function [samples targets labels classes] = get_MNIST_data(mnist_file, inp_rows, inp_cols, ...

classes_to_select, samples_per_class, resize_samples, samples_coeff)

if ~exist('mnist_file', 'var') || isempty(mnist_file), mnist_file = 'D:/datasets/MNIST/train_data.mat'; end

if ~exist('inp_rows', 'var') || isempty(inp_rows), inp_rows = 32; end

if ~exist('inp_cols', 'var') || isempty(inp_cols), inp_cols = inp_rows; end

if ~exist('classes_to_select', 'var') || isempty(classes_to_select), classes_to_select = '0123456789'; end

if ~exist('samples_per_class', 'var') || isempty(samples_per_class), samples_per_class = 100; end

if ~exist('resize_samples', 'var') || isempty(resize_samples), resize_samples = false; end

if ~exist('samples_coeff', 'var') || isempty(samples_coeff), samples_coeff = 1; end

if ~exist('data_', 'var') || isempty(samples_coeff), samples_coeff = 1; end

n_classes = numel(classes_to_select);

if ~exist(mnist_file, 'file'),

mnist_dir = [fileparts(mnist_file) '/'];

mnist2matlab(mnist_dir)

end

load(mnist_file);

selection = zeros(n_classes, samples_per_class);

for c_i = 1:n_classes,

class_label = uint8(classes_to_select(c_i)-'0');

class_indices = find(labels == class_label); %#ok<*NODEF> % select samples labelled as class_label

class_indices = class_indices(randperm(numel(class_indices))); % randomly permute found indices

class_indices = class_indices(1:samples_per_class); % get only what is needed

selection(c_i,:) = class_indices; % add to the index selection

end

selection = selection(randperm(numel(selection)));

labels = labels(selection);

samples = double(images(:,:,selection)); %#ok<NODEF>

[in_rows in_cols n_samples] = size(samples);

classes = unique(labels);

n_classes = numel(classes);

targets = -ones(n_classes, n_samples);

for cl_i = 1:n_classes,

targets(cl_i, labels == classes(cl_i) ) = +1;

end

if (in_rows~=inp_rows) || (in_cols~=inp_cols),

if resize_samples,

fprintf('Resizing samples [%gx%g]->[%gx%g]:', in_rows, in_cols, inp_rows, inp_cols)

transformed_samples = zeros(inp_rows, inp_cols, n_samples, class(samples));

for i = 1:n_samples,

transformed_samples(:,:,i) = imresize(samples(:,:,i), [inp_rows inp_cols]);

end

fprintf(' done.\n');

else

if (inp_rows < in_rows) || (inp_cols < in_cols),

fprintf('Cropping samples [%gx%g]->[%gx%g]:', in_rows, in_cols, inp_rows, inp_cols)

transformed_samples = samples((1:inp_rows)+floor((in_rows-inp_rows)/2), ...

(1:inp_cols)+floor((in_cols-inp_cols)/2), :);

fprintf(' done.\n');

else

fprintf('Padding samples [%gx%g]->[%gx%g]:', in_rows, in_cols, inp_rows, inp_cols)

transformed_samples = zeros(inp_rows, inp_cols, n_samples);

transformed_samples((1:in_rows)+floor(abs(in_rows-inp_rows)/2), ...

(1:in_cols)+floor(abs(in_cols-inp_cols)/2), :) = samples;

fprintf(' done.\n');

end

end

samples = transformed_samples;

clear transformed_samples

end

fprintf('Normalizing samples:')

for i = 1:n_samples,

sample = samples(:,:,i);

sample = sample - mean(sample(:));

sample_std = std(sample(:));

if sample_std > 0,

sample = sample / sample_std;

end

samples(:,:,i) = sample;

end

samples = samples * samples_coeff;

fprintf(' done.\n');

end