0

Here is my loss and accuracy after each epoch. It's a sequence-to-sequence model with 4 input and output tokens. enter image description here

If I were to implement early stopping, where would I stop training? Dev loss and accuracy haven't yet started to decline, what would you do?

Thanks.

mdurrant
  • 1

2 Answers2

0

First, keep saving your best model, and once you find for couple of epochs your training loss is going down but at the same time your validation loss is going up, it's time to stop (early stopping). You have your best model saved, that is your trained model.

Pranit Kothari
  • 9,721
  • 10
  • 61
  • 137
0

It's about your target: best accuracy? Best outlier rejection? Best log_loss? If your final target is well defined, then use the right metric for your purpose. Then, train your model until overfitting occurs, or when your metric isn't improving anymore.

Simon Caby
  • 191
  • 6