Related
I compute the regression map of a time series A(t) on a field B(x,y,t) in the following way:
A=1:10; %time
B=rand(100,100,10); %x,y,time
rc=nan(size(B,1),size(B,2));
for ii=size(B,1)
for jj=1:size(B,2)
tmp = cov(A,squeeze(B(ii,jj,:))); %covariance matrix
rc(ii,jj) = tmp(1,2); %covariance A and B
end
end
rc = rc/var(A); %regression coefficient
Is there a way to vectorize/speed up code? Or maybe some built-in function that I did not know of to achieve the same result?
In order to vectorize this algorithm, you would have to "get your hands dirty" and compute the covariance yourself. If you take a look inside cov you'll see that it has many lines of input checking and very few lines of actual computation, to summarize the critical steps:
y = varargin{1};
x = x(:);
y = y(:);
x = [x y];
[m,~] = size(x);
denom = m - 1;
xc = x - sum(x,1)./m; % Remove mean
c = (xc' * xc) ./ denom;
To simplify the above somewhat:
x = [x(:) y(:)];
m = size(x,1);
xc = x - sum(x,1)./m;
c = (xc' * xc) ./ (m - 1);
Now this is something that is fairly straightforward to vectorize...
function q51466884
A = 1:10; %time
B = rand(200,200,10); %x,y,time
%% Test Equivalence:
assert( norm(sol1-sol2) < 1E-10);
%% Benchmark:
disp([timeit(#sol1), timeit(#sol2)]);
%%
function rc = sol1()
rc=nan(size(B,1),size(B,2));
for ii=1:size(B,1)
for jj=1:size(B,2)
tmp = cov(A,squeeze(B(ii,jj,:))); %covariance matrix
rc(ii,jj) = tmp(1,2); %covariance A and B
end
end
rc = rc/var(A); %regression coefficient
end
function rC = sol2()
m = numel(A);
rB = reshape(B,[],10).'; % reshape
% Center:
cA = A(:) - sum(A)./m;
cB = rB - sum(rB,1)./m;
% Multiply:
rC = reshape( (cA.' * cB) ./ (m-1), size(B(:,:,1)) ) ./ var(A);
end
end
I get these timings: [0.5381 0.0025] which means we saved two orders of magnitude in the runtime :)
Note that a big part of optimizing the algorithm is assuming you don't have any "strangeness" in your data, like NaN values etc. Take a look inside cov.m to see all the checks that we skipped.
I need to do function that works like this :
N1 = size(X,1);
N2 = size(Xtrain,1);
Dist = zeros(N1,N2);
for i=1:N1
for j=1:N2
Dist(i,j)=D-sum(X(i,:)==Xtrain(j,:));
end
end
(X and Xtrain are sparse logical matrixes)
It works fine and passes the tests, but I believe it's not very optimal and well-written solution.
How can I improve that function using some built Matlab functions? I'm absolutely new to Matlab, so I don't know if there really is an opportunity to make it better somehow.
You wanted to learn about vectorization, here some code to study comparing different implementations of this pair-wise distance.
First we build two binary matrices as input (where each row is an instance):
m = 5;
n = 4;
p = 3;
A = double(rand(m,p) > 0.5);
B = double(rand(n,p) > 0.5);
1. double-loop over each pair of instances
D0 = zeros(m,n);
for i=1:m
for j=1:n
D0(i,j) = sum(A(i,:) ~= B(j,:)) / p;
end
end
2. PDIST2
D1 = pdist2(A, B, 'hamming');
3. single-loop over each instance against all other instances
D2 = zeros(m,n);
for i=1:n
D2(:,i) = sum(bsxfun(#ne, A, B(i,:)), 2) ./ p;
end
4. vectorized with grid indexing, all against all
D3 = zeros(m,n);
[x,y] = ndgrid(1:m,1:n);
D3(:) = sum(A(x(:),:) ~= B(y(:),:), 2) ./ p;
5. vectorized in third dimension, all against all
D4 = sum(bsxfun(#ne, A, reshape(B.',[1 p n])), 2) ./ p;
D4 = permute(D4, [1 3 2]);
Finally we compare all methods are equal
assert(isequal(D0,D1,D2,D3,D4))
I have an n x p matrix - mX which is composed of n points in R^p.
I have another m x p matrix - mY which is composed of m reference points in R^p.
I would like to create an n x m matrix - mD which is the Mahalanobis Distance matrix.
D(i, j) means the Mahalanobis Distance between point i in mX, mX(i, :) and point j in mY, mY(j, :).
Namely, is computes the following:
mD(i, j) = (mX(i, :) - mY(j, :)) * inv(mC) * (mX(i, :) - mY(j, :)).';
Where mC is the given Mahalanobis Distance PSD Matrix.
It is easy to be done in a loop, is there a way to vectorize it?
Namely, is the a function which its inputs are mX, mY and mC and its output is mD and fully vectorized without using any MATLAB toolbox?
Thank You.
Approach #1
Assuming infinite resources, here's one vectorized solution using bsxfun and matrix-multiplication -
A = reshape(bsxfun(#minus,permute(mX,[1 3 2]),permute(mY,[3 1 2])),[],p);
out = reshape(diag(A*inv(mC)*A.'),n,m);
Approach #2
Here's a comprise solution trying to reduce the loop complexity -
A = reshape(bsxfun(#minus,permute(mX,[1 3 2]),permute(mY,[3 1 2])),[],p);
imC = inv(mC);
out = zeros(n*m,1);
for ii = 1:n*m
out(ii) = A(ii,:)*imC*A(ii,:).';
end
out = reshape(out,n,m);
Sample run -
>> n = 3; m = 4; p = 5;
mX = rand(n,p);
mY = rand(m,p);
mC = rand(p,p);
imC = inv(mC);
>> %// Original solution
for i = 1:n
for j = 1:m
mD(i, j) = (mX(i, :) - mY(j, :)) * inv(mC) * (mX(i, :) - mY(j, :)).'; %//'
end
end
>> mD
mD =
-8.4256 10.032 2.8929 7.1762
-44.748 -4.3851 -13.645 -9.6702
-4.5297 3.2928 0.11132 2.5998
>> %// Approach #1
A = reshape(bsxfun(#minus,permute(mX,[1 3 2]),permute(mY,[3 1 2])),[],p);
out = reshape(diag(A*inv(mC)*A.'),n,m); %//'
>> out
out =
-8.4256 10.032 2.8929 7.1762
-44.748 -4.3851 -13.645 -9.6702
-4.5297 3.2928 0.11132 2.5998
>> %// Approach #2
A = reshape(bsxfun(#minus,permute(mX,[1 3 2]),permute(mY,[3 1 2])),[],p);
imC = inv(mC);
out1 = zeros(n*m,1);
for ii = 1:n*m
out1(ii) = A(ii,:)*imC*A(ii,:).'; %//'
end
out1 = reshape(out1,n,m);
>> out1
out1 =
-8.4256 10.032 2.8929 7.1762
-44.748 -4.3851 -13.645 -9.6702
-4.5297 3.2928 0.11132 2.5998
Instead if you had :
mD(j, i) = (mX(i, :) - mY(j, :)) * inv(mC) * (mX(i, :) - mY(j, :)).';
The solutions would translate to the versions listed next.
Approach #1
A = reshape(bsxfun(#minus,permute(mY,[1 3 2]),permute(mX,[3 1 2])),[],p);
out = reshape(diag(A*inv(mC)*A.'),m,n);
Approach #2
A = reshape(bsxfun(#minus,permute(mY,[1 3 2]),permute(mX,[3 1 2])),[],p);
imC = inv(mC);
out1 = zeros(m*n,1);
for i = 1:n*m
out(i) = A(i,:)*imC*A(i,:).'; %//'
end
out = reshape(out,m,n);
Sample run -
>> n = 3; m = 4; p = 5;
mX = rand(n,p); mY = rand(m,p); mC = rand(p,p); imC = inv(mC);
>> %// Original solution
for i = 1:n
for j = 1:m
mD(j, i) = (mX(i, :) - mY(j, :)) * inv(mC) * (mX(i, :) - mY(j, :)).'; %//'
end
end
>> mD
mD =
0.81755 0.33205 0.82254
1.7086 1.3363 2.4209
0.36495 0.78394 -0.33097
0.17359 0.3889 -1.0624
>> %// Approach #1
A = reshape(bsxfun(#minus,permute(mY,[1 3 2]),permute(mX,[3 1 2])),[],p);
out = reshape(diag(A*inv(mC)*A.'),m,n); %//'
>> out
out =
0.81755 0.33205 0.82254
1.7086 1.3363 2.4209
0.36495 0.78394 -0.33097
0.17359 0.3889 -1.0624
>> %// Approach #2
A = reshape(bsxfun(#minus,permute(mY,[1 3 2]),permute(mX,[3 1 2])),[],p);
imC = inv(mC);
out1 = zeros(m*n,1);
for i = 1:n*m
out1(i) = A(i,:)*imC*A(i,:).'; %//'
end
out1 = reshape(out1,m,n);
>> out1
out1 =
0.81755 0.33205 0.82254
1.7086 1.3363 2.4209
0.36495 0.78394 -0.33097
0.17359 0.3889 -1.0624
Here is one solution that eliminates one loop
function d = mahalanobis(mX, mY)
n = size(mX, 2);
m = size(mY, 2);
data = [mX, mY];
mc = cov(transpose(data));
dist = zeros(n,m);
for i = 1 : n
diff = repmat(mX(:,i), 1, m) - mY;
dist(i,:) = sum((mc\diff).*diff , 1);
end
d = sqrt(dist);
end
You would invoke it as:
d = mahalanobis(transpose(X),transpose(Y))
Reduce to L2
It seems that Mahalanobis Distance can be reduced to ordinary L2 distance if you are allowed to preprocess matrix mC and you are not afraid of numerical differences.
First of all, compute Cholesky decomposition of mC:
mR = chol(mC) % C = R^t * R, where R is upper-triangular
Now we can use these factors to reformulate Mahalanobis Distance:
(Xi-Yj) * inv(C) * (Xi-Yj)^t = || (Xi-Yj) inv(R) ||^2 = ||TXi - TYj||^2
where: TXi = Xi * inv(R)
TYj = Yj * inv(R)
So the idea is to transform points Xi, Yj to TXi, TYj first, and then compute euclidean distances between them. Here is the algorithm outline:
Compute mR - Cholesky factor of covariance matrix mC (takes O(p^3) time).
Invert triangular matrix mR (takes O(p^3) time).
Multiply both mX and mY by inv(mR) on the right (takes O(p^2 (m+n)) time).
Compute squared L2 distances between pairs of points (takes O(m n p) time).
Total time is O(m n p + (m + n) p^2 + p^3) versus original O(m n p^2). It should work faster when 1 << p << n,m. In such case step 4 would takes most of the time and should be vectorized.
Vectorization
I have little experience of MATLAB, but quite a lot of SIMD vectorization on x86 CPUs. In raw computations, it would be enough to vectorize along one sufficiently large array dimension, and make trivial loops for the other dimensions.
If you expect p to be large enough, it may probably be OK to vectorize along coordinates of points, and make two nested loops for i <= n and j <= m. That's similar to what #Daniel posted.
If p is not sufficiently large, you can vectorize along one of the point sequences instead. This would be similar to solution posted by #dpmcmlxxvi: you have to subtract single row of one matrix from all the rows of the second matrix, then compute squared norms of the resulting rows. Repeat n times (
or m times).
As for me, full vectorization (which means rewriting with matrix operations instead of loops in MATLAB) does not sound like a clever performance goal. Most likely partially vectorized solutions would be optimally fast.
I came to the conclusion that vectorizing this problem is not efficient. My best idea for vectorizing this problem would require m x n x p x p working memory, at least if everything is processed at once. This means with n=m=p=152 the code would already require 4GB Ram. At these dimensions, my system can run the loop in less than a second:
mD=zeros(size(mX,1),size(mY,1));
ImC=inv(mC);
for i=1:size(mX,1)
for j=1:size(mY,1)
d=mX(i, :) - mY(j, :);
mD(i, j) = (d) * ImC * (d).';
end
end
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've found myself needing to do a least-squares (or similar matrix-based operation) for every pixel in an image. Every pixel has a set of numbers associated with it, and so it can be arranged as a 3D matrix.
(This next bit can be skipped)
Quick explanation of what I mean by least-squares estimation :
Let's say we have some quadratic system that is modeled by Y = Ax^2 + Bx + C and we're looking for those A,B,C coefficients. With a few samples (at least 3) of X and the corresponding Y, we can estimate them by:
Arrange the (lets say 10) X samples into a matrix like X = [x(:).^2 x(:) ones(10,1)];
Arrange the Y samples into a similar matrix: Y = y(:);
Estimate the coefficients A,B,C by solving: coeffs = (X'*X)^(-1)*X'*Y;
Try this on your own if you want:
A = 5; B = 2; C = 1;
x = 1:10;
y = A*x(:).^2 + B*x(:) + C + .25*randn(10,1); % added some noise here
X = [x(:).^2 x(:) ones(10,1)];
Y = y(:);
coeffs = (X'*X)^-1*X'*Y
coeffs =
5.0040
1.9818
0.9241
START PAYING ATTENTION AGAIN IF I LOST YOU THERE
*MAJOR REWRITE*I've modified to bring it as close to the real problem that I have and still make it a minimum working example.
Problem Setup
%// Setup
xdim = 500;
ydim = 500;
ncoils = 8;
nshots = 4;
%// matrix size for each pixel is ncoils x nshots (an overdetermined system)
%// each pixel has a matrix stored in the 3rd and 4rth dimensions
regressor = randn(xdim,ydim, ncoils,nshots);
regressand = randn(xdim, ydim,ncoils);
So my problem is that I have to do a (X'*X)^-1*X'*Y (least-squares or similar) operation for every pixel in an image. While that itself is vectorized/matrixized the only way that I have to do it for every pixel is in a for loop, like:
Original code style
%// Actual work
tic
estimate = zeros(xdim,ydim);
for col=1:size(regressor,2)
for row=1:size(regressor,1)
X = squeeze(regressor(row,col,:,:));
Y = squeeze(regressand(row,col,:));
B = X\Y;
% B = (X'*X)^(-1)*X'*Y; %// equivalently
estimate(row,col) = B(1);
end
end
toc
Elapsed time = 27.6 seconds
EDITS in reponse to comments and other ideas
I tried some things:
1. Reshaped into a long vector and removed the double for loop. This saved some time.
2. Removed the squeeze (and in-line transposing) by permute-ing the picture before hand: This save alot more time.
Current example:
%// Actual work
tic
estimate2 = zeros(xdim*ydim,1);
regressor_mod = permute(regressor,[3 4 1 2]);
regressor_mod = reshape(regressor_mod,[ncoils,nshots,xdim*ydim]);
regressand_mod = permute(regressand,[3 1 2]);
regressand_mod = reshape(regressand_mod,[ncoils,xdim*ydim]);
for ind=1:size(regressor_mod,3) % for every pixel
X = regressor_mod(:,:,ind);
Y = regressand_mod(:,ind);
B = X\Y;
estimate2(ind) = B(1);
end
estimate2 = reshape(estimate2,[xdim,ydim]);
toc
Elapsed time = 2.30 seconds (avg of 10)
isequal(estimate2,estimate) == 1;
Rody Oldenhuis's way
N = xdim*ydim*ncoils; %// number of columns
M = xdim*ydim*nshots; %// number of rows
ii = repmat(reshape(1:N,[ncoils,xdim*ydim]),[nshots 1]); %//column indicies
jj = repmat(1:M,[ncoils 1]); %//row indicies
X = sparse(ii(:),jj(:),regressor_mod(:));
Y = regressand_mod(:);
B = X\Y;
B = reshape(B(1:nshots:end),[xdim ydim]);
Elapsed time = 2.26 seconds (avg of 10)
or 2.18 seconds (if you don't include the definition of N,M,ii,jj)
SO THE QUESTION IS:
Is there an (even) faster way?
(I don't think so.)
You can achieve a ~factor of 2 speed up by precomputing the transposition of X. i.e.
for x=1:size(picture,2) % second dimension b/c already transposed
X = picture(:,x);
XX = X';
Y = randn(n_timepoints,1);
%B = (X'*X)^-1*X'*Y; ;
B = (XX*X)^-1*XX*Y;
est(x) = B(1);
end
Before: Elapsed time is 2.520944 seconds.
After: Elapsed time is 1.134081 seconds.
EDIT:
Your code, as it stands in your latest edit, can be replaced by the following
tic
xdim = 500;
ydim = 500;
n_timepoints = 10; % for example
% Actual work
picture = randn(xdim,ydim,n_timepoints);
picture = reshape(picture, [xdim*ydim,n_timepoints])'; % note transpose
YR = randn(n_timepoints,size(picture,2));
% (XX*X).^-1 = sum(picture.*picture).^-1;
% XX*Y = sum(picture.*YR);
est = sum(picture.*picture).^-1 .* sum(picture.*YR);
est = reshape(est,[xdim,ydim]);
toc
Elapsed time is 0.127014 seconds.
This is an order of magnitude speed up on the latest edit, and the results are all but identical to the previous method.
EDIT2:
Okay, so if X is a matrix, not a vector, things are a little more complicated. We basically want to precompute as much as possible outside of the for-loop to keep our costs down. We can also get a significant speed-up by computing XT*X manually - since the result will always be a symmetric matrix, we can cut a few corners to speed things up. First, the symmetric multiplication function:
function XTX = sym_mult(X) % X is a 3-d matrix
n = size(X,2);
XTX = zeros(n,n,size(X,3));
for i=1:n
for j=i:n
XTX(i,j,:) = sum(X(:,i,:).*X(:,j,:));
if i~=j
XTX(j,i,:) = XTX(i,j,:);
end
end
end
Now the actual computation script
xdim = 500;
ydim = 500;
n_timepoints = 10; % for example
Y = randn(10,xdim*ydim);
picture = randn(xdim,ydim,n_timepoints); % 500x500x10
% Actual work
tic % start timing
picture = reshape(picture, [xdim*ydim,n_timepoints])';
% Here we precompute the (XT*Y) calculation to speed things up later
picture_y = [sum(Y);sum(Y.*picture)];
% initialize
est = zeros(size(picture,2),1);
picture = permute(picture,[1,3,2]);
XTX = cat(2,ones(n_timepoints,1,size(picture,3)),picture);
XTX = sym_mult(XTX); % precompute (XT*X) for speed
X = zeros(2,2); % preallocate for speed
XY = zeros(2,1);
for x=1:size(picture,2) % second dimension b/c already transposed
%For some reason this is a lot faster than X = XTX(:,:,x);
X(1,1) = XTX(1,1,x);
X(2,1) = XTX(2,1,x);
X(1,2) = XTX(1,2,x);
X(2,2) = XTX(2,2,x);
XY(1) = picture_y(1,x);
XY(2) = picture_y(2,x);
% Here we utilise the fact that A\B is faster than inv(A)*B
% We also use the fact that (A*B)*C = A*(B*C) to speed things up
B = X\XY;
est(x) = B(1);
end
est = reshape(est,[xdim,ydim]);
toc % end timing
Before: Elapsed time is 4.56 seconds.
After: Elapsed time is 2.24 seconds.
This is a speed up of about a factor of 2. This code should be extensible to X being any dimensions you want. For instance, in the case where X = [1 x x^2], you would change picture_y to the following
picture_y = [sum(Y);sum(Y.*picture);sum(Y.*picture.^2)];
and change XTX to
XTX = cat(2,ones(n_timepoints,1,size(picture,3)),picture,picture.^2);
You would also change a lot of 2s to 3s in the code, and add XY(3) = picture_y(3,x) to the loop. It should be fairly straight-forward, I believe.
Results
I sped up your original version, since your edit 3 was actually not working (and also does something different).
So, on my PC:
Your (original) version: 8.428473 seconds.
My obfuscated one-liner given below: 0.964589 seconds.
First, for no other reason than to impress, I'll give it as I wrote it:
%%// Some example data
xdim = 500;
ydim = 500;
n_timepoints = 10; % for example
estimate = zeros(xdim,ydim); %// initialization with explicit size
picture = randn(xdim,ydim,n_timepoints);
%%// Your original solution
%// (slightly altered to make my version's results agree with yours)
tic
Y = randn(n_timepoints,xdim*ydim);
ii = 1;
for x = 1:xdim
for y = 1:ydim
X = squeeze(picture(x,y,:)); %// or similar creation of X matrix
B = (X'*X)^(-1)*X' * Y(:,ii);
ii = ii+1;
%// sometimes you keep everything and do
%// estimate(x,y,:) = B(:);
%// sometimes just the first element is important and you do
estimate(x,y) = B(1);
end
end
toc
%%// My version
tic
%// UNLEASH THE FURY!!
estimate2 = reshape(sparse(1:xdim*ydim*n_timepoints, ...
builtin('_paren', ones(n_timepoints,1)*(1:xdim*ydim),:), ...
builtin('_paren', permute(picture, [3 2 1]),:))\Y(:), ydim,xdim).'; %'
toc
%%// Check for equality
max(abs(estimate(:)-estimate2(:))) % (always less than ~1e-14)
Breakdown
First, here's the version that you should actually use:
%// Construct sparse block-diagonal matrix
%// (Type "help sparse" for more information)
N = xdim*ydim; %// number of columns
M = N*n_timepoints; %// number of rows
ii = 1:N;
jj = ones(n_timepoints,1)*(1:N);
s = permute(picture, [3 2 1]);
X = sparse(ii,jj(:), s(:));
%// Compute ALL the estimates at once
estimates = X\Y(:);
%// You loop through the *second* dimension first, so to make everything
%// agree, we have to extract elements in the "wrong" order, and transpose:
estimate2 = reshape(estimates, ydim,xdim).'; %'
Here's an example of what picture and the corresponding matrix X looks like for xdim = ydim = n_timepoints = 2:
>> clc, picture, full(X)
picture(:,:,1) =
-0.5643 -2.0504
-0.1656 0.4497
picture(:,:,2) =
0.6397 0.7782
0.5830 -0.3138
ans =
-0.5643 0 0 0
0.6397 0 0 0
0 -2.0504 0 0
0 0.7782 0 0
0 0 -0.1656 0
0 0 0.5830 0
0 0 0 0.4497
0 0 0 -0.3138
You can see why sparse is necessary -- it's mostly zeros, but will grow large quickly. The full matrix would quickly consume all your RAM, while the sparse one will not consume much more than the original picture matrix does.
With this matrix X, the new problem
X·b = Y
now contains all the problems
X1 · b1 = Y1
X2 · b2 = Y2
...
where
b = [b1; b2; b3; ...]
Y = [Y1; Y2; Y3; ...]
so, the single command
X\Y
will solve all your systems at once.
This offloads all the hard work to a set of highly specialized, compiled to machine-specific code, optimized-in-every-way algorithms, rather than the interpreted, generic, always-two-steps-away from the hardware loops in MATLAB.
It should be straightforward to convert this to a version where X is a matrix; you'll end up with something like what blkdiag does, which can also be used by mldivide in exactly the same way as above.
I had a wee play around with an idea, and I decided to stick it as a separate answer, as its a completely different approach to my other idea, and I don't actually condone what I'm about to do. I think this is the fastest approach so far:
Orignal (unoptimised): 13.507176 seconds.
Fast Cholesky-decomposition method: 0.424464 seconds
First, we've got a function to quickly do the X'*X multiplication. We can speed things up here because the result will always be symmetric.
function XX = sym_mult(X)
n = size(X,2);
XX = zeros(n,n,size(X,3));
for i=1:n
for j=i:n
XX(i,j,:) = sum(X(:,i,:).*X(:,j,:));
if i~=j
XX(j,i,:) = XX(i,j,:);
end
end
end
The we have a function to do LDL Cholesky decomposition of a 3D matrix (we can do this because the (X'*X) matrix will always be symmetric) and then do forward and backwards substitution to solve the LDL inversion equation
function Y = fast_chol(X,XY)
n=size(X,2);
L = zeros(n,n,size(X,3));
D = zeros(n,n,size(X,3));
B = zeros(n,1,size(X,3));
Y = zeros(n,1,size(X,3));
% These loops compute the LDL decomposition of the 3D matrix
for i=1:n
D(i,i,:) = X(i,i,:);
L(i,i,:) = 1;
for j=1:i-1
L(i,j,:) = X(i,j,:);
for k=1:(j-1)
L(i,j,:) = L(i,j,:) - L(i,k,:).*L(j,k,:).*D(k,k,:);
end
D(i,j,:) = L(i,j,:);
L(i,j,:) = L(i,j,:)./D(j,j,:);
if i~=j
D(i,i,:) = D(i,i,:) - L(i,j,:).^2.*D(j,j,:);
end
end
end
for i=1:n
B(i,1,:) = XY(i,:);
for j=1:(i-1)
B(i,1,:) = B(i,1,:)-D(i,j,:).*B(j,1,:);
end
B(i,1,:) = B(i,1,:)./D(i,i,:);
end
for i=n:-1:1
Y(i,1,:) = B(i,1,:);
for j=n:-1:(i+1)
Y(i,1,:) = Y(i,1,:)-L(j,i,:).*Y(j,1,:);
end
end
Finally, we have the main script which calls all of this
xdim = 500;
ydim = 500;
n_timepoints = 10; % for example
Y = randn(10,xdim*ydim);
picture = randn(xdim,ydim,n_timepoints); % 500x500x10
tic % start timing
picture = reshape(pr, [xdim*ydim,n_timepoints])';
% Here we precompute the (XT*Y) calculation
picture_y = [sum(Y);sum(Y.*picture)];
% initialize
est2 = zeros(size(picture,2),1);
picture = permute(picture,[1,3,2]);
% Now we calculate the X'*X matrix
XTX = cat(2,ones(n_timepoints,1,size(picture,3)),picture);
XTX = sym_mult(XTX);
% Call our fast Cholesky decomposition routine
B = fast_chol(XTX,picture_y);
est2 = B(1,:);
est2 = reshape(est2,[xdim,ydim]);
toc
Again, this should work equally well for a Nx3 X matrix, or however big you want.
I use octave, thus I can't say anything about the resulting performance in Matlab, but would expect this code to be slightly faster:
pictureT=picture'
est=arrayfun(#(x)( (pictureT(x,:)*picture(:,x))^-1*pictureT(x,:)*randn(n_ti
mepoints,1)),1:size(picture,2));