% Change the filenames if you've saved the files under different names

% On some platforms, the files might be saved as

% train-images.idx3-ubyte / train-labels.idx1-ubyte

% fp12 = fopen('train-resize', 'rb');

images = loadMNISTImages('train-resize');

labels = loadMNISTLabels('train-labels-idx1-ubyte');

test_images = loadMNISTImages('t10k-resize');

test_labels = loadMNISTLabels('t10k-labels-idx1-ubyte');

% We are using display_network from the autoencoder code

display_network(images(:,1:100)); % Show the first 100 images

% disp(labels(1:10));

n = 60000;

% test_n = ceil(n*15/85);

test_n = 10000;

scale = 2;

ori_size = 28;

size = ori_size/scale;

% p = zeros(size*size, n);

% for i = 1:n

% for j = 0:size-1

% for k = 1:size

% p(j*size+k,i) = images(j*ori_size*2+k*2,i);

% end

% end

% end

p = images(:,1:n);

t = zeros(10,n);

% test_p = zeros(size*size, test_n);

% for i = 1:test_n

% for j = 0:size-1

% for k = 1:size

% test_p(j*size+k,i) = test_images(j*ori_size*2+k*2,i);

% end

% end

% end

test_p = test_images(:,1:test_n);

test_t = zeros(10,test_n);

for i = 1:n

t(labels(i)+1, i) = 1;

end

for i = 1:test_n

test_t(test_labels(i)+1, i) = 1;

end

net = newff(p, t, [100], {'tansig' 'logsig'}, 'trainrp', ...

'', 'mse', {}, {}, 'divideblock');

net = init(net);

% net.trainParam.lr = 2;

net.trainParam.delta0 = 0.07;

net.trainParam.deltamax = 100;

net.trainParam.epochs = 400;

% net.trainParam.max_fail = 1000;

% net.trainParam.min_grad = 0;

% net.trainParam.epochs = 100;

% net.divideParam.Q = n+test_n;

net.divideParam.trainRatio = 0.75;

net.divideParam.valRatio = 0.15;

net.divideParam.testRatio = 0;

[trained_net, stats] = train(net, p, t);

plotconfusion(test_t, sim(trained_net, test_p))