karolpiczak/ESC-50

How to divide the dataset into train/val/test?

fake-warrior8 opened this issue · 3 comments

fake-warrior8 commented

Hi, I'd like to know how to divide the dataset into train/val/test. It seems to be divided as train/val=8:1. Is this true?

karolpiczak commented

Hi, I am not sure why my previous e-mail reply didn't go through, I missed that.

There are no predefined train/test splits, but the dataset is split into 5 folds of equal size.

A common procedure is to use 4 folds for training and 1 for testing, and perform this 5 times for cross-validation (by switching the fold used for testing). If you need additional validation, you can use 3 folds for training, 1 for validation and 1 for testing.

karolpiczak commented
fake-warrior8 commented

Thank you for your reply!

There are no predefined train/test splits, but the dataset is split into 5 folds of equal size. A common procedure is to use 4 folds for training and 1 for testing, and perform this 5 times for cross-validation (by switching the fold used for testing). If you need validation, you can use 3 folds for training, 1 for validation and 1 for testing.
…
On Wed, Jun 15, 2022, 14:41 LDong @.> wrote: Hi, I'd like to know how to divide the dataset into train/val/test. It seems to be divided as train/val=8:1. Is this true? — Reply to this email directly, view it on GitHub <#18>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AABUUQT6GWXRHAYNF64COVTVPHFPXANCNFSM5Y3FD6DA . You are receiving this because you are subscribed to this thread.Message ID: @.>