clear
close all

load('LPALIC_Latin_ManSplit.mat')

n=length(trainimages);
k=ceil(0.2*length(trainimages));
rng('shuffle');
ind=randperm(n,k);
testimages=trainimages(:,:,1,ind) ;
testlabels=trainlabels(ind);
indx=1:n;
indx(ind)=[];
trainimage=trainimages(:,:,:,indx);
trainlabels(ind)=[];

%% Define the Network Layers

layers = [
    imageInputLayer([28 28 1],'Normalization','none','Name','input')
    convolution2dLayer(5,64,'Name','conv1')
    batchNormalizationLayer('Name','BN1')
    reluLayer('Name','relu1')
% 24x24
    convolution2dLayer(5,128,'Name','conv2')
    batchNormalizationLayer('Name','BN2')
    reluLayer('Name','relu2')
% 20x20
    maxPooling2dLayer(2,'Stride',2,'Name','max1')
% 10x10 
    convolution2dLayer(5,176,'Name','conv3')
    batchNormalizationLayer('Name','BN3')
    reluLayer('Name','relu3')
% 6x6 
    convolution2dLayer(5,208,'Name','conv4')
    batchNormalizationLayer('Name','BN4')
    reluLayer('Name','relu4')
% 2x2   
    maxPooling2dLayer(2,'Stride',2,'Name','max2')
  
    fullyConnectedLayer(35,'Name','fc')
    softmaxLayer('Name','softmax')
    classificationLayer('Name','classOutput')];

%% Specify the Training Options
% After defining the layers (network structure), specify the training
% options. Set the options to default settings for the stochastic gradient
% descent with momentum. Set the maximum number of epochs at 15 (an epoch
% is a full training cycle on the whole training data), and start the
% training with an initial learning rate of 0.0001.
options = trainingOptions('sgdm','LearnRateSchedule','piecewise',...
      'MiniBatchSize',120,'LearnRateDropFactor',0.7,... 
      'LearnRateDropPeriod',2,... 
      'MaxEpochs',15, ...
	'InitialLearnRate',1/10,...
    'Shuffle','every-epoch');  

   convnet = trainNetwork(trainimage,trainlabels,layers,options);

YTest = classify(convnet,testimages);
accuracy = sum(YTest == testlabels)/numel(testlabels);
Error=(1-accuracy)*100

% YTrainTest = classify(convnet,trainimage);
% Taccuracy = sum(YTrainTest == trainlabels)/numel(trainlabels);
% TError=(1-Taccuracy)*100