This question is somewhat related to: Is it correct to join training and validation set before inferring on test-set?
As far as I understand, normalisation in general is done in the following way:
- Split the data
- Normalise train data
- Use mean+std from 2 for normalising validation and test data
- Train model and tune hyperparameters
Now we have a model we are happy with and hyperparams. Above question suggests that it's good then train a model using the train+validation data together. But I am confused, do I leave them normalised as they are and combine them together? Or do I calculate a fresh new normalisation on the combined sets and then recalculate test data with the new normalisation values?