2

I was wondering whether it would be possible to have multiple conditions that need to be true for early stopping to take place. So for example, would it be possible to only stop training if the validation accuracy at the given epoch was > 90 % AND the validation loss has not been decreasing for the 10 epochs after it (patience = 10)?

1337_N00B
  • 47
  • 8
  • Is there a reason that you're monitoring both accuracy and loss? Your loss already captures the accuracy performance of the model. In fact it contains more information because it also quantifies the certainty of the model, whereas your model could make accurate predictions but it could be uncertain about most of them. If you were monitoring one metric, you can pass it multiple parameters including min_delta and patience: https://stackoverflow.com/questions/43906048/which-parameters-should-be-used-for-early-stopping – Desmond Cheong May 28 '20 at 18:07
  • @DesmondCheong I was not aware that the validation loss included that! In that case monitoring only the validation loss should work just fine. Thank you so much! – 1337_N00B May 29 '20 at 09:47

0 Answers0