0
$\begingroup$

In orden to avoid overfitting, how I can choose the correct amount of epochs?

I should be create 3 splits (train, valid and test) for each fold??

$\endgroup$
0

1 Answer 1

1
$\begingroup$

If you want to do this to avoid overfitting, than standard, and less computationally demanding, approach would be to use early stopping, i.e. periodically check test error and stop training when it does not decrease anymore. All the major deep learning libraries have functionalities for enabling early stopping during training, but it is trivial to implement it from the scratch as well.

$\endgroup$
2
  • $\begingroup$ I undestand, but can I use early stop in a cross-validation setup? In that case, it is common have 3 split in cross-validation (train/validation/test)? $\endgroup$
    – Xtalker
    Aug 19, 2020 at 21:14
  • $\begingroup$ @Xtalker you need separate test/validation set for early stopping, since you need to evaluate the test/validation set performance to use it. $\endgroup$
    – Tim
    Aug 20, 2020 at 4:44

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