clear all
display 'start....'

%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%% We create our arbitary CNN and train it with MNIST dataset
%%%%% The architecture of CNN is arbitrarily choosen for experimental purpose
%%%%% The architecture may be revised for better result.

%   cnnAddConvLayer - Add convolution layer
%   cnn, no_of_feature_maps, sizeof(kernels), activation function -'sigm' 
%   for sigmoid, 'tanh' for tanh, 'rect' for ReLu, 'soft' for softmax, 
%  'none' for none, 'plus' for softplus.

% cnnAddPoolLayer - Add Pool layer
% cnn, subsampling factor, subsampling type. Presently only 'mean'
% subsampling is implemented.

%cnnAddFCLayer - Add fully connected neural network layer
% cnn, no of NN nodes, activation function.


%% xx = xx - mean(xx(:));

%predicting on raw signals of EEG data ................
% load('trainedClassifier_guassiansvm.mat', 'class');
% load('trainedClassifier_guassiansvm.mat', 'features');
%load('C:\Users\d035190\Documents\MATLAB\EEG_sanam\EEG_sanam\EEGeye1.mat')



 inputtrain= inputdata(1:120000,:);
 inputtest= inputdata(120000:end,:);


 outputtrain= outputrain3hour(1:120000,1);
 outputtest=  outputrain3hour(120000:end,1);

train_x=inputtrain';
train_y=outputtrain';

% 
% train_x= x(1:14,1:14000);
% train_y= x(15,1:14000);
% 
% test_x= x(1:14,14001:end);
% test_y= x(15,14001:end);

% initialize cnn
cnn.namaste=1; % just intiationg cnn object
% [x,t] = cancer_dataset;
%  train_x=x(:,1:500);
%  train_y=t(:,1:500);
% % 
%  test_x=x(:,500:end);
%  test_y=t(:,500:end);

h=size(train_x,1);
w= 1;


% initializing the cnn with input layer.........

cnn=initcnn(cnn,[h w]);

%....adding 1 con layer 
cnn=cnnAddConvLayer(cnn, 3, [3 1], 'rect');
% adding pooling layer .......................
cnn=cnnAddPoolLayer(cnn, 1, 'mean');
% 
 cnn=cnnAddConvLayer(cnn, 10, [3 1], 'sigm');
 cnn=cnnAddPoolLayer(cnn, 1, 'mean');

% cnn=cnnAddConvLayer(cnn, 12, [9 1], 'sigm');
% cnn=cnnAddPoolLayer(cnn, 1, 'mean');

% adding fully connected layers.........
cnn=cnnAddFCLayer(cnn,5, 'tanh' ); %add fully connected layer
cnn=cnnAddFCLayer(cnn,1, 'none' ); %add fully connected layer % last layer no of nodes = no of lables



%%
%  below is an another example of CNN that can be used.
% Example 1
%  cnn=cnnAddConvLayer(cnn, 10, [7 7], 'tanh'); 
%  cnn=cnnAddPoolLayer(cnn, 2, 'mean'); %cnn, subsampling factor
%  cnn=cnnAddConvLayer(cnn, 15, [3 3], 'tanh');
%  cnn=cnnAddPoolLayer(cnn, 3, 'mean');
% cnn=cnnAddFCLayer(cnn,150, 'tanh' ); %add fully connected layer
% cnn=cnnAddFCLayer(cnn,10, 'sigm' ); %add fully connected layer
% 

%% Example 2
%  
%  cnn=cnnAddConvLayer(cnn, 40, [5 5], 'sigm');
%  cnn=cnnAddPoolLayer(cnn, 3, 'mean');
%  cnn=cnnAddConvLayer(cnn, 50, [3 3], 'sigm');
%  cnn=cnnAddPoolLayer(cnn, 2, 'mean');
% cnn=cnnAddFCLayer(cnn,90, 'sigm' ); %add fully connected layer
% cnn=cnnAddFCLayer(cnn,10, 'sigm' ); %add fully connected layer
%% 

%%% Example 3
% cnn=cnnAddConvLayer(cnn, 10, [9 1], 'sigm');
%  cnn=cnnAddPoolLayer(cnn, 2, 'mean');
% cnn=cnnAddConvLayer(cnn, 20, [3 3], 'tanh');
%  %cnn=cnnAddPoolLayer(cnn, 2, 'mean');
% cnn=cnnAddFCLayer(cnn,150, 'tanh' ); %add fully connected layer
% cnn=cnnAddFCLayer(cnn,10, 'sigm' ); %add fully connected layer % last layer no of nodes = no of lables

%%
%%%more parameters
%cnn.loss_func = 'cros';

cnn.loss_func = 'quad'; 
no_of_epochs = 550;
batch_size=5;
display 'training started...Wait for ~200 seconds...'
tic
cnn=traincnn(cnn,train_x,train_y, no_of_epochs,batch_size);
toc
display '...training finished.'
display 'testing started....'
tic
err=testcnn(cnn, test_x, test_y);
toc
display '... testing finished. To get minimum error, increase no of epochs while training.'