NeuralForecast cross validation refitting models

[egolub22]

I have a few questions about the implementation of cross validation in Neuralforecast. Currently, I am instantiating a NeuralForecasting object with many AutoModels and then running cross validation with a defined number of windows. I take the output of this crossval and run a custom eval function which gives me the best models. I want to take these best models and train them on all of the training data and then use them for predictions, without changing any of the hyperparameters. How can I extract the model parameters so that running a fit function would not cause them to be replaced with another set of hyper-parameter tuning?

Alternatively I could first run the fit function, then extract the best parameters and run a crossval on those models to find the best ones, However in this case, should I exclude the data that will be used for the crossval forecasts from the training data to ensure out of sample forecasts?