how do i use k-fold cross validation in deep network designer?

hello, i'm working for my project by using deep network designer to create U-net architecture model adapted of image regression. i need to do k-fold cross validation due to my train dataset;2D image pairs are small (just 30 pairs). can anybody please tell me what should i do for using k-fold for valid the model.

Answer 1

[Rahul]

"deepNetworkDesigner" do not provide k-fold cross validation as such. But as a workaround, you can perform following steps:

Assuming your input images if of size 28x28x1000 and your labels of size 28x28x1000.

imgs: 28 x 28 x 1000

labels: 28 x 28 x 1000

imgs = randi([0,255], 28, 28, 1, 1000); % 1000 input images of size 28x28 single plane
labels = randi([0,255], 28, 28, 1, 1000); % 1000 labeled images of size 28x28 single plane

params.LR = 0.001;
params.Maxepochs = 100;
params.num_batch = 16;

layers  = % your CNN model layers
options = trainingOptions('adam',...
    'InitialLearnRate',params.LR,...
    'MaxEpochs', params.Maxepochs,...
    'MiniBatchSize', params.num_batch); % change this as per your requirement

%% k-fold cross validation
kfold_val = 10; % 10-fold cross validation value
num_samples = 1000;  %  = size(labels, 4);
fold = cvpartition(num_samples, 'kfold', kfold_val); % performing k-fold cross validation

for ii = 1:kfold_val
  train_idx = fold.training(ii);
  validation_idx = fold.test(ii);
  
  % extract training images and labels using train_idx
  xtrain = imgs(:, :, :, train_idx);
  ytrain = labels(:, :, :, train_idx);

  % extract validation images and labels using validation_idx
  xvalid = imgs(:, :, :, validation_idx); 
  yvalid = labels(:, :, :, validation_idx);

  % train the CNN model
  trained_net = trainNetwork(xtrain, ytrain, layers, options);

  % test on validation images
  Pred = predict(trained_net, xvalid);

  % calculate loss between yvalid and Pred
end

Please check documentation links below for your reference:

Image-to-Image Regression in Deep Network Designer - MATLAB & Simulink - MathWorks India

Partition data for cross-validation - MATLAB - MathWorks India

Generate indices for training and test sets - MATLAB crossvalind - MathWorks India

You can also go through below MATLAB central page for more information.

Simple Deep Learning Algorithms with K-fold Cross-Validation - File Exchange - MATLAB Central (mathworks.com)

Please note that this workflow is NOT designed by MathWorks and contact the author in case of any issues.

[patthanee]

thank you so much! i will discuss this with my professor