Split dataset into three different size sets without overlapping

I am working on image processing using Matlab. I need to split a large dataset into three non-overlapped subsets (25%, 25% and 50%). The dataset (let's say has 1K images) has 10 classes (each has 100 images). from class 1, 25% of images should be in the training set, other 25% should be stored in the validation set and the rest (50%) should be stored in the testset. there should not repetition. I mean if an image from a class has been stored in a subset, it must not be stored in other subsets of the class. How do I do that in Matlab?

My code is as follows:

      load ('data.mat')    
      for i = 1:size(data, 1)
              for j = 1:78
                  if mod(i,2)==0
                      trainingset(i/2,j) = data(i,j);
                  else
                      remainset((i-1)/2+1,j) = data(i,j);
                  end
              end
      end

          for i = 1:size(remainset, 1)
              for j = 1:78
                  if mod(i,2)==0
                      testset(i/2,j) = remainset(i,j);
                  else
                      validationset((i-1)/2+1,j) = remainset(i,j);
                  end
              end
          end

Although it somehow works, I am looking for a better algorithm as some parts of data are lost.

[George]

Can you use datasample ?

[david jones]

As I need to split the data into three subsets, using 'datasample', it calculates indices for one subset. But, if I use it again to calculate indices of other subsets, is likely to have duplicate indices in different subset. I can use randperm, but the same issue exists. I need to split the dataset into three different subsets that each of the subsets contains a percentage of each class of data. using simple sampling method like using

   1:250,1:250,1:500

does not work as the subsets have members of some of the classes. Example: subset 1 should have 25% of class1, 25% of class 2, 25% of class 3, ... , 25% of class n. subset 2 should have 25% of class1, 25% of class 2, 25% of class 3, ... , 25% of class n. subset 3 should have 50% of class1, 50% of class 2, 50% of class 3, ... , 50% of class n.

intersection of subset 1,subset 2 and subset 3 must be zero and union of subsets must cover the whole dataset.

Answer 1

[Frank B.]

Here is a quick answer using datasample, for a single vector named data. Loop over your classes or use indexes if they have to be shared.

 load ('data.mat')

 % Declaring data division ratio
 % 25% for training, 25% for validation, 50% for test
 dataset_div=[0.25 0.25 0.5];

 % Number of data in each set
 nb_train=(dataset_div(1)/sum(dataset_div))*length(data);
 nb_valid=(dataset_div(2)/sum(dataset_div))*length(data);
 nb_test=(dataset_div(3)/sum(dataset_div))*length(data);

 % Splitting data in 3 un-overlapping vector

 % Training data
 [data_train,idx_sample]=datasample(data,nb_train,'Replace',false);
 % Removing used values
 idx_left=1:length(data);
 idx_left(idx_sample)=[];
 val_left=data(idx_left);

 % Validation data
 [data_valid,idx_sample]=datasample(val_left,nb_valid,'Replace',false);
 % Removing used values
 idx_left=1:length(val_left);
 idx_left(idx_sample)=[];
 val_left=data(idx_left);

 % Test data
 [data_test,idx_sample]=datasample(val_left,nb_test,'Replace',false);

Cheers