Early stopping at loss increassing or accuracy decreasing?

What do you think that is the better approach for EarlyStopping?

Is a good idea to earlystop at higher valid_loss at validation? And early_stop at accuracy decreasing in test?

1 Like

it depends.Initially the early stopping was used bcz if we train our model too deep then it may not generalize well on real life situations as it is learning from training samples.So we are stopping the train once loss is below the particular threshold.It may be helpful for real life machine learning examples and if you want to get higher accuracy at test samples it is better to train for longer time.