More realistically, the code shown is not like what I would write. Correct me if I'm mistaken please; Our code does not call only one time to polynomialdecay() and prune_low_magnitude() using only one hyperparameter set as shown, but instead would iterate many times to choose satisfactory if not optimal hyperparameters for poly...() and prune...(). So you have to have another round of train, validate, and test on unbiased data just like you did in training. It's a whole iterative processing cycle again. Or if I'm wrong and the parameter values you show in the video here are actually good on the first try, the library designer would probably have provided these as default values for the parameters in these two functions.