0
$\begingroup$

I have 5 classes of pictures to classify:

0 -> ~3 200 (~800 initial number before interference and duplication)

1 -> ~9 000 (I reduced from ~90 000)

2 -> ~8 000

3 -> ~3 000

4 -> ~7 200

How to divide the data?

Now I have divided the data giving 2 000 to test and 2 000 to validation set by taking a fixed number of images (400) from each class. I don't have much knowledge so I don't know if this is a good division of data. The attached picture shows the results on the test data after about 60 epoch of CNN with 15 layers.

enter image description here

The network continues to overfiting, and the results of validation and test set do not improve. I know that I could definitely improve my model but I would like to divide the data in some thoughtful and reasonable way. Pictures are spectrograms and are in RGB format.

$\endgroup$
0
$\begingroup$

I would make the distribution of the classes in test and validation sets the same as in the training set (and as in the whole original set). Anyway all your metrics are relative, not absolute and designed to provide reasonable results when classes are not ideally balanced.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.