‣ Using preprocessing layers for data augmentation has the following two advantages: The data augmentation will run on GPU in batches, so the training will not be bottlenecked by the data pipeline in environments with constrained CPU resources (such as a Colab Notebook, or a personal machine) Deployment is easier as the data preprocessing pipeline is encapsulated in the model, and does not have to be reimplemented when deploying it (참고1:연관)