How to prepare imagedatastore for Image to single value regression

I would like to train a CNN for image regression using a datastore. The output is a single value. I set the Labels in imagedatastore to a value vector(Y_train). But the debug information show the labels in imagedatastore must be a categorical vetcor. Is it possible to use the imagedatastore for regression problems.

ims_train = imageDatastore('F:\imagestore\train_image224', 'FileExtensions', '.tiff', 'Labels', Y_train, 'ReadSize', 128);

% here I defined my network architecture

% here I defined my training options

[net,Info] = trainNetwork(ims_train,lgraph,options);

[Cris LaPierre]

OP's comments moved here

Y_train is labels of the images in the folder named 'train_image224'. Each image in the 'train_image224' has a single value label. The numbel of the images in the 'train_image224' equals to length(Y_train).

To solve the image regression problem, I have tried the following codes :

[net,Info] = trainNetwork(X_train,Y_train,lgraph,options).

It works well.

But I want to use 'imagestore' to solve the regression problem. Since overfitting issues occur in my model, I want to augment images using the following codes:

augImdsTrain = augmentedImageDatastore([227 227], ims_train, ...'DataAugmentation', aug);

[Cris LaPierre]

Question from Sho Wright

Hi Qi Lu, I'm also trying to use imds and augimds to preprocess images for an image-to-value regression network, could you share your code if possible?

[Cris LaPierre]

Qi lu's reply

Sorry. I don't find any solutions or workaround. It seems it's impossible to use imds to solve regression problems since the he labels in imagedatastore must be a categorical vetcor. But the labels in regression problems are values. Now I just load all image in the RAM and process image.

Answer 1

[Joe Chalfoun]

This solution worked for me:

Just get the list of files name from the regular imagedatastore and create a table of two columns, the first one is the files name and the second column is the regression response. You can then input that table to trainNetwork function or for the validation option too, like this:

For training:

[net, info] = trainNetwork(table(ims_train.Files,Y_train),lgraph,training_options);

to add validation data:

options = trainingOptions('ValidationData',table(ims_val.Files,Y_val));

Hope this helps.

[David]

The internet and helpful people save the day. Thank you!