The MNIST dataset of handwritten digits

Demonstrates loading the MNIST dataset.

Sources:

Contents

MNIST dataset

download/extract files if needed

dirMNIST = fullfile(mexopencv.root(), 'test', 'mnist');
if ~isdir(dirMNIST)
    % download
    baseURL = 'http://yann.lecun.com/exdb/mnist/';
    files = {
        'train-images-idx3-ubyte.gz'
        'train-labels-idx1-ubyte.gz'
        't10k-images-idx3-ubyte.gz'
        't10k-labels-idx1-ubyte.gz'
    };
    disp('Downloading MNIST database, and extracting files...');
    mkdir(dirMNIST);
    for i=1:numel(files)
        gzFile = fullfile(dirMNIST, files{i});
        if exist(gzFile, 'file') ~= 2
            url = [baseURL, files{i}];
            urlwrite(url, gzFile);
        end
        %HACK: unfortunately MATLAB's gunzip (with Java's GZIPInputStream)
        % ignore the stored filename when extracting a GZ file, instead they
        % simply remove the .gz extension! So we use Apache Commons to get the
        % original filename (a Java library that ships with MATLAB).
        % Alternatively we could use system tools like GNU gunzip or 7zip.
        if false
            gunzip(gzFile, dirMNIST);
        elseif false
            system(['gunzip --keep --name ' gzFile]);
        else
            % adapted from gunzip.m
            in = org.apache.commons.compress.compressors.gzip.GzipCompressorInputStream(...
                java.io.FileInputStream(java.io.File(gzFile)));
            fname = char(in.getMetaData().getFilename());
            if isempty(fname)
                [~,fname,~] = fileparts(gzFile);
            end
            out = java.io.FileOutputStream(java.io.File(fullfile(dirMNIST, fname)));
            %org.apache.commons.io.IOUtils.copy(in, out);
            streamCopier = com.mathworks.mlwidgets.io.InterruptibleStreamCopier.getInterruptibleStreamCopier();
            streamCopier.copyStream(in, out);
            in.close();
            out.close();
            clear in out streamCopier fname
        end
    end
end

Load dataset from disk

we give the path to the extracted files to the load method

tic
ds = cv.Dataset('OR_mnist');
ds.load(fullfile(dirMNIST,'/'));  %HACK: path must end with a slash!
toc

Elapsed time is 0.431709 seconds.

Data

dataset contains for each object its image and label as a structure array of images (8-bit 28x28 grayscale) and corresponding labels (0..9)

fprintf('NumSplits = %d\n', ds.getNumSplits());
dtrain = ds.getTrain()
dtest = ds.getTest()
whos dtrain dtest

NumSplits = 1
dtrain = 
  1×60000 struct array with fields:
    label
    image
dtest = 
  1×10000 struct array with fields:
    label
    image
  Name        Size                  Bytes  Class     Attributes

  dtest       1x10000            10160128  struct              
  dtrain      1x60000            60960128  struct              

Display

show one instance from each train/test sets

subplot(121), imshow(dtrain(1).image)
title(sprintf('label = %d',dtrain(1).label))
subplot(122), imshow(dtest(1).image)
title(sprintf('label = %d',dtest(1).label))

Display

show a sample of the first 100 train images corresponding to digit 8

if mexopencv.require('images')
    idx = find([dtrain.label] == 8);
    idx(101:end) = [];
    figure, montage(cat(4, dtrain(idx).image))
end