Saving the best model on epoch validation loss or epoch validation accuracy

I am using the epoch validation accuracy as a metric to save the model current status if there is a progress but I was wondering if it is the right way or should I use the lowest epoch validation loss or a combination of both ?
I am working on multi class classification problem

From my own experience, I always save all model after each epoch so that I can select the best one after training based on validation accuracy curve, validation loss curve and training loss curve.

1 Like

yes but how do you proceed if you dont have the same epoch with the lowest validation loss and the highest validation accuracy?

When everything else is normal, I prefer accuracy.
Others are welcome to answer this question.

1 Like

in my case the validation plots accuracy and loss are not looking like the exemplar curve I suspect an overffiting so i am only releying on the accuracy so far

As what I know, selecting the checkpoint at the lowest loss is the most common way. And as @Eta_C says, I would also recommend to save all checkpoints and do model averaging after the training is done. Here is an example fairseq/ at main · pytorch/fairseq (

let us say I care more about accuracy this is what I should use as improvement and I dont get the best epoch in which the lowest loss is and the highest validation , + the fact that I am getting overfitting

@David_Harvey torchcontrib.optim.SWA also implements model averaging :wink: