Divide a 4D array into training set and validation set for CNN (regression)
显示 更早的评论
Hi everybody, I am trying to design a CNN for regression following this Matlab example. It uses a 4D array to store the images and vector to store the values associated to every picture. I am using this code to create a 4D array called 'database' that contains my images and a vector 'labels' that contains the values.
k = 1;
%2cm
for i = 1:1000
str = sprintf('images/2cm/%d.jpg', i);
image_to_store = imread(str);
database(:,:,1,k) = (image_to_store(:,:)); % images are in grey scale
labels(k) = 2;
k = k+1;
end
%20cm
for i = 1:1000
str = sprintf('images/20cm/%d.jpg', i);
image_to_store = imread(str);
database(:,:,1,k) = (image_to_store(:,:));
labels(k) = 20;
k = k+1;
end
% ...
Now, I have my 4D array and the vector, so I am trying to divide them into a Training Set and a Validation Set as suggested in the example linked. Can anyone please help me to understand how can I do that?
Thanks
Vale
1 个评论
Muhammad Waris
2021-3-31
hello, what was the size of image? I am trying to do the similar work, made 4Darraydata and also divided that for training and testing.But when i strat to train this code https://ch.mathworks.com/help/deeplearning/ug/train-a-convolutional-neural-network-for-regression.html i encounter with an error "Insufficient number of outputs from right hand side of equal sign to satisfy assignment".
采纳的回答
Jakub Sikorowski
2019-3-9
0 个投票
Hi Vale,
Hope this does what you wanted:
% Your data set:
% The first 1000 entries with labels 2cm,
% The second 1000 entries with labels 20cm,
database = rand(28,28,1,2000);
% percentage of training points = 70%, validation = 30%, test = 0%
p=0.7;
% One way to divide the 2000 database entries
[trainInd,valInd,testInd] = dividerand(2000,p,1-p,0);
trainDatabaseBad = database(:,:,:,trainInd);
valDatabaseBad = database(:,:,:,valInd);
size(trainDatabaseBad) % output: 28 28 1 1400
size(valDatabaseBad) % output: 28 28 1 600
% A better way to divide, which ensures that
% there is equal propotion of 2cm to 20cm samples in
% the training set, validation set, and the whole set
[trainInd1,valInd1,testInd1] = dividerand(1000,p,1-p,0);
[trainInd2,valInd2,testInd2] = dividerand(1000,p,1-p,0);
trainDatabase = cat(4, database(:,:,:,trainInd1), database(:,:,:,trainInd2));
valDatabase = cat(4, database(:,:,:,valInd1), database(:,:,:,valInd2));
size(trainDatabase) % output: 28 28 1 1400
size(valDatabase) % output: 28 28 1 600
Just a word of caution that holding data set in memory might not be optimal if the it is a large data set. If that's the case one can use data stores e.g. image data store. They make the above much simpler with splitEachLabel function, and help with things like data augmentation.
Hope that helps,
Jakub
更多回答（0 个）
类别
在 帮助中心 和 File Exchange 中查找有关 Object Detection 的更多信息
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!
Translated by 
