First CNN experiment (w. matlab)

% download image data
url='http://download.tensorflow.org/example_images/flower_photos.tgz';
downloadFolder=tempdir;
filename=fullfile(downloadFolder, 'flower_dataset.tgz');

imageFolder=fullfile(downloadFolder, 'flower_photos');%to uncompress datas

if ~exist(imageFolder,'dir') % download only once
    disp('Downloading Flower Dataset (218 MB)...');
    websave(filename,url);
    untar(filename,downloadFolder)
end

% load images
imds=imageDatastore(imageFolder, 'LabelSource', 'foldernames', 'IncludeSubFolders', true);
daisy=find(imds.Labels=='daisy', 1);
% figure
% imshow(readimage(imds, daisy));

tbl=countEachLabel(imds)%summarize the # of images per cat
%       Label       Count
%     __________    _____
% 
%     daisy          633 
%     dandelion      898 
%     roses          641 
%     sunflowers     699 
%     tulips         799 
% % make images number balanced
% % soing some pruning

minSetCount=min(tbl{:,2});
% limit images number for sake of time
maxNumImages=1000;
minSetCount=min(maxNumImages, minSetCount);
imds=splitEachLabel(imds,minSetCount, 'randomize'); %to tirm the set
countEachLabel(imds) %ora imgs bilanciate

% load pretrained NETWORK (from matlab)
net=resnet50();
net.Layers(end) %per ispezione ultimo layer rete

numel(net.Layers(end).ClassNames) %classi di net
% % nel nostro caso la usiamo per op di classificazione di altre classi....
% %     non le sue native

% division in train and test set
% percentages are 30-70%
[trainingSet, testSet]=splitEachLabel(imds, 0.3, 'randomize'); %randomize per non avere ris statico(?)

% preprocess image for cnn

% % our network only manages rgb images 224x224. without saving 'em all in this format
% % we re gonna use augmentedimageDatastore for the resize and gray to rgb converison on the fly
% % possible to do also other augmentation technique if desired

imageSize=net.Layers(1).InputSize;
augmentedTrainSet=augmentedImageDatastore(imageSize, trainingSet, 'ColorPreprocessing', 'gray2rgb');
augmentedTestSet=augmentedImageDatastore(imageSize, testSet, 'ColorPreprocessing', 'gray2rgb');

% extract training features using cnn

w1=net.Layers(2).Weights;

w1=mat2gray(w1);
w1=imresize(w1,5);

% figure
% montage(w1) %it displays a montage of network weights
% title('first convolutional layer weights') 
% deep layers features available using :  deepDreamImage

% % see mathworks example.......

% extraction form deeper one
featureLayer='fc1000';
trainingFeatures=activations(net, augmentedTrainSet, featureLayer, ...
                        'MiniBatchSize', 32, 'OutputAs', 'columns');
% output is arranged in columns cause it will speed up the svm that follows

% Train a MultiClass SVM classifier using cnn features
trainingLabels=trainingSet.Labels;

% % we train fast svm classifier using a linear solver(SGD solver). it speeds up the training with high
% % dimensional cnn features vector

size(trainingFeatures)
size(trainingLabels)
classifier=fitcecoc(trainingFeatures, trainingLabels, ...
                'Learners', 'Linear', 'Coding', 'onevsall', 'ObservationsIn', 'columns');
            
% valutazione del classificatore
testFeatures=activations(net, augmentedTestSet, featureLayer, ...
                    'MiniBatchSize', 32, 'outputAs', 'columns');
predictedLabels=predict(classifier, testFeatures, 'ObservationsIn', 'columns');

testLabels=testSet.Labels;
confMat=confusionmat(testLabels, predictedLabels);
% converto valori in percentuale
confMat=bsxfun(@rdivide, confMat, sum(confMat,2))

mean(diag(confMat))

% use the classifier on the test set
cnt=0;
for i=1:1:numel(testSet.Labels)
    disp(i);
    testImage=readimage(testSet, i);
    testLabel=testSet.Labels(i);
    
    ds=augmentedImageDatastore(imageSize, testImage, 'ColorPreprocessing', 'gray2rgb');
    imgFeatures=activations(net, ds, featureLayer, 'OutputAs', 'columns');
    predictedLabel=predict(classifier, imgFeatures, 'ObservationsIn', 'columns');
    
    if(predictedLabel==testLabel)
        cnt=cnt+1;
    end
end
acc=cnt/numel(testSet.Labels);
disp(acc)