terms without loops. Now, the two loops run for N times each, thus giving us a total loop count of NxN. Thus, we must have NxN times the data in each such term as compared to when these terms were inside the nested loops.
This is basically the essence of the vectorization done here, as the three terms are represented by ‘term1’,’term2’ and ‘fc’ itself.
In order to give a self-contained answer, I'll copy the original initialization
N = 3;
rho_g = linspace(1e-3,1,N);
phi_g = linspace(0,2*pi,N);
n = 1:3;
tau = [1 2.*ones(1,length(n)-1)];
and generate some missing data (k(3) and rho_s and phi_s in the dimension of n)
rho_s = rand(size(n));
phi_s = rand(size(n));
k(3) = rand(1);
then you can compute the same Ez_t with multidimensional arrays:
[RHO_G, PHI_G, N] = meshgrid(rho_g, phi_g, n);
[~, ~, TAU] = meshgrid(rho_g, phi_g, tau);
[~, ~, RHO_S] = meshgrid(rho_g, phi_g, rho_s);
[~, ~, PHI_S] = meshgrid(rho_g, phi_g, phi_s);
FC = cos(N.*(PHI_G - PHI_S));
FS = sin(N.*(PHI_G - PHI_S)); % not used
EZ_T = sum(TAU.*besselj(N, k(3)*RHO_S).*besselh(N, 2, k(3)*RHO_G).*FC, 3).';
You can check afterwards that both matrices are the same
norm(Ez_t - EZ_T)
I am doing a very large calculation (atmospheric absorption) that has a lot of individual narrow peaks that all get added up at the end. For each peak, I have pre-calculated the range over which the value of the peak shape function is above my chosen threshold, and I am then going line by line and adding the peaks to my spectrum. A minimum example is given below:
X = 1:1e7;
K = numel(a); % count the number of peaks I have.
spectrum = zeros(size(X));
for k = 1:K
grid = X >= rng(1,k) & X <= rng(2,k);
spectrum(grid) = spectrum(grid) + peakfn(X(grid),a(k),b(k),c(k)]);
end
Here, each peak has some parameters that define the position and shape (a,b,c), and a range over which to do the calculation (rng). This works great, and on my machine it benchmarks at around 220 seconds to do a complete data set. However, I have a 4 core machine and I would eventually like to run this on a cluster, so I'd like to parallelize it and make it scaleable.
Because each loop relies on the results of the previous iteration, I cannot use parfor, so I am taking my first step into learning how to use spmd blocks. My first try looked like this:
X = 1:1e7;
cores = matlabpool('size');
K = numel(a);
spectrum = zeros(size(X),cores);
spmd
n = labindex:cores:K
N = numel(n);
for k = 1:N
grid = X >= rng(1,n(k)) & X <= rng(2,n(k));
spectrum(grid,labindex) = spectrum(grid,labindex) + peakfn(X(grid),a(n(k)),b(n(k)),c(n(k))]);
end
end
finalSpectrum = sum(spectrum,2);
This almost works. The program crashes at the last line because spectrum is of type Composite, and the documentation for 2013a is spotty on how to turn Composite data into a matrix (cell2mat does not work). This also does not scale well because the more cores I have, the larger the matrix is, and that large matrix has to get copied to each worker, which then ignores most of the data. Question 1: how do I turn a Composite data type into a useable array?
The second thing I tried was to use a codistributed array.
spmd
spectrum = codistributed.zeros(K,cores);
disp(size(getLocalPart(spectrum)))
end
This tells me that each worker has a single vector of size [K 1], which I believe is what I want, but when I try to then meld the above methods
spmd
spectrum = codistributed.zeros(K,cores);
n = labindex:cores:K
N = numel(n);
for k = 1:N
grid = X >= rng(1,n(k)) & X <= rng(2,n(k));
spectrum(grid) = spectrum(grid) + peakfn(X(grid),a(n(k)),b(n(k)),c(n(k))]); end
finalSpectrum = gather(spectrum);
end
finalSpectrum = sum(finalSpectrum,2);
I get Matrix dimensions must agree errors. Since it's in a parallel block, I can't use my normal debugging crutch of stepping through the loop and seeing what the size of each block is at each point to see what's going on. Question 2: what is the proper way to index into and out of a codistributed array in an spmd block?
Regarding question#1, the Composite variable in the client basically refers to a non-distributed variant array stored on the workers. You can access the array from each worker by {}-indexing using its corresponding labindex (e.g: spectrum{1}, spectrum{2}, ..).
For your code that would be: finalSpectrum = sum(cat(2,spectrum{:}), 2);
Now I tried this problem myself using random data. Below are three implementations to compare (see here to understand the difference between distributed and nondistributed arrays). First we start with the common data:
len = 100; % spectrum length
K = 10; % number of peaks
X = 1:len;
% random position and shape parameters
a = rand(1,K); b = rand(1,K); c = rand(1,K);
% random peak ranges (lower/upper thresholds)
ranges = sort(randi([1 len], [2 K]));
% dummy peakfn() function
fcn = #(x,a,b,c) x+a+b+c;
% prepare a pool of MATLAB workers
matlabpool open
1) Serial for-loop:
spectrum = zeros(size(X));
for i=1:size(ranges,2)
r = ranges(:,i);
idx = (r(1) <= X & X <= r(2));
spectrum(idx) = spectrum(idx) + fcn(X(idx), a(i), b(i), c(i));
end
s1 = spectrum;
clear spectrum i r idx
2) SPMD with Composite array
spmd
spectrum = zeros(1,len);
ind = labindex:numlabs:K;
for i=1:numel(ind)
r = ranges(:,ind(i));
idx = (r(1) <= X & X <= r(2));
spectrum(idx) = spectrum(idx) + ...
feval(fcn, X(idx), a(ind(i)), b(ind(i)), c(ind(i)));
end
end
s2 = sum(vertcat(spectrum{:}));
clear spectrum i r idx ind
3) SPMD with co-distributed array
spmd
spectrum = zeros(numlabs, len, codistributor('1d',1));
ind = labindex:numlabs:K;
for i=1:numel(ind)
r = ranges(:,ind(i));
idx = (r(1) <= X & X <= r(2));
spectrum(labindex,idx) = spectrum(labindex,idx) + ...
feval(fcn, X(idx), a(ind(i)), b(ind(i)), c(ind(i)));
end
end
s3 = sum(gather(spectrum));
clear spectrum i r idx ind
All three results should be equal (to within an acceptably small margin of error)
>> max([max(s1-s2), max(s1-s3), max(s2-s3)])
ans =
2.8422e-14
I have implemented cosine similarity in Matlab like this. In fact, I have a two-dimensional 50-by-50 matrix. To obtain a cosine should I compare items in a line by line form.
for j = 1:50
x = dat(j,:);
for i = j+1:50
y = dat(i,:);
c = dot(x,y);
sim = c/(norm(x,2)*norm(y,2));
end
end
Is this correct?
and The question is this: wath is the complexity or O(n) in this state?
Just a note on an efficient implementation of the same thing using vectorized and matrix-wise operations (which are optimized in MATLAB). This can have huge time savings for large matrices:
dat = randn(50, 50);
OP (double-for) implementation:
sim = zeros(size(dat));
nRow = size(dat,1);
for j = 1:nRow
x = dat(j, :);
for i = j+1:nRow
y = dat(i, :);
c = dot(x, y);
sim(j, i) = c/(norm(x,2)*norm(y,2));
end
end
Vectorized implementation:
normDat = sqrt(sum(dat.^2, 2)); % L2 norm of each row
datNorm = bsxfun(#rdivide, dat, normDat); % normalize each row
dotProd = datNorm*datNorm'; % dot-product vectorized (redundant!)
sim2 = triu(dotProd, 1); % keep unique upper triangular part
Comparisons for 1000 x 1000 matrix: (MATLAB 2013a, x64, Intel Core i7 960 # 3.20GHz)
Elapsed time is 34.103095 seconds.
Elapsed time is 0.075208 seconds.
sum(sum(sim-sim2))
ans =
-1.224314766369880e-14
Better end with 49. Maybe you should also add an index to sim?
for j = 1:49
x = dat(j,:);
for i = j+1:50
y = dat(i,:);
c = dot(x,y);
sim(j) = c/(norm(x,2)*norm(y,2));
end
end
The complexity should be roughly like o(n^2), isn't it?
Maybe you should have a look at correlation functions ... I don't get what you want to write exactly, but it looks like you want to do something similar. There are built-in correlation functions in Matlab.