How can I train a convolutional neural network for both classification and regression?

I would like to use the same convolutional neural network to classify and perform regression on images. In other words, I would like to have shared input and hidden layers, but then branch off into a regression output layer and a classification output layer. How can I do this?

Part of this problem is that I have a lot of float-valued images stored as .mat files, so I would like to use their file names instead of storing all of my data in memory. Is it possible to use an image datastore with 2 labels for each image, or something like it?

As an example, I would like to train a convolutional neural network to classify digits and determine their rotation. MathWorks already has examples for the classification task and for the regression task. I would like to couple the two problems.

Answer 1

[KH TOHIDUL ISLAM]

HI,

If you have not found any solution for this, now you can have one! Please visit the following link!

https://au.mathworks.com/help/deeplearning/ug/train-network-with-multiple-outputs.html

Regards,

ISLAM

[Matthew Fall]

Unfortunately, the example in this link relies on holding all of the data in memory. I am hoping for an out-of-memory solution.