1 Objective

In this lab, we will go through convolutional neural network (CNN) to classify the letter in a image. The code "test_example_letter.m" is the one you should execute to train the CNN to identify hand-written numbers.

2 Prerequisties

• Matlab

3 Procedure of CNN

1. Download ./CNN_lab.zip and unzip it
2. Add the path of CNN functions and data into the working path by typing

   clear all,close all

   addpath ./CNN

   addpath ./util

   addpath ./data

3. Load the training data and test data by typing

   load mnist_uint8;

   train_x = double(reshape(train_x’,28,28,60000))/255;

   test_x = double(reshape(test_x’,28,28,10000))/255;

   train_y = double(train_y’);

   test_y = double(test_y’);

4. Visualizingthe data by typing

   train_n=reshape(train_x,784,60000)’;

   sel = randperm(size(train_n, 1));

   sel = sel(1:100);

   displayData(train_n(sel,:));

5. Setup CNN parameters by typing

   rand(’state’,0);

   opts.alpha = 1; %learning rate

   opts.batchsize = 50; %minibatch size

   opts.numepochs = 2; %iteration number

6. Setup CNN structure by typing

   cnn.layers = { struct(’type’, ’i’) %input layer

   struct(’type’, ’c’, ’outputmaps’, 6, ’kernelsize’, 5) %convolution layer

   struct(’type’, ’s’, ’scale’, 2) %sub sampling layer

   struct(’type’, ’c’, ’outputmaps’, 12, ’kernelsize’, 5) %convolution layer

   struct(’type’, ’s’, ’scale’, 2)} %subsampling layer ;

7. Check the input parameters and train CNN by typing

   cnn = cnnsetup(cnn, train_x, train_y);

   cnn = cnntrain(cnn, train_x, train_y, opts);

8. Apply trained CNN on the test data by typing

   [er, bad, nn] = cnntest(cnn, test_x, test_y);

9. Plot the residue curve by typing

   figure;plot(cnn.rL);

10. Visualizing test data and its classification by typing

    sel = randperm(size(test_x, 3));

    sel= sel(1:6);

    figure(’rend’,’painters’,’pos’,[10 10 1200 250]);

    for ii=1:6

    subplot(1,6,ii);

    imagesc(test_x(:,:,sel(ii)));

    axis off;

    [num,letter]=max(nn.o(:,sel(ii)));

    title([’letter is ’, num2str(letter-1)],’fontsize’,24);

    end

11. Visualizing the weights in the first convolution layer

    figure(’rend’,’painters’,’pos’,[10 10 1200 250]);

    for ii=1:6

    subplot(1,6,ii)

    imagesc(cnn.layers{2,1}.k{1,1}{1,ii})

    axis off;

    end

4 Questions

1. Modify the structure of CNN. See how the CNN parameter will change the result.
2. Modify minibatch size and learning rate. See how these parameter will change the convergence rate and the result.

5 PS

If there are any errors in this Lab, please contact: shihang.feng@kaust.edu.sa