It depends.
Whether you need to train your model on your whole data set or not depends on what you want to achieve.
In the blog that you have linked, their intention was to get an estimate of what they call the “skill” of any given model that is trained on the dataset. This is useful when you are considering using multiple different kinds of models on your dataset and have to pick the best one. This is found by partitioning the data. Train on one part and validate on the other. Once they have the estimate, the model is chosen. They refer to this choosing as model finalization. Then, they leverage the entire dataset and retrain the model.
In the workflow from the MathWorks website which you are referring to, the data is partitioned into train, validation and test. The validation set is used to help with training and then the test set is used to measure performance and check for overfitting later. It’s mentioned in this link on: https://www.mathworks.com/help/deeplearning/ug/analyze-neural-network-performance-after-training.html Also, check the last paragraph of this section which talks about the performance plots: https://www.mathworks.com/help/deeplearning/ug/train-and-apply-multilayer-neural-networks.html#bss331l-17 In this case, there is no need for any model finalization.
You can even change the way the data set is split if you want to use all your data for training and validation alone and none for testing. https://www.mathworks.com/help/deeplearning/ref/dividerand.html#f8-2887855
