The format of data for a deep learning model CAE.

Question

영광 박 2024-7-18

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2138302-the-format-of-data-for-a-deep-learning-model-cae

回答： Sahas 2024-7-18

I'm looking to create a deep learning CAE model. Inside the "trainnet" folder are JPEG files that I intend to use as input and output images for training the model. How should I format these images to trainnet for learning? For example, please explain the format and method for data within "trainnet(X)". Thank you.

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Sahas 2024-7-18

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2138302-the-format-of-data-for-a-deep-learning-model-cae#answer_1487481

在 MATLAB Online 中打开

Hi @영광 박,

As per my understanding, you want to create a Convolutional Autoencoder model, and need guidance in formatting the images that you have and use “trainnet” function for training the network.

Please go through the provided sample script below which makes use of image formatting functions such as “imageDatastore”, “augmentedimageDatastore” various image formatting that MATLAB provides and in the end use the “trainnet” function to train the network. The “trainnet” function is supported in MATLAB R2023b onwards.

% Name of the folder where input images are stored
imageFolder = 'trainnet';
% Using imageDatastore to store and accessing the images
imds = imageDatastore(imageFolder, 'FileExtensions', '.jpg', 'LabelSource', 'none');
% Read the first image from imageDatastore and get it's metadata
sampleImage = readimage(imds, 1);
[imageHeight, imageWidth, numChannels] = size(sampleImage);
% Preprocess the images using "augmentedImageDatastore" if necessary
% Initialize an array to store the images
numImages = numel(imds.Files);
X = zeros(imageHeight, imageWidth, numChannels, numImages, 'single');
for i = 1:numImages
    img = readimage(imds, i);
    % Convert to single precision
    img = im2single(img); 
    X(:, :, :, i) = img;
end
% Split the data into training and validation sets, if required
[trainInd, valInd] = dividerand(numImages, 0.8, 0.2);
XTrain = X(:, :, :, trainInd);
XVal = X(:, :, :, valInd);
% Define the CAE Layers according to your architechture
% NOTE: This is just a sample CAE architecture
layers = [
    imageInputLayer([imageHeight, imageWidth, numChannels], 'Normalization', 'none')
    convolution2dLayer(3, 16, 'Padding', 'same')
    reluLayer
    maxPooling2dLayer(2, 'Stride', 2)
    convolution2dLayer(3, 8, 'Padding', 'same')
    reluLayer
    maxPooling2dLayer(2, 'Stride', 2)
    transposedConv2dLayer(3, 8, 'Cropping', 'same')
    reluLayer
    transposedConv2dLayer(3, 16, 'Cropping', 'same')
    reluLayer
    transposedConv2dLayer(3, numChannels, 'Cropping', 'same')
    regressionLayer];
% Specify training options
options = trainingOptions('adam', ...
    'MaxEpochs', 50, ...
    'InitialLearnRate', 1e-3, ...
    'MiniBatchSize', 32, ...
    'Shuffle', 'every-epoch', ...
    'ValidationData', {XVal, XVal}, ...
    'ValidationFrequency', 10, ...
    'Verbose', false, ...
    'Plots', 'training-progress');
% Train the CAE
net = trainnet(XTrain, XTrain, layers, options);