Currently overfitting is detected if the validation error does not decrease in the past 32 epochs.
This is not correct as the validation error can decrease after a larger number of epochs (see using policy=all_epochs in trainers).
Potential solution: use the average loss value to check for overfitting?!
Currently overfitting is detected if the validation error does not decrease in the past 32 epochs. This is not correct as the validation error can decrease after a larger number of epochs (see using policy=all_epochs in trainers).
Potential solution: use the average loss value to check for overfitting?!