SO SVM

%% LOAD TRAINING DATA %%

load('ocr_names.mat');
A = 26;
pixels = size(TrnData(1).X, 1);

% training data
y_trn = [];
for i = 1:length(TrnData)
    y_trn = strcat(y_trn,TrnData(i).Y);
end

N_trn = length(y_trn);
x_trn = zeros(pixels, N_trn);
i_tmp = 1;


for i = 1:length(TrnData)
    x_trn(:,i_tmp:(i_tmp+length(TrnData(i).Y)-1)) = TrnData(i).X;
    i_tmp = i_tmp+length(TrnData(i).Y);
end
y_trn = uint8(y_trn)-96;

%% LOAD TESTING DATA %%
y_tst = [];
for i = 1:length(TstData)
    y_tst = strcat(y_tst,TstData(i).Y);
end

N_tst = length(y_tst);
x_tst = zeros(pixels, N_tst);
i_tmp = 1;


for i = 1:length(TstData)
    x_tst(:,i_tmp:(i_tmp+length(TstData(i).Y)-1)) = TstData(i).X;
    i_tmp = i_tmp+length(TstData(i).Y);
end
y_tst = uint8(y_tst)-96;

%% #1 - simple perceptron algorithm %%

w1 = zeros(A, pixels+1);    % classifier in matrix form

max_wrong = 0;
last_wrong = 0;
done = 0;
x_trn = [x_trn; zeros(1, N_trn)];

while ~done

    [~, classified] = max(w1*x_trn);
    missclassified = classified ~= y_trn;
    disp(sum(missclassified));
    missclassified_idx = find(missclassified);
    done = sum(missclassified) == 0;

    for i = 1:sum(missclassified)
        idx = missclassified_idx(i);
        w1(classified(idx),:) = w1(classified(idx),:) - x_trn(:, idx)';
        w1(y_trn(idx),:) = w1(y_trn(idx),:) + x_trn(:, idx)';
    end
end


%% clssify and measure label error %%
wrong = 0;
% classified = zeros(1, N_tst);

[~, classified] = max(w1*[x_tst;zeros(1, size(x_tst, 2))]);
error = mean(classified == y_tst); % 0.7384
disp(error);

%% classify and measure sequence error%%

ok_num = 0;
for i = 1:length(TstData)
    y = uint8(TstData(i).Y) - 96;
    X = [TstData(i).X ; ones(1, length(y))];
    [~, y_class] = max((w1*X));
    ok_num = ok_num + all(y_class == y);
end
error = 1 - (ok_num/length(TstData)); % 0.6660
disp(error);


%% learn sequences %%