Question

Am new to matlab.
Can someone explain me the following code. this code is used for training the neural network

N = xlsread('data.xls','Sheet1');
N = N(1:150,:);
UN = xlsread('data.xls','Sheet2');
UN = UN(1:150,:);
traindata = [N ; UN];
save('traindata.mat','traindata');
label = [];
for i = 1 : size(N,1)*2
if( i <= size(N,1))
%        label = [label ;sum(traindata(i,:))/size(traindata(i,:),2)];
     label = [label ;sum(traindata(i,:))/10];
else
%        label = [label ;sum(traindata(i,:))/size(traindata(i,:),2)];
     label = [label ;sum(traindata(i,:))/10];
end
end
weightMat = BpTrainingProcess(4,0.0001,0.1,0.9,15,[size(traindata,1) 1],traindata,label);

Answer 1

I cannot find a Neural Network toolbox built-in that corresponds to BpTrainingProcess(), so this must be a file you have access to locally (or you need to obtain from the person who gave you this code). It likely strings together several function calls to Neural Network toolbox functions, or perhaps is an original implementation of a back-propagation training method.

Otherwise, the code has some drawbacks. For one, it doesn’t appear that the interior if-else statement actually does anything. Even the lines that are commented out would leave a totally useless if-else setup. It looks like the if-else is intended to let you do different label normalization for the data loaded from Sheet1 of the Excel file vs. data loaded from Sheet2. Maybe that is important for you, but it’s currently not happening in the program.

Lastly, the code uses an empty array for label and the proceeds to append rows to the empty array. This is unneeded because you already know how many rows there will be (it will total up to size(N,1)*2 = 150*2 = 300 rows. You could just as easily set label=zeros(300,1) and then use usual indexing at each iteration of the for-loop: label(i) = .... This saves time and space, but arguably won’t matter much for a 300-row data set (assuming that the length of each row is not too large).

I put documentation next to the code below.

% The functionn 'xlsread()' reads data from an Excel file.
% Here it is storing the values from Sheet 1 of the file 'data.xls'
% into the variable N, and then using the syntax N = N(1:150,:) to
% change N from being all of the data into being only the first
% 150 rows of the data
N = xlsread('data.xls','Sheet1');
N = N(1:150,:);

% Now do the same thing for Sheet 2 from the Excel file.
UN = xlsread('data.xls','Sheet2');
UN = UN(1:150,:);

% This concatenates the two different data arrays together, making
% one large array where N is the top half and UN is the bottom half.
% This is basically just stacking N on top of UN into one array.
traindata = [N ; UN];

% This saves a copy of the newly stacked array into the Matlab data file
% 'traindata.mat'. From now on, you should be able to load the data from
% this file, without needing to read it from the Excel sheet above.
save('traindata.mat','traindata');

% This makes an empty array which will have new things appended to it below.
label = [];

% Because UN and N have the same number of rows, then the training data
% has twice as many rows. So this sets up a for loop that will traverse
% all of these rows of the training data. The 'size()' function can be
% used to get the different dimensions of an array.
for i = 1 : size(N,1)*2

    % Here, an if statement is used to check if the current row number, i,
    % is less than or equal to than the number of rows in N. This implies
    % that this part of the if-statement is only for handling the top half
    % of 'trainingdata', that is, the stuff coming from the variable N.

    if( i <= size(N,1))
       % The line below was already commented out. Maybe it had an old use
       % but is no longer needed?
       % label = [label ;sum(traindata(i,:))/size(traindata(i,:),2)];

       % This syntax will append new rows to the variable 'label', which
       % started out as an empty array. This is usually bad practice, memory-wise
       % and also for readability.

       % Here, the sum of the training data is being computed, and divided by 10
       % in every case, and then appended as a new row in 'label'. Hopefully,
       % if you are familiar with the data, you will know why the data in 'N'
       % always needs to be divided by 10.
       label = [label ;sum(traindata(i,:))/10];

    % Otherwise, if i > # of rows then handle the data differently.
    % Really this means the code below treats only data from the variable UN.
    else
       % The line below was already commented out. Maybe it had an old use
       % but is no longer needed?
       % label = [label ;sum(traindata(i,:))/size(traindata(i,:),2)];

       % Just like above, the data is being divided by 10. Given that there
       % is nothing different about the code here, and how it modifies 'label'
       % there is no need for the if-else statements, and they only waste time.
       label = [label ;sum(traindata(i,:))/10];

    % This is needed to show the end of the if-else block.
    end

% This is needed to show the end of the for-loop.
end


% This appears to be a Back-Propagation Neural Network training function.
% This doesn't match any built-in Matlab function I can find, but you might
% check in the Neural Network toolbox to see if the local function
% BpTrainingProcess is a wrapper for a collection of built-in training functions.

weightMat = BpTrainingProcess(4, 0.0001, 0.1, 0.9, 15,
                              [size(traindata,1) 1], traindata,label);

Here is a link to an example Matlab Neural Network toolbox function for back-propagation training. You might want to look around the documentation there to see if any of it resembles the interior of BpTrainingProcess().