Some sources consider a test/train split, such as with sklearn, to be expected practice, and validation is more or less reserved for k-fold validation.

However, Keras has a somewhat different approach with its validation_split parameter. Different sources report different things on the subject, some suggesting that this replaces test/train splitting, and it seems it should obviously not be confused with k-fold cross-validation.

Can anyone confirm or clarify what is generally expected among keras users on the subject?


1 Answer 1


After some additional digging I came across this issue at the Keras source repository which seems to outline the usage and some of the confusion surrounding the nomenclature of Keras' validation set. According to this, it appears it is correct to say that the validation set is equivalent to a test set, and the naming reflects how it is used to help assess the training process itself during training.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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