Create a multiclass SVM classification with templateSVM and a custom kernel

hi to everybody,

I would like to build a multiclass SVM classificator (20 different classes) using templateSVM() and chi_squared kernel, but I don't know how to define the custom kernel: I tryin the folowing way:

""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""

function gram = compute_gram_matrix(U,V)

[sample_U,~] = size(U);

[sample_V,~] = size(V);

gram = zeros(sample_U,sample_V);

for r=1:sample_U

for t =r:sample_V

temp = chi_squared_kernel(U(r,:),V(t,:));

gram(r,t) = temp;

gram(t,r) = temp;

end

end

% calcolo la media

mean = 0;

for i=1:size(gram,1)

for j=1:size(gram,2)

if i <= j

mean = mean + gram(i,j);

else

continue

end

end

end

mean = mean / ((sample_U^2 + sample_U)/2);

gram = exp(-gram/mean);

end

function value = chi_squared_kernel(hist_1,hist_2)

value = 0;

k = size(hist_1,1);

for i=1:k

if hist_1(i)==0 && hist_2(i)==0

continue

else

value = value + (hist_1(i) - hist_2(i))^2 / (hist_1(i) + hist_2(i));

end

end

value = 0.5 * value;

end

"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""

but it says that the kernel is not of the correct type: How can I do to solve this problem?

Answer 1

[Shashank Gupta]

Hi Alberto,

I just took your gram matrix and able to define it properly, can you elaborate what you all did? Just for the reference I will attach a piece of code which I used and it worked.

% take some data.
load fisheriris
% define SVM
t = templateSVM('KernelFunction','compute_gram_matrix');

% Specify template t to binary leaner.
Mdl = fitcecoc(meas,species,'Learners',t);

I took your "compute_gram_matrix" function and able to execute it properly without fail. Check out this and let me know if it make sense.

Cheers.

[Alberto Presta]

thanks for your time.

I am very new with Matlab. I want to buid a multiclass svm classificator with custom kernel (I have 20 different species to classify).

1-I extract dense descriptors (Dense sift descriptors) foe each image and group all toghether with bag of visual words tech (300 words).

2. I extract histograms of bag of visual words and I want to use them to fit my chi-sqaured kernel svm model.

I attached main code here:

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

extractor = @phowFeatures; % function which exracts dense descriptors

bag = bagOfFeatures(trainingSet, 'CustomExtractor',extractor,'VocabularySize',300);

[trainingSet, validationSet] = splitEachLabel(imds, 0.70, 'randomize');

opts = templateSVM('KernelFunction', 'compute_gram_matrix', 'BoxConstraint', SVM_C, 'kernelScale', SVM_RBF_Gamma);

classifier = trainImageCategoryClassifier(trainingSet, bag, 'LearnerOptions', opts);

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

My question is : can I use fitcecoc also for multiclass svm?