How to deploy my code on raspberry pi as a standalone?

vid = videoinput('winvideo', 1);
set(vid, 'ReturnedColorSpace', 'RGB');
img = getsnapshot(vid);
 imshow(img)

h =  findobj('type','figure');
n = length(h);
for k=1:n
    
    baseFileName = sprintf('Img #%d.png', k);
fullFileName = fullfile('C:\Users\Cv\Desktop\image classification2 - copy',['img' '.bmp']);
imwrite(img, fullFileName);
end



outputFolder = fullfile('caltech102');
rootFolder = fullfile(outputFolder, '101_ObjectCategories');
 
categories = {'Bottles', 'NotBottles'};
imds = imageDatastore(fullfile(rootFolder,categories),'LabelSource', 'foldernames');
 
tb1 = countEachLabel(imds)
minSetCount = min(tb1{:,2})
 
imds = splitEachLabel(imds, minSetCount, 'randomize');
countEachLabel(imds);
 
Bottles = find(imds.Labels == 'Bottles', 1);
NotBottles = find(imds.Labels == 'NotBottles', 1);

 
% figure
% subplot(2,2,1);
% imshow(readimage(imds,airplanes));
% subplot(2,2,2);
% imshow(readimage(imds,ferry));
% subplot(2,2,3);
% imshow(readimage(imds,laptop));

net = resnet50();

 figure
 plot(net)
 title('Architecture of ResNet-50');
set(gca, 'YLim', [150 170]);

net.Layers(1);
net.Layers(end);
numel(net.Layers(end).ClassNames);
[trainingSet, testSet] = splitEachLabel(imds, 0.3, 'randomize');
 
imageSize = net.Layers(1).InputSize;
 
augmentedTrainingSet = augmentedImageDatastore(imageSize, ...
    trainingSet, 'ColorPreprocessing', 'gray2rgb');
 
augmentedTestSet = augmentedImageDatastore(imageSize, ...
    testSet, 'ColorPreprocessing', 'gray2rgb');
 
w1 = net.Layers(2).Weights;
w1 = mat2gray(w1);
 
 figure
 montage(w1)
title('First Convolutional Layer Weight')
 
featureLayer = 'fc1000';
trainingFeatures = activations(net, ...
    augmentedTrainingSet, featureLayer, 'MiniBatchSize', 32, 'OutputAs', 'columns');
 
trainingLables = trainingSet.Labels;
 
classifier  = fitcecoc(trainingFeatures,trainingLables, ...
    'Learner', 'Linear', 'Coding', 'onevsall','ObservationsIn', 'columns');
 
testFeatures = activations(net, ...
    augmentedTestSet, featureLayer, 'MiniBatchSize', 32, 'OutputAs', 'columns');
 
predictLabels = predict(classifier, testFeatures, 'ObservationsIn','columns');
 
testLables = testSet.Labels;
confMat = confusionmat(testLables,predictLabels);
confMat = bsxfun(@rdivide, confMat, sum(confMat,2));
 
mean(diag(confMat));
 
newImage = imread(fullfile('img.bmp'));


ds = augmentedImageDatastore(imageSize, ...
    newImage, 'ColorPreprocessing', 'gray2rgb');

imageFeatures = activations(net, ...
    ds, featureLayer, 'MiniBatchSize', 32, 'OutputAs', 'columns');

Label = predict(classifier, imageFeatures, 'ObservationsIn','columns');
 
sprintf('The Loaded image belongs to %s class', Label)

[Ahmed abbasi]

This code basically uses CNN and detects whether the image is bottle or not using image processing.

Answer 1

[Walter Roberson]

It is not possible to deploy CNN training to hardware.

It is not possible to deploy augmentedImageDatastore to hardware.

Your strategy would have to be to train on the host, and save the net and activations, and load() it in the code that was deployed to hardware, where you would use it only to predict()

[Ahmed abbasi]

Thank you so much.

i am a bit confused as well. Basically i am develoing a small machine that detects plastic bottles from the Raspberry camera and as the bottle gets detected as plastic, which i am doing using training and CNN, Raspberry turns on the the motors to drive conveyer belt to throw the bottle in the dump.

Should i be using this code?

Main problem with this is that it is not a deploable.

[Walter Roberson]

You need a small machine to detect plastic bottles from the Raspberry camera. But you do not need to train the CNN on that small machine. The small machine needs to be able to run predict() based on the training results, but it does not need to be able to train the CNN.

What you need to do is split the code into two pieces.

• One of the pieces runs on the host computer and uses the augmented data store to train the CNN and store that in a .mat file.
• The second piece is to be deployed to the raspberry pi. It loads the .mat file, and starts the camera, reads frames, does any resizing and rgb2gray that might be needed, and runs predict() using the loaded network. NO TRAINING

[Ahmed abbasi]

Okay now i got your point. Training will be done in the Host computer and than deploy and ras pi for prediction. Thank you so much. Appreciated SIR.

[amgad]

A small question:

how to deploy 'load' to Raspberry Pi in order to have the code run as standalone