How refit on train+val set?

[EdgedSquirrels]

How should I refit the model on the whole dataset containing both training and validation sets?

It seems that both nf.fit and nf.cross_validation don't support this feature, but AutoModels have refit_with_val. Could I set refit_with_val on non-auto models?

[marcopeix]

You can simply combine both sets into a single dataframe and pass it to the fit method. Alternatively, if you used val_size, simply set it to 0 to use all of the data.

With cross_validation, you can use the refit parameter and it to True.

[EdgedSquirrels]

Hi @marcopeix,

Thanks for your answer. I’m not sure that when early stopping is enabled, will the number of training steps recorded and used during refit, or it will use max_steps during refit?

Also I’ve tried cross_validation with refit=True, but I only saw a progress bar running without another progress bar running after that. I’m not sure that whether cross_validation supports refit when val_size is given instead of n_windows.

[andmadeira]

Hi @marcopeix !

Taking @EdgedSquirrels's question about early stopping with refit, I noticed that once the early stopping is triggered, the fit for that window stops, and when moving to the following refit window, the model is fitted at least until the first validation step, where the refit is imemdiately stopped in case the validation loss is above the previous window's lowest, not considering patience.
I understand that a refit will continue to fit the model from where it last stopped. Shouldn't the patience counter be reset on every refit?