When to stop training

Hi ! I am new to deep learning and have some questions about training networks. Hope someone can help me clear my doubts.

When i train my network i usually stop training when i see that the training loss plot have converged. However this always produces a validation loss plot that decreases and increases after a point. Meaning the network have over fitted right ? Should i stop training at that point when validation loss starts to increases even though training loss has not converged

Thanks in advance

You can save the state of your model regularly during training, in different files, so at the end you can choose which one to use, by looking at your losses plot.

To get the best results on data that is not in the training set, you’d want to use the model state saved at the lowest validation loss (just before the over-fitting starts to increase validation loss).

1 Like

Thank you for the advice :smile: