0
$\begingroup$

I use 2-fold cross-validation to determine the optimal number of epochs to train a neural network. Lets assume this number is 100 epochs.

As a next step, I want to train the network using the full data set. Should I adjust the number of epochs to 50 epochs or should I train it for 100 epochs?

If I train it for 100 epochs the network effectively sees double the number of observations that it saw during cross-validation, which does not feel right to me. What do you think?

Improve this question
$\endgroup$
2
  • $\begingroup$ @JanKukacka Thanks, that answers the question. However, I wonder why most neural network packages still use epochs instead of training iteratiosn/weights updates. $\endgroup$
    – Funkwecker
    Commented Mar 5, 2020 at 12:43
  • 1
    $\begingroup$ Hard to say... at the end of the day, 1 epoch = n_samples / batch_size iterations, so it's easy to convert between these two, one just has to properly understand what each of them means. $\endgroup$
    – Jan Kukacka
    Commented Mar 5, 2020 at 12:51

0

Reset to default

Browse other questions tagged or ask your own question.