Is it possible to save model which gets trained during cross-validation? Basically I do not have test data, so I would like to train the model using cross-validation. How can I do that?
Hi, I also wonder if there is a simple way of basically “stopping early” while training nlu or core models? It seems that the current pipelines support a fixed number of epochs, but with deep models it is very important to have a way to not overfit.
Is there right now a pipeline which handles that by implementing early stopping? (for example based on validation error on validation set). Or a pipeline which would save only the best performning model on the validation set?
I’m also interested in an early stop parameter.
@MetcalfeTom May be someone can address this question? Seems like there are a few community members who have similar concerns.
For early stopping: do you have examples of your models overfitting? Currently we train with dropout and L2-regularization. I’m not sure if early stopping would bring significant benefit.
And as for saving these models, this functionality is not yet included, but mirrors a community PR which we (unfortunately) had to close due to our repository merge. We currently don’t have plans to add this but it would be a welcome contribution if you decided to implement it.