I have a variable pth which is a cell array of dimension 1xn where n is a user input. Each of the elements in pth is itself a cell array and length(pth{k}) for k=1:n is variable (result of another function). Each element pth{k}{kk} where k=1:n and kk=1:length(pth{k}) is a 1D vector of integers/node numbers of again variable length. So to summarise, I have a variable number of variable-length vectors organised in a avriable number of cell arrays.
I would like to try and find all possible intersections when you take a vector at random from pth{1}, pth{2}, {pth{3}, etc... There are various functions on the File Exchange that seem to do that, for example this one or this one. The problem I have is you need to call the function this way:
mintersect(v1,v2,v3,...)
and I can't write all the inputs in the general case because I don't know explicitly how many there are (this would be n above). Ideally, I would like to do some thing like this;
mintersect(pth{1}{1},pth{2}{1},pth{3}{1},...,pth{n}{1})
mintersect(pth{1}{1},pth{2}{2},pth{3}{1},...,pth{n}{1})
mintersect(pth{1}{1},pth{2}{3},pth{3}{1},...,pth{n}{1})
etc...
mintersect(pth{1}{1},pth{2}{length(pth{2})},pth{3}{1},...,pth{n}{1})
mintersect(pth{1}{1},pth{2}{1},pth{3}{2},...,pth{n}{1})
etc...
keep going through all the possible combinations, but I can't write this in code. This function from the File Exchange looks like a good way to find all possible combinations but again I have the same problem with the function call with the variable number of inputs:
allcomb(1:length(pth{1}),1:length(pth{2}),...,1:length(pth{n}))
Does anybody know how to work around this issue of function calls with variable number of input arguments when you can't physically specify all the input arguments because their number is variable? This applies equally to MATLAB and Octave, hence the two tags. Any other suggestion on how to find all possible combinations/intersections when taking a vector at random from each pth{k} welcome!
EDIT 27/05/20
Thanks to Mad Physicist's answer, I have ended up using the following which works:
disp('Computing intersections for all possible paths...')
grids = cellfun(#(x) 1:numel(x), pth, 'UniformOutput', false);
idx = cell(1, numel(pth));
[idx{:}] = ndgrid(grids{:});
idx = cellfun(#(x) x(:), idx, 'UniformOutput', false);
idx = cat(2, idx{:});
valid_comb = [];
k = 1;
for ii = idx'
indices = reshape(num2cell(ii), size(pth));
selection = cellfun(#(p,k) p{k}, pth, indices, 'UniformOutput', false);
if my_intersect(selection{:})
valid_comb = [valid_comb k];
endif
k = k+1;
end
My own version is similar but uses a for loop instead of the comma-separated list:
disp('Computing intersections for all possible paths...')
grids = cellfun(#(x) 1:numel(x), pth, 'UniformOutput', false);
idx = cell(1, numel(pth));
[idx{:}] = ndgrid(grids{:});
idx = cellfun(#(x) x(:), idx, 'UniformOutput', false);
idx = cat(2, idx{:});
[n_comb,~] = size(idx);
temp = cell(n_pipes,1);
valid_comb = [];
k = 1;
for k = 1:n_comb
for kk = 1:n_pipes
temp{kk} = pth{kk}{idx(k,kk)};
end
if my_intersect(temp{:})
valid_comb = [valid_comb k];
end
end
In both cases, valid_comb has the indices of the valid combinations, which I can then retrieve using something like:
valid_idx = idx(valid_comb(1),:);
for k = 1:n_pipes
pth{k}{valid_idx(k)} % do something with this
end
When I benchmarked the two approaches with some sample data (pth being 4x1 and the 4 elements of pth being 2x1, 9x1, 8x1 and 69x1), I got the following results:
>> benchmark
Elapsed time is 51.9075 seconds.
valid_comb = 7112
Elapsed time is 66.6693 seconds.
valid_comb = 7112
So Mad Physicist's approach was about 15s faster.
I also misunderstood what mintersect did, which isn't what I wanted. I wanted to find a combination where no element present in two or more vectors, so I ended writing my version of mintersect:
function valid_comb = my_intersect(varargin)
% Returns true if a valid combination i.e. no combination of any 2 vectors
% have any elements in common
comb_idx = combnk(1:nargin,2);
[nr,nc] = size(comb_idx);
valid_comb = true;
k = 1;
% Use a while loop so that as soon as an intersection is found, the execution stops
while valid_comb && (k<=nr)
temp = intersect(varargin{comb_idx(k,1)},varargin{comb_idx(k,2)});
valid_comb = isempty(temp) && valid_comb;
k = k+1;
end
end
Couple of helpful points to construct a solution:
This post shows you how to construct a Cartesian product between arbitrary arrays using ndgrid.
cellfun accepts multiple cell arrays simultaneously, which you can use to index specific elements.
You can capture a variable number of arguments from a function using cell arrays, as shown here.
So let's get the inputs to ndgrid from your outermost array:
grids = cellfun(#(x) 1:numel(x), pth, 'UniformOutput', false);
Now you can create an index that contains the product of the grids:
index = cell(1, numel(pth));
[index{:}] = ndgrid(grids{:});
You want to make all the grids into column vectors and concatenate them sideways. The rows of that matrix will represent the Cartesian indices to select the elements of pth at each iteration:
index = cellfun(#(x) x(:), index, 'UniformOutput', false);
index = cat(2, index{:});
If you turn a row of index into a cell array, you can run it in lockstep over pth to select the correct elements and call mintersect on the result.
for i = index'
indices = num2cell(i');
selection = cellfun(#(p, i) p{i}, pth, indices, 'UniformOutput', false);
mintersect(selection{:});
end
This is written under the assumption that pth is a row array. If that is not the case, you can change the first line of the loop to indices = reshape(num2cell(i), size(pth)); for the general case, and simply indices = num2cell(i); for the column case. The key is that the cell from of indices must be the same shape as pth to iterate over it in lockstep. It is already generated to have the same number of elements.
I believe this does the trick. Calls mintersect on all possible combinations of vectors in pth{k}{kk} for k=1:n and kk=1:length(pth{k}).
Using eval and messing around with sprintf/compose a bit. Note that typically the use of eval is very much discouraged. Can add more comments if this is what you need.
% generate some data
n = 5;
pth = cell(1,n);
for k = 1:n
pth{k} = cell(1,randi([1 10]));
for kk = 1:numel(pth{k})
pth{k}{kk} = randi([1 100], randi([1 10]), 1);
end
end
% get all combs
str_to_eval = compose('1:length(pth{%i})', 1:numel(pth));
str_to_eval = strjoin(str_to_eval,',');
str_to_eval = sprintf('allcomb(%s)',str_to_eval);
% use eval to get all combinations for a given pth
all_combs = eval(str_to_eval);
% and make strings to eval in intersect
comp = num2cell(1:numel(pth));
comp = [comp ;repmat({'%i'}, 1, numel(pth))];
str_pattern = sprintf('pth{%i}{%s},', comp{:});
str_pattern = str_pattern(1:end-1); % get rid of last ,
strings_to_eval = cell(length(all_combs),1);
for k = 1:size(all_combs,1)
strings_to_eval{k} = sprintf(str_pattern, all_combs(k,:));
end
% and run eval on all those strings
result = cell(length(all_combs),1);
for k = 1:size(all_combs,1)
result{k} = eval(['mintersect(' strings_to_eval{k} ')']);
%fprintf(['mintersect(' strings_to_eval{k} ')\n']); % for debugging
end
For a randomly generated pth, the code produces the following strings to evaluate (where some pth{k} have only one cell for illustration):
mintersect(pth{1}{1},pth{2}{1},pth{3}{1},pth{4}{1},pth{5}{1})
mintersect(pth{1}{1},pth{2}{1},pth{3}{1},pth{4}{2},pth{5}{1})
mintersect(pth{1}{1},pth{2}{1},pth{3}{1},pth{4}{3},pth{5}{1})
mintersect(pth{1}{1},pth{2}{1},pth{3}{2},pth{4}{1},pth{5}{1})
mintersect(pth{1}{1},pth{2}{1},pth{3}{2},pth{4}{2},pth{5}{1})
mintersect(pth{1}{1},pth{2}{1},pth{3}{2},pth{4}{3},pth{5}{1})
mintersect(pth{1}{2},pth{2}{1},pth{3}{1},pth{4}{1},pth{5}{1})
mintersect(pth{1}{2},pth{2}{1},pth{3}{1},pth{4}{2},pth{5}{1})
mintersect(pth{1}{2},pth{2}{1},pth{3}{1},pth{4}{3},pth{5}{1})
mintersect(pth{1}{2},pth{2}{1},pth{3}{2},pth{4}{1},pth{5}{1})
mintersect(pth{1}{2},pth{2}{1},pth{3}{2},pth{4}{2},pth{5}{1})
mintersect(pth{1}{2},pth{2}{1},pth{3}{2},pth{4}{3},pth{5}{1})
mintersect(pth{1}{3},pth{2}{1},pth{3}{1},pth{4}{1},pth{5}{1})
mintersect(pth{1}{3},pth{2}{1},pth{3}{1},pth{4}{2},pth{5}{1})
mintersect(pth{1}{3},pth{2}{1},pth{3}{1},pth{4}{3},pth{5}{1})
mintersect(pth{1}{3},pth{2}{1},pth{3}{2},pth{4}{1},pth{5}{1})
mintersect(pth{1}{3},pth{2}{1},pth{3}{2},pth{4}{2},pth{5}{1})
mintersect(pth{1}{3},pth{2}{1},pth{3}{2},pth{4}{3},pth{5}{1})
mintersect(pth{1}{4},pth{2}{1},pth{3}{1},pth{4}{1},pth{5}{1})
mintersect(pth{1}{4},pth{2}{1},pth{3}{1},pth{4}{2},pth{5}{1})
mintersect(pth{1}{4},pth{2}{1},pth{3}{1},pth{4}{3},pth{5}{1})
mintersect(pth{1}{4},pth{2}{1},pth{3}{2},pth{4}{1},pth{5}{1})
mintersect(pth{1}{4},pth{2}{1},pth{3}{2},pth{4}{2},pth{5}{1})
mintersect(pth{1}{4},pth{2}{1},pth{3}{2},pth{4}{3},pth{5}{1})
As Madphysicist pointed out, I misunderstood the initial structure of your initial cell array, however the point stands. The way to pass an unknown number of arguments to a function is via comma-separated-list generation, and your function needs to support it by being declared with varargin. Updated example below.
Create a helper function to collect a random subcell from each main cell:
% in getRandomVectors.m
function Out = getRandomVectors(C) % C: a double-jagged array, as described
N = length(C);
Out = cell(1, N);
for i = 1 : length(C)
Out{i} = C{i}{randi( length(C{i}) )};
end
end
Then assuming you already have an mintersect function defined something like this:
% in mintersect.m
function Intersections = mintersect( varargin )
Vectors = varargin;
N = length( Vectors );
for i = 1 : N; for j = 1 : N
Intersections{i,j} = intersect( Vectors{i}, Vectors{j} );
end; end
end
Then call this like so:
C = { { 1:5, 2:4, 3:7 }, {1:8}, {2:4, 3:9, 2:8} }; % example double-jagged array
In = getRandomVectors(C); % In is a cell array of randomly selected vectors
Out = mintersect( In{:} ); % Note the csl-generator syntax
PS. I note that your definition of mintersect differs from those linked. It may just be you didn't describe what you want too well, in which case my mintersect function is not what you want. What mine does is produce all possible intersections for the vectors provided. The one you linked to produces a single intersection which is common to all vectors provided. Use whichever suits you best. The underlying rationale for using it is the same though.
PS. It is also not entirely clear from your description whether what you're after is a random vector k for each n, or the entire space of possible vectors over all n and k. The above solution does the former. If you want the latter, see MadPhysicist's solution on how to create a cartesian product of all possible indices instead.
I wrote this matlab code in order to concatenate the results of the integration of all the columns of a matrix extracted form a multi matrix array.
"datimf" is a matrix composed by 100 matrices, each of 224*640, vertically concatenated.
In the first loop i select every single matrix.
In the second loop i integrate every single column of the selected matrix
obtaining a row of 640 elements.
The third loop must concatenate vertically all the lines previously calculated.
Anyway i got always a problem with the third loop. Where is the error?
singleframe = zeros(224,640);
int_frame_all = zeros(1,640);
conc = zeros(100,640);
for i=0:224:(22400-224)
for j = 1:640
for k = 1:100
singleframe(:,:) = datimf([i+1:(i+223)+1],:);
int_frame_all(:,j) = trapz(singleframe(:,j));
conc(:,k) = vertcat(int_frame_all);
end
end
end
An alternate way to do this without using any explicit loops (edited in response to rayryeng's comment below. It's also worth noting that using cellfun may not be more efficient than explicitly looping.):
nmats = 100;
nrows = 224;
ncols = 640;
datimf = rand(nmats*nrows, ncols);
% convert to an nmats x 1 cell array containing each matrix
cellOfMats = mat2cell(datimf, ones(1, nmats)*nrows, ncols);
% Apply trapz to the contents of each cell
cellOfIntegrals = cellfun(#trapz, cellOfMats, 'UniformOutput', false);
% concatenate the results
conc = cat(1, cellOfIntegrals{:});
Taking inspiration from user2305193's answer, here's an even better "loop-free" solution, based on reshaping the matrix and applying trapz along the appropriate dimension:
datReshaped = reshape(datimf, nrows, nmats, ncols);
solution = squeeze(trapz(datReshaped, 1));
% verify solutions are equivalent:
all(solution(:) == conc(:)) % ans = true
I think I understand what you want. The third loop is unnecessary as both the inner and outer loops are 100 elements long. Also the way you have it you are assigning singleframe lots more times than necessary since it does not depend on the inner loops j or k. You were also trying to add int_frame_all to conc before int_frame_all was finished being populated.
On top of that the j loop isn't required either since trapz can operate on the entire matrix at once anyway.
I think this is closer to what you intended:
datimf = rand(224*100,640);
singleframe = zeros(224,640);
int_frame_all = zeros(1,640);
conc = zeros(100,640);
for i=1:100
idx = (i-1)*224+1;
singleframe(:,:) = datimf(idx:idx+223,:);
% for j = 1:640
% int_frame_all(:,j) = trapz(singleframe(:,j));
% end
% The loop is uncessary as trapz can operate on the entire matrix at once.
int_frame_all = trapz(singleframe,1);
%I think this is what you really want...
conc(i,:) = int_frame_all;
end
It looks like you're processing frames in a video.
The most efficent approach in my experience would be to reshape datimf to be 3-dimensional. This can easily be achieved with the reshape command.
something along the line of vid=reshape(datimf,224,640,[]); should get you far in this regard, where the 3rd dimension is time. vid(:,:,1) then would display the first frame of the video.
I have the following function that works perfectly, but I would like to apply vectorization to it...
for i = 1:size(centroids,1)
centroids(i, :) = mean(X(idx == i, :));
end
It checks if idx matches the current index and if it does, it calculates the mean value for all the X values that correspond to that index.
This is my attempt at vectorization, my solution does not work and I know why...
centroids = mean(X(idx == [1:size(centroids,1)], :));
The following idx == [1:size(centroids,1)] breaks the code. I have no idea how to check if idx equals to either of the numbers from 1 to size(centroids,1).
tl:dr
Get rid of the for loop through vectorization
One option is to use arrayfun;
nIdx = size(centroids,1);
centroids = arrayfun(#(ii) mean(X(idx==ii,:)),1:nIdx, 'UniformOutput', false);
centroids = vertcat(centroids{:})
Since the output of a single function call is not necessarily a scalar, the UniformOutput option has to be set to false. Thus, arrayfun returns a cell array and you need to vertcat it to get the desired double array.
you can split the matrix into cells and take the mean from each cell using cellfun (which applies a loop in its inner operation):
generate data:
dim = 10;
N = 400;
nc = 20;
idx = randi(nc,[N 1]);
X = rand(N,dim);
centroids = zeros(nc,dim);
mean using loop (the question's method)
for i = 1:size(centroids,1)
centroids(i, :) = mean(X(idx == i, :));
end
vectorizing:
% split X into cells by idx
A = accumarray(idx, (1:N)', [nc,1], #(i) {X(i,:)});
% mean of each cell
C = cell2mat(cellfun(#(x) mean(x,1),A,'UniformOutput',0));
maximum absolute error between the methods:
max(abs(C(:) - centroids(:))) % about 1e-16
I was trying to loop over an image in MATLAB but my code is running to slow. I am fairly knew to MATLAB but I suspect that it's because it's making a copy of my randomly selected image. My code is:
function patches = sampleIMAGES()
load IMAGES; % load images from disk
patchsize = 8; % we'll use 8x8 patches
numpatches = 10000;
patches = zeros(patchsize*patchsize, numpatches);
size_img = size(IMAGES);
num_rows_img = size_img(1);
num_cols_img = size_img(2);
num_images = size_img(3);
for i=1:numpatches,
%get random image
rand_img_number = randi(num_images);
rand_img = IMAGES(:, :, rand_img_number);
%get random patch patchsizexpatchsize
rand_row = randi(num_rows_img - patchsize);
rand_col = randi(num_cols_img - patchsize);
rand_patch = rand_img(rand_row:rand_row+patchsize-1, rand_col:rand_col+patchsize-1);
patches(:, i) = rand_patch(:)';
end
end
How is it possible to loop over this without making a copy if MATLAB does not allow to index twice into a matrix/array?
Approach #1 - im2col based
numpatches = 10000; %//Number of patches
blksz = 8; %// Blocksize
[m,n,r] = size(IMAGES); %// Get sizes
%// Store blocks from IMAGES as columns, so that they could be processed in
%// a vectorized fashion later on
blks_col(blksz*blksz,(m-blksz+1)*(n-blksz+1),r)=0; %// Pre-allocate
for k1=1:r
blks_col(:,:,k1) = im2col(IMAGES(:,:,k1),[blksz blksz],'sliding');
end
blks_col = reshape(blks_col,size(blks_col,1),[]);
%// Get rand row, column and dimension-3 indices to be used for indexing
%// into blks_col in one go
rand_row = randi(size(IMAGES,1)-blksz+1,numpatches,1);
rand_col = randi(size(IMAGES,2)-blksz+1,numpatches,1);
rand_dim3 = randi(size(IMAGES,3),numpatches,1);
%// Select the specific column from blks_col that represents the
%// [blksz x blksz] used to make a single patch in each iteration from
%// original code
num_cols_im2col = (m-blksz+1)*(n-blksz+1);
col_ind = (rand_dim3-1)*num_cols_im2col + (rand_col-1)*(m-blksz+1) + rand_row;
patches = blks_col(:,col_ind);
Example
As an example I assumed IMAGES as the 3D data obtained from reading one of the images provided in the image gallery of Image Processing Toolbox and increased the number of patches to 100000, i.e. -
IMAGES = imread('peppers.png');
numpatches = 100000;
The runtime with original code - 22.376446 seconds.
The runtime with im2col based code - 2.237993 seconds
Then, I doubled the number of patches to 200000, for which the runtime with original code literally doubled and im2col based approach's runtime stayed around that ~2.3 sec mark.
Thus, this im2col based approach would make sense when you are working with lots of patches as opposed to when working with lots of images (that are put in the third dimension of IMAGES).
Approach #2 - Indexing based
Being a purely indexing based approach, this is expected to be memory-efficient and good with performance too.
numpatches = 10000; %//Number of patches
blksz = 8; %// Blocksize
[m,n,r] = size(IMAGES); %// Get sizes
%// Get rand row, column and dimension-3 indices to be used for indexing
rand_row = randi(size(IMAGES,1)-blksz+1,numpatches,1);
rand_col = randi(size(IMAGES,2)-blksz+1,numpatches,1);
rand_dim3 = randi(size(IMAGES,3),numpatches,1);
%// Starting indices for each patch
start_ind = (rand_dim3-1)*m*n + (rand_col-1)*m + rand_row;
%// Row indices for each patch
lin_row = permute(bsxfun(#plus,start_ind,[0:blksz-1])',[1 3 2]); %//'
%// Get linear indices based on row and col indices
lin_rowcol = reshape(bsxfun(#plus,lin_row,[0:blksz-1]*m),blksz*blksz,[]);
%// Finally get the patches
patches = IMAGES(lin_rowcol);
To no longer copy the images, instead of these two lines:
rand_img = IMAGES(:, :, rand_img_number);
rand_patch = rand_img(rand_row:rand_row+patchsize-1, rand_col:rand_col+patchsize-1);
combine both to one line:
rand_patch = IMAGES(rand_row:rand_row+patchsize-1, rand_col:rand_col+patchsize-1, rand_img_number);
Another way to improve the performance: Generating 100 random numbers at onece is faster than generating 1 number 100 times. Generate all numbers you need outside the loop:
rand_img_number = randi(num_images,numpatches,1);
Then use rand_img_number(i) instead of rand_img_number inside the loop. Do the same for the two other random numbers.