Although it does not have any mention in the assignment explicitly, I just wanted to confirm if we need to use validation_split in the final training phase of the best model (part h) or not.

Edit: My bad. I tried to fit the best model again, when it was already a fitted one return by the grid search. Please ignore the comment above.

I think my original question holds.

After grid search in part-g, it returns the best params and epochs. It also returns best model found.

Now in last step i.e. part h, it says:

... Now, test your model with the best batch_size and epochs on the testing set...

Does this mean that we have to fit the best found model again (with best batch_size and epochs) on the training data, and then test it against testing data?
Or, do we just use the best found model in grid search from previous step as it is, to test it against testing data? Doing this makes sense to me because the best model is already trained as mentioned in the assignment doc in the very next line:

... grid.best_estimator_.model gives you the best model found and trained in the gridsearch.

Yes, just test (evaluate) the best model that you found in the Gridsearch on the testing data. Let me know if you have any other questions.