clear all
close all

Data=importfile('fashion-mnist_train.csv',2,inf);
input = Data(:,2:end);
target= Data(:,1);

Data=importfile('fashion-mnist_test.csv',2,inf);
input_test = Data(:,2:end);
target_test= Data(:,1);

clearvars Data;

target_string={0 'T-shirt/top';
1 'Trouser';
2 'Pullover';
3 'Dress';
4 'Coat';
5 'Sandal';
6 'Shirt';
7 'Sneaker';
8 'Bag';
9 'Ankle boot'};

figure;
colormap(gray);

perm = randperm(60000,20);
for i = 1:20
    subplot(4,5,i);
    display_array=reshape(input(perm(i),:),[28 28]);
    imshow(display_array',[]);
    xlabel(target_string{target(perm(i))+1,2});
end

% A partir deste ponto os dados já estão importados para o ambiente.
% Entretanto você tem de atentar que:
%     - as entradas e saídas para uma rede neural no MATLAB estão nas linhas
% da matriz;
%     - se o problema é de classificação é desejável codificar as
% saídas categóricas usando o esquema "onehot" onde cada neurônio da camada
% de saída codifica uma dada categoria.

% esquema onehot para a entrada

train_output = zeros(10, 60000);
test_output  = zeros(10, 10000);
confusion_matrix = zeros(10, 60000);
confusion_test_matrix = zeros(10, 10000);

for i = 1:60000
    if target(i, 1) == 0
        train_output(1, i) = 1;

    elseif target(i, 1) == 1
        train_output(2, i) = 1;

    elseif target(i, 1) == 2
        train_output(3, i) = 1;

    elseif target(i, 1) == 3
        train_output(4, i) = 1;

    elseif target(i, 1) == 4
        train_output(5, i) = 1;

    elseif target(i, 1) == 5
        train_output(6, i) = 1;

    elseif target(i, 1) == 6
        train_output(7, i) = 1;

    elseif target(i, 1) == 7
        train_output(8, i) = 1;

    elseif target(i, 1) == 8
        train_output(9, i) = 1;

    else
        train_output(10, i) = 1;

    end

end

%esquema onehot para os testes

for i = 1:10000

    if target_test(i, 1) == 0
        test_output(1, i) = 1;

    elseif target_test(i, 1) == 1
        test_output(2, i) = 1;

    elseif target_test(i, 1) == 2
        test_output(3, i) = 1;

    elseif target_test(i, 1) == 3
        test_output(4, i) = 1;

    elseif target_test(i, 1) == 4
        test_output(5, i) = 1;

    elseif target_test(i, 1) == 5
        test_output(6, i) = 1;

    elseif target_test(i, 1) == 6
        test_output(7, i) = 1;

    elseif target_test(i, 1) == 7
        test_output(8, i) = 1;

    elseif target_test(i, 1) == 8
        test_output(9, i) = 1;

    else
        test_output(10, i) = 1;

    end
end

% normalização dos dados
a_train = min(input(:));
b_train = max(input(:));
ra_train = 1;
rb_train = -1;

a_test = min(input_test(:));
b_test = max(input_test(:));
ra_test = 1;
rb_test = -1;

data_train = transpose((((ra_train-rb_train)*(input-a_train))/(b_train - a_train)) + rb_train);
data_test  = transpose((((ra_test-rb_test)*(input_test-a_test))/(b_test - a_test)) + rb_test);

% criação da rede neural
net = newff(data_train, train_output, [11 7], {'tansig', 'tansig'}, 'traingdx');

net.trainParam.epochs = 500;
net.trainParam.goal = 0.01;
net.trainParam.lr = 0.01;

net.divideParam.trainRatio = 0.8;
net.divideParam.valRatio = 0.2;
net.divideParam.testRatio = 0;

% treinamento da rede
net = train(net, data_train, train_output);

% simulação da rede
Y = sim(net, data_train);
Y_t = sim(net, data_test);

% confusion matrix
for i = 1:60000
    x = max([Y(1,i), Y(2,i), Y(3, i), Y(4, i), Y(5, i), Y(6, i), Y(7, i), Y(8, i), Y(9, i), Y(10, i)]);
    for j = 1:10
        if Y(j, i) == x
            confusion_matrix(j, i) = 1;
        end
    end
end

figure;
plotconfusion(train_output, confusion_matrix);

for i = 1:10000
    x = max([Y_t(1,i), Y_t(2,i), Y_t(3, i), Y_t(4, i), Y_t(5, i), Y_t(6, i), Y_t(7, i), Y_t(8, i), Y_t(9, i), Y_t(10, i)]);
    for j = 1:10
        if Y_t(j, i) == x
            confusion_test_matrix(j, i) = 1;
        end
    end
end

figure;
plotconfusion(test_output, confusion_test_matrix);